Good writing takes real work to produce responsibly: checking facts, weighing conflicting evidence, deciding what actually matters, then rewriting the argument until a stranger to the topic can follow the reasoning without help. Journalists and independent writers cannot keep doing that work if readers expect the result to be free forever.

I can understand why a well-researched article should cost money (even using GenAI as co-writer). I have to admit, though, that I started myself using paywalls for my paid articles by simply copying industry standards blindly and always felt uncomfortable while placing them and where. I do not object to paying for writing. What bothers me, however—often even more than the price itself—is the product I’m apparently supposed to buy: an article behind the paywall that offers me little value.

I click a link because a headline raises a question I actually care about. The opening paragraphs set up the situation and get me invested enough to want the answer. Then the article simply stops. A paywall appears, and to continue I am asked to subscribe to an entire publication, even though I may have arrived because of one specific question and might not return for weeks.

From the publisher’s side, this looks like an article and a potential subscriber. From mine, it looks different. I did not click because I wanted access to an article. I clicked because I wanted help understanding something I did not yet understand. I may want to know why an industry is shifting, whether some new technology actually matters, or how a political decision will ripple through a wider system. In my own professional reading, I am usually trying to understand a subject well enough to connect it to what I already know, notice its implications, and make a sharper judgment call at work the next day.

The job I am hiring the article for is rarely “help me finish this article.” It is closer to something like: help me understand this subject in its wider context, connect it with what I already know, and become better able to judge what it means. That distinction changed how I think about paid writing, and it points to a broader product lesson than any single paywall decision.

Products Are Visible. Jobs Are Not.

Clayton Christensen and his colleagues argued that customers hire products and services to make progress in a particular circumstance.[1] Someone buys a drill, but the drill was never the desired outcome. Someone enrolls in a course, but sitting through lessons is not the progress they were after. Someone commissions a report, though owning another document was rarely the real objective.

Jobs to Be Done asks us to look past the visible transaction and into what is actually happening in the customer’s life at the moment they choose a product. What progress are they trying to make? What are they using instead right now? What has made the situation uncomfortable enough that they are willing to change their routine or spend money to fix it?

This sounds obvious until you try applying it to your own work. Companies describe what they produce almost by reflex. Software firms sell platforms. Consultants sell projects. Trainers sell workshops. Publishers sell articles and subscriptions. Internal innovation teams offer methods, processes, and support. Customers, though, rarely experience any of that as the job itself. They might be trying to avoid an expensive mistake, build confidence before committing to something bigger, settle a disagreement, cut down the time it takes to understand a complicated situation, or become capable enough to handle similar decisions without outside help the next time.

When the product and the job get confused, teams keep improving the thing they produce while the customer’s actual progress barely moves. A publisher writes more articles. A course adds more lessons. A software company bolts on more features. A consultancy hands over a thicker report. The customer ends up with more product, not necessarily more progress.

The Trouble With the Single-Article Paywall

A paywall can make perfectly good commercial sense. Journalism needs revenue, and outlets with distinctive reporting, a strong reputation, or an established reading habit among their audience can reasonably ask readers to pay for continued access.

Still, the wider market shows how hard this has become. The Reuters Institute’s 2026 Digital News Report found that the share of people paying for online news across the twenty countries it tracks held at 17 percent, and personal benefit was cited more often than any sense of civic duty toward journalism, though both played some role.[2] A separate study of 42 newspapers that introduced paywalls found daily page views fell by roughly 30 percent on average, with the size of the drop depending on the publication and how distinctive its content actually was.[3] A paywall can still lift revenue even as it costs a publication a third of its readers, but the pattern is telling: plenty of readers respond to a paywall by simply leaving, not by paying to stay.

It would be a mistake to conclude from this that readers just refuse to pay for knowledge. People spend real money on books, professional education, training, conferences, consulting, coaching, specialist software, and communities, all the time. They pay when they believe the offer will help them make progress that actually matters to them. The conventional single-article paywall usually makes a weaker case than any of those. It asks the reader to pay because the publisher has stopped granting access, not because a new and more valuable job is about to begin. The payment removes an obstacle. It does not, by itself, improve the outcome.

That was the discomfort I eventually had to admit about my own writing. I was trying to build a modest income from work that takes real effort, and yet I had backed into a model where the paid benefit was simply “the article keeps going.” I could justify why I needed to get paid. I had not yet explained why the reader should want to pay. Those turned out to be two very different questions.

Readers May Be Buying an Educational Journey, Not an Article

An article is a fine delivery format. It can raise a question, build an argument, and expose a reader to evidence they would not otherwise have encountered. But one article rarely builds real capability on its own.

When I pay for serious professional writing, I am not really after another piece of information. I want each piece to add something to what I already understand, to surface connections I had not noticed, to challenge an assumption I have been quietly carrying into my work, and to gradually sharpen the questions I know how to ask. The value is supposed to accumulate. The tenth article should be worth more than the first, precisely because I have read and applied the previous nine. Concepts should start to connect. Distinctions that once blurred together should become easy to tell apart. Examples pulled from different industries should build into a kind of repertoire I can reach for the next time an unfamiliar problem shows up.

Seen this way, the reader is not really buying access to a pile of articles. The reader is investing in an educational journey. The immediate job might be understanding one topic in its wider context. The longer job is becoming more capable of making sound judgments in that domain over time.

That reframing widens the competitive field considerably. A publication built around that job is not competing only with other newsletters or newspapers. It is competing with business schools, executive education, professional training, workshops, consultants, coaches, books, conferences, specialist communities, AI assistants, and plain lived experience. These alternatives look nothing alike on the surface, yet a reader might hire any of them for roughly the same underlying progress: help me understand situations I will keep running into at work and get better at handling them. That is a considerably higher bar than producing premium content.

Information Gets More Valuable Once It Enters the Work

Bloomberg is a useful example here precisely because it operates deep inside the information business without treating the single article as its core unit of value. Financial professionals can get news, market data, and company announcements from plenty of sources. Bloomberg instead weaves news together with data, analytics, communication, and tools, right inside the workflows where people actually investigate situations and make decisions. The company frames its news product around turning insight into action, not around producing more stories.[4]

The lesson here is not that every publisher should go build a terminal. Bloomberg serves a specialized institutional market at a scale that has almost nothing in common with independent writing. But there is a narrower, transferable idea underneath it: once information becomes abundant, paid value can shift from controlling access toward helping people actually use that information.

Research on service and value creation backs this up. Value is not simply manufactured by a provider and handed intact to a passive customer. It emerges in use, inside the customer’s own context, and it can be strengthened further through the interaction between provider and user.[5][6] An article can carry good research and an original argument, but its real value only shows up once a reader uses it to make sense of a situation, question an assumption, or change a decision they were about to make.

Which suggests the paywall might simply be sitting in the wrong place. The free article can do the intellectual work of helping a reader recognize and understand a problem. The paid product can start where the reader wants to apply that understanding, connect it to what they already know, and build better judgment through practice. A paywall should not charge someone to resume a journey that got cut off mid-sentence. It should mark the start of a different one.

A Test for Whatever You Are Selling

(I hope this is a fair place to put the paywall after you got enough value and insights)

Read more

Netflix is usually presented as one big pivot: from DVDs to streaming.

That interpretation is too neat.

Netflix did not replace one business model once. It repeatedly changed whichever element was becoming the constraint on its next stage of growth.

Subscription removed transaction friction. Streaming removed physical distribution. Device partnerships expanded access. Originals reduced dependence on licensed content. Global expansion increased reach. Advertising, paid sharing, games and live programming created new forms of engagement and revenue.

Each adaptation solved a constraint. Each also created another.

The familiar story hides the more useful pattern

The familiar Netflix story moves cleanly from DVD rental to streaming. It is attractive because it compresses a complicated history into one decisive strategic move.

But the transition to streaming did not complete the transformation. It changed the constraint.

Physical distribution became less important, but access to licensed content became more important. Original programming reduced part of that dependence, but required larger and less reversible content commitments. Global expansion increased the potential market while introducing localisation, regulation and content-production complexity.

More recently, slowing subscriber growth moved the constraint again. Netflix responded with an advertising tier, tighter household access, games and live programming. The problem was no longer simply how to reach more viewers. It was also how to capture more value and increase engagement from the audience already within reach.

Four phases of adaptation

1. Remove transaction and distribution friction

The original subscription model replaced individual rental decisions, due dates and late fees with recurring access. Streaming then removed the delay and cost of physical fulfilment, while device partnerships made the service available beyond the computer.

2. Reduce dependence and differentiate the offer

Streaming gave Netflix a new channel, but much of the value still depended on content owned by others. Original programming made the service more distinctive and gave Netflix greater control over availability, timing and global distribution.

3. Expand the addressable market

International expansion turned Netflix from a US service into a global platform. This increased potential scale but required local content, language capabilities, market knowledge and a much more complex production system.

4. Expand engagement and revenue capture

Games broadened the entertainment proposition. Advertising introduced advertisers as a second customer group. Paid-sharing rules converted part of informal access into revenue. Live programming added urgency, appointment viewing and additional advertising inventory.

Three lessons from the case

1. Continuous adaptation includes reversals

Qwikster belongs in the chronology because business-model adaptation is not a clean sequence of correct decisions. Netflix separated DVD and streaming pricing, proposed a separate service and then reversed the brand split after encountering strong contrary evidence.

The reversal was not separate from the adaptation process. It was part of it.

2. Removing one constraint creates another

Streaming reduced fulfillment friction but increased dependence on licensed content. Originals created differentiation but increased fixed commitments. Global scale increased reach but also increased localization complexity.

Business-model adaptation rarely eliminates uncertainty. It relocates it.

3. The object of adaptation changes over time

Netflix initially changed how customers accessed and paid for entertainment. Later changes increasingly concerned content ownership, market scope, engagement and revenue capture.

The relevant question was not simply whether the existing model still worked. It was whether it could support the company’s next stage of growth.

Future cases will examine Adobe and other organizations that changed their business models as old assumptions became constraints. Subscribe for source-backed cases, strategic interpretation and practical decision tools.

Netflix — Continuous Business Model Innovation, Case 01

54.7KB ∙ PDF file

Download

A two-page, source-backed chronology of fourteen business-model adaptations, including a Business Model Canvas View, Core Pattern and Strategic Tension.

Download

What would you challenge?

Which Netflix adaptation created the most consequential new constraint?

I am particularly interested in interpretations that challenge the chronology, the business-model elements I assigned or the core pattern I identified.

Use this case with your team

I am developing Continuous Business Model Innovation as a format for executive briefings, strategy off-sites, workshops, bootcamps and teaching.

The purpose is not to copy Netflix. It is to help a team determine which part of its own business model has become a constraint, which assumptions require evidence and what should be tested before committing further resources.

If your organization is facing such a decision, send me the concrete situation you are working through.

This visual reflects my personal interpretation of Netflix’s business-model adaptations based on available public data and public sources. Netflix and related marks are trademarks of Netflix, Inc. This independent analysis is not affiliated with or endorsed by Netflix.

This morning I read a piece on SRF about a team at ETH Zürich that has spent nearly a decade building a chip to fight deepfakes. The chip sits directly on a camera sensor and generates a cryptographic signature the moment an image is captured. Anyone who alters the image after the fact invalidates the signature. Felix Franke, the professor leading the project, says he saw the problem coming ten years ago and decided to work on it rather than wait for someone else. The chip is still a prototype. The problem it addresses is not.

That piece landed differently than most I have read on the subject. Not because the technology surprised me. I have watched deepfakes improve steadily for years, and the trajectory is not subtle. What struck me was the framing: authenticity now needs to be verifiable at the hardware level, because human eyes passed their useful detection threshold some time ago.

My spam folder has been telling me the same story. What used to arrive as obviously fraudulent, misspelled names, impossible offers, transparent pretexts, has gradually become more careful. Some messages now use language calibrated to sound like a colleague, a trusted institution, or a service I actually use. The writing has gotten better. The targeting has improved. The cost of producing a convincing fake has dropped, and the volume has gone up. That combination is not an accident. It is the signature of a business model that found a technology it could use.

The question that tends to get avoided

Whenever deepfakes come up in public discussion, the conversation lands quickly on the technology: how realistic it is, how fast it is improving, whether detection can keep pace, and what regulators might eventually do. Those are reasonable questions. But there is a prior one that gets less attention.

Who owns a person’s face, voice, image, and identity? And who gave anyone else the right to use them for advertising, content, entertainment, fraud, or profit?

The law is genuinely unclear in some places. Enforcement is weak in others. Platforms move slowly, and by the time a policy is updated, the next version of the tool has already changed shape. But the basic question is not that complicated. It is about consent and about who carries the cost when something goes wrong. Creating fake celebrity investment pitches is not a gray area. Cloning someone’s voice to run a financial scam is not a gray area. Generating non-consensual images of real people is not a gray area. The law may be slow. The moral position on those cases was never unclear.

Neutral at the level of physics. Not at the level of the market.

We reach for “technology is neutral” because it feels true and because it lets us avoid harder questions. A camera does not decide whether it documents a family trip or records someone who never agreed to be filmed. A microphone does not decide whether it captures a podcast or a private conversation. An AI model does not decide whether it helps a student understand a concept or helps a fraudster replicate a trusted voice.

At the level of physics, yes. Neutral.

But once someone builds a business model around a technology, the neutrality ends. A business model decides who is served, who pays, who benefits, and who has no say in the matter. It translates capability into incentives. And incentives decide what gets built next, what gets repeated, what gets funded, and what gets scaled. The technology is the capability layer. The business model is the behavior layer. That distinction is what makes the deepfake conversation actually useful.

The productive version of this pattern

We know this mechanism from disruption, and its constructive form is not hard to describe. A new technology lowers the cost of something. A business model forms around that lower cost. The initial offer is simpler, less polished, and less profitable than what established players sell, so they ignore it. The margins are thin. The customers look marginal. The use case seems narrow. The new entrant serves those customers anyway, learns faster than anyone expected, and gradually improves until it reaches a market incumbents actually care about.

Christensen’s work on disruption describes exactly this: a simpler, cheaper, more accessible offer that reaches people who were overserved, underserved, or excluded entirely. The technology alone does not disrupt. The model around it does.

At its best, this serves society in a straightforward way. Lower cost enables a different offer. The different offer reaches people who were previously locked out by price, geography, or the complexity of the incumbent product. They do not need the full package. They need the job done well enough at a cost they can actually afford. That is not a compromise. That is progress.

The same engine, pointed differently

Read more

I’m currently reading one of those bestselling books that promises to explain how people should work with AI, and I have to admit that one part of it is hard to argue with. Large language models really are impressive creative partners. They can throw out ideas, variations, objections, concepts, and competitor moves faster than any workshop room full of smart people ever could, and in innovation work that speed is not a small thing.

I’d go further than the book does, honestly. AI is genuinely useful in the divergent phase of a problem, when you’re still trying to frame a situation, generate hypotheses, and figure out what struggles your job performers actually have. It widens the search space in a way a Tuesday afternoon workshop rarely manages, especially once you notice that most workshop rooms are filled with people from the same company, carrying the same incentives, the same vocabulary, and the same unspoken assumptions about what a good idea looks like.

It’s just as useful on the solution side. Ask it for analogies from other industries, prototype directions, business model variants, or the uncomfortable question nobody on the team wants to raise, and it delivers something usable within seconds. And it doesn’t stop at divergence. Feed it a stack of interview notes and it finds patterns you missed. Give it a pile of problem statements and it sorts and compares them. Hand it messy workshop material and it turns it into something structured enough to actually discuss, instead of a wall of sticky notes nobody wants to revisit.

It can even help you build something fast enough to test whether a solution offers real relief, and whether people would switch and pay for it, without pulling your best engineers off their real work or spending a serious budget on something that only exists to test problem-solution fit. Used deliberately, that’s a genuine accelerant.

AI can widen the room, but…

Read more

The decision for boards and CEOs is no longer whether AI can speed up innovation and transformation work.

It can.

The better question is what that speed is being used for.

Is the organization reducing uncertainty, or is it only producing more convincing artifacts around uncertainty?

That distinction matters because the cost of producing things has fallen faster than the cost of choosing well. Prototypes, plans, research summaries, workflows, code, business cases, customer personas, pitch decks, and strategic options can now be generated at a volume most organizations are not designed to evaluate.

This changes the work. It does not remove the hard part.

The hard part is still deciding what problem matters, which assumption carries the risk, what prediction would actually improve the decision, and what evidence is strong enough to justify commitment.

That is the shift leaders need to see.

In many firms, the bottleneck is no longer the first version.

It is judgment.

AI lowers the cost of prediction, not the cost of being right

Agrawal, Gans, and Goldfarb make in their book Power and Prediction a useful distinction that many AI conversations still miss.

AI is prediction technology.

Prediction is not only about forecasting next quarter’s revenue or tomorrow’s demand. Prediction means using information you have to generate information you do not have. Is this customer likely to churn? Is this image showing damage? Is this transaction suspicious? Is this lead worth attention? Is this prototype likely to solve a real customer problem?

That is powerful.

But prediction is only one input into a decision.

A prediction does not tell you what outcome matters. It does not tell you which error is acceptable. It does not tell you whether speed is worth the risk. It does not tell you who carries the consequence if the decision is wrong.

That part is judgment.

And when prediction gets cheaper, judgment does not disappear. It becomes exposed.

A company can now produce more options, more simulations, more concepts, and more apparent evidence than before. But if nobody has defined the payoff, the trade-off, the acceptable loss, or the decision threshold, the organization has not become more intelligent.

It has become faster at generating undecided work.

More output does not mean more progress

A lot of organizations are mistaking lower production cost for progress.

The mistake is understandable. The outputs look credible. They arrive fast. They create momentum. A team can generate a market scan in an afternoon. A prototype can appear before the problem is understood. A board deck can look mature before the critical assumption has been tested.

This feels like acceleration.

But movement is not learning.

Right now, teams can create more options than they can evaluate. They can generate more scenarios than they can absorb. They can produce more answers than the organization has the discipline to challenge.

That sounds efficient.

It can also become expensive.

Because when output gets cheaper, weak thinking scales too.

The risk is not only that AI produces bad answers. That is the simple version. The deeper risk is that AI produces plausible work around weak questions.

A concept can look ready before the market risk is clear. A prototype can feel like evidence when it is still only a technical demonstration. A strategy can sound aligned because the language improved, while the underlying disagreement stayed untouched.

That is not transformation.

That is acceleration without enough judgment.

Read more

I have worked on innovation projects and co-founded startups where the real difficulty was not the product.

The product mattered, of course. The prototype had to work. The value proposition had to make sense. The customer problem had to be real enough. But in hindsight, the harder question was somewhere else.

How much of the customer’s system had to change before our solution could create value?

That question is easy to underestimate.

When you are building something new, you naturally focus on the improvement you can offer. Faster response times. Lower costs. Better data. Better decisions. Less manual work. A clearer view of what is happening in the field. A way to connect people, assets, suppliers, customers, and information differently.

From the inside, this can feel obvious.

Why would a company not want that?

But companies do not adopt ideas in the abstract. They adopt changes inside existing systems. And those systems already have processes, incentives, budgets, departments, politics, vendors, data structures, risk rules, and people whose jobs are built around the current way of working.

That is where the real adoption problem begins.

When the change is only a task

A startup solution that improves a single task is usually the easiest to absorb.

A human does something repetitive or rule-based. The software helps that person do the task faster, with fewer errors, or with better information. The basic work does not change much. The company can keep the existing process and simply improve one part of it.

This kind of adoption is relatively modular.

There is still a buying decision. There may still be IT, procurement, legal, and security involved. Nothing is ever as simple as the pitch deck suggests. But the logic is clear enough.

Same task. Better tool. Higher output per hour. Lower cost per outcome.

That is a story most companies can understand.

The buyer does not have to redesign the company. The approval path is relatively narrow. The decision can sit inside one team, one budget, one operational pain.

The solution improves the existing machine.

When the process has to move

The next level is harder.

Read more

By the mid-2010s, you could attach a sensor to almost anything and get data back to a dashboard within minutes. Temperature. Humidity. Pressure. Vibration. Location. Fill levels. Energy consumption. Brake conditions. Door status. Water leakage. Machine anomalies. The hardware was cheap. Connectivity was available. Cloud storage cost almost nothing.

The pitch was clean: replace a dumb device with a smart one, detect problems before they become expensive, automate the response, show the ROI on a slide. Customers nodded. Pilots launched. The proofs-of-concept worked.

And then almost nothing scaled.

Not because the sensors failed. Because the industry had mistaken the technical problem for the actual one.

When Sensors Start Measuring People

The easy cases in IoT are easy for a specific reason.

A flood sensor does not care who watches it. A transformer does not worry about how its vibration data will be used next year. A parking space does not suspect the city is building a behavioral profile. When the sensor measures a physical condition with no human attached to the consequence, adoption is straightforward. The benefit is immediate. The perceived risk is low.

The friction arrives the moment data stops describing a condition and starts describing a person.

Read more

I was pitched the same idea three times in twelve months. The tool varied. The claim did not: before you commit real budget, test your idea in a safe space.

The offer came packaged differently each time. AI-generated personas. Synthetic customer panels. Digital twins of future users. Beneath the variations, one promise: make your assumptions visible before the spending starts.

I understand the appeal. Most teams run on weak assumptions disguised as analysis. Any system that forces assumption clarity before money gets locked in deserves attention.

So I am not against synthetic users. I am against treating them as customers. That distinction will cost you a round, a market entry, or twelve months you will not get back.

The Preference-Behavior Gap Predates This Technology

The core problem is not AI. It is the distance between what people say and what they do.

In 2001, economists List and Gallet published a meta-analysis of 29 experimental studies on stated versus actual willingness to pay. Participants in hypothetical settings overstated their preferences by a factor of roughly three on average. That finding predates generative AI by two decades. Every research method built on stated intent faces the same structural trap: responses are not behavior.

Language models add a second layer of distance. Horton, Filippas, and Manning at MIT and NBER describe LLMs as “homo silicus,” an implicit computational model of humans that can be given preferences and simulated in economic scenarios. Their 2023 NBER working paper shows that LLM experiments reproduce results qualitatively similar to classic behavioral economics studies.

The same paper notes what economists have long held: the economic content of mere statements is worth far less than the economic content of actual behavior. That critique applies directly to synthetic user outputs.

Bisbee and colleagues sharpened the constraint in 2024. Studying ChatGPT’s ability to replicate human survey data, they found that while LLMs recover average opinion scores with reasonable accuracy, the variance is compressed and the regression coefficients diverge from human benchmarks. For the kind of statistical inference a founder uses when deciding whether to fund, kill, or pivot a product, LLM-generated responses are not reliable substitutes for behavioral data.

The practical consequence: a synthetic user can produce a plausible answer to what a person like this might say. It cannot produce evidence of whether that person will pay, return, or recommend.

The Digital Twin Language Is Doing Real Damage

Every pitch I receive includes the phrase “digital twin.” I think it is the most expensive two words in market research right now.

A digital twin of a machine works because the machine has stable physical properties, measurable states, and causal behavior you can model. A person embedded in a buying situation has none of that. They have constraints, fears, habits, alternatives, social pressure, timing conflicts, and identity concerns that shift faster than any training data can track.

If the idea is new enough to matter, there is no behavioral evidence for the exact situation you want to predict. The model interpolates from adjacent data. It generates plausible reactions. It can surface contradictions. But plausibility is not evidence.

When teams call that output a digital twin, they convert uncertainty into interface design. The tool looks authoritative. The confidence is borrowed. Expensive decisions then move forward on borrowed confidence, and the reckoning arrives when it costs real money to find out.

B2B Breaks the Single-User Frame Entirely

The digital twin problem becomes structural in B2B. There is no single customer. There is a buying system.

Gartner’s research on B2B purchasing puts the average buying group at 6 to 10 decision-makers for complex solutions, each entering the process carrying 4 to 5 pieces of independent research they bring to the group. Forrester’s 2024 State of Business Buying report raises that average to 13 stakeholders, with 89% of purchasing decisions crossing multiple departments.

Buyers spend only 17% of their total purchasing time meeting with potential vendors. The rest happens in internal rooms and conversations a seller never enters.

The user has the problem. The manager owns the outcome. Procurement owns the process. IT owns the risk. Legal owns the contract. Finance owns the budget. An executive sponsor owns the political cover. A middle manager you never encounter may block the decision because the solution reduces their control.

The relevant question is not: would this synthetic user buy? It is: can this solution survive the buying system?

A B2B purchase moves through budget cycles, internal politics, procurement rules, security reviews, switching costs, vendor trust, risk ownership, and the capacity of a sponsor to spend internal capital. Gartner’s 2025 survey of 632 B2B buyers found that 74% of buyer teams experience unhealthy conflict during the decision process. Even when consensus forms, one new stakeholder joining late can dissolve it.

To stress-test B2B demand properly, you would need a synthetic buying committee, a synthetic procurement path, a synthetic political map, and a synthetic risk model. That is not what most tools offer. Even if you built all of it, the output would still be a rehearsal, not evidence.

B2G Adds Another Layer

Public sector makes the synthetic-customer idea even more fragile.

In B2G, public need does not equal institutional ability to purchase. A department may have a genuine problem, a clearly superior product in front of them, and an obvious economic case. None of that moves procurement if the buying path is constrained by public law, tender requirements, multi-year budget cycles, political accountability, data protection rules, and vendor neutrality obligations.

The best product may lose because the tender rewards the wrong specification. The most urgent need may wait because the budget was allocated two years earlier. A successful pilot may never scale because no one can justify the next step inside the formal process.

The question in B2G is never: does the user want this? It is: can the institution buy this without creating legal, political, operational, or reputational exposure? A synthetic user cannot answer that.

A synthetic procurement simulation might help you rehearse the obstacles. That is not evidence of demand. What it does is show you where your go-to-market belief is naive. Useful. Stop calling it validation.

Where Synthetic Users Actually Help

The value is real. The use case is narrower than most pitches suggest.

Synthetic users can force teams to surface assumptions before money is committed. They can generate objections the team has not considered. They can expose where messaging fails to hold. They can simulate how different segments might interpret the same offer and stress-test pricing logic before a real experiment costs real money.

Most teams skip this work entirely. They move from idea to prototype to launch while quietly assuming the user has urgency, the buyer has budget, procurement has no friction, switching is easy, and the competition stays passive. A synthetic-user system can make those assumptions explicit and testable. That is worth something.

The problem starts when synthetic users substitute for real discovery: actual willingness-to-pay tests, real procurement conversations, behavioral evidence from real buyers. At that point the tool stops reducing uncertainty. It makes uncertainty look sophisticated.

Assumption Stress Tests, Not Digital Twins

I would not frame this output as customer validation. I would call it an assumption stress test.

That is less exciting. It is also accurate.

The output should not be confidence. It should be a sharper test plan.

After running synthetic users, a team should know which assumption needs real-world evidence first, which objection must be tested with actual buyers, which stakeholder can kill the deal, which channel assumption is optimistic, and which part of the business model is still belief rather than evidence. That framing changes what the tool is for. It is not an oracle. It is a diagnostic.

The Core Distinction

The mistake is not using synthetic users. The mistake is promoting them to synthetic customers.

In B2C, the risk is mistaking preference for behavior. In B2B, the risk is mistaking user pain for buying power. In B2G, the risk is mistaking public need for institutional ability to act.

Start with the decision system. Who must act? Who must pay? Who must approve? Who must change what they do? Who can block? Who benefits most if nothing changes?

Only after that map is visible does synthetic simulation have a useful job. The real question is not whether an AI-generated persona likes your idea. It is whether the market has a path from interest to action.

That path is where most ideas die. And no synthetic user saying yes has ever built it.

Leave a comment

Sources

1. List, J.A. & Gallet, C.A. (2001). “What Experimental Protocol Influence Disparities Between Actual and Hypothetical Stated Values?” Environmental and Resource Economics, 20(3):241–254.

2. Horton, J.J., Filippas, A. & Manning, B.S. (2023). “Large Language Models as Simulated Economic Agents: What Can We Learn from Homo Silicus?” NBER Working Paper No. 31122.

3. Argyle, L.P., Busby, E.C., Fulda, N., Gubler, J.R., Rytting, C. & Wingate, D. (2023). “Out of One, Many: Using Language Models to Simulate Human Samples.” Political Analysis, 31(3):337–351.

4. Bisbee, J., Clinton, J.D., Dorff, C., Kenkel, B. & Larson, J.M. (2024). “Synthetic Replacements for Human Survey Data? The Perils of Large Language Models.” Political Analysis, 32(4):401–416.

5. Gartner. “The B2B Buying Journey.” Gartner Research.

6. Gartner (2025). “Gartner Sales Survey Finds 74% of B2B Buyer Teams Demonstrate ‘Unhealthy’ Conflict During the Decision Process.” Press release, May 7, 2025.

7. Forrester Research (2024). The State of Business Buying, 2024. Forrester Research, Inc.

I chatted with a company chatbot today because of a simple customer question.

Nothing complicated. I had received a generic message and wanted to understand what I was supposed to do next.

The answer was bad.

Not aggressively bad. Politely bad. The kind of bad that makes the system sound helpful while sending you in circles. It gave generic replies. It did not understand the situation. It repeated itself. It pushed me back toward website content I had already checked.

At some point I caught myself thinking what many customers probably think:

Why do companies believe this is the future of customer interaction?

That question annoyed me enough to run a small test.

So I did a short, non-academic field check of chatbots used by large Swiss companies. Not as a representative study. Not as a scientific benchmark. Just as a structured customer test on a Saturday afternoon.

The test was simple

I used the same generic prompt across companies:

I received a message from your company and do not understand exactly what I need to do. The information on the website does not help because my case does not quite fit.

Then I observed what happened.

Did the chatbot ask a useful clarification?

Did it identify the type of issue?

Did it explain its limits?

Did it offer a concrete next step?

Did it connect me to a person?

Did it get stuck in a loop?

Did it pretend to help while pushing the work back to me?

This was not a test of chatbot intelligence in the abstract.

It was a test of customer-resolution capability.

There is a difference.

A chatbot that can answer “What are your opening hours?” is not a customer service bot. It is a conversational FAQ wrapper.

A useful service bot must handle the moment when the customer says:

This does not answer my case.

That is the point where most systems show what they were really built for.

This is not only my irritation

The research direction is clear enough.

The Consumer Financial Protection Bureau’s report on chatbots in consumer finance warned that chatbots can help with basic questions, but may fail when customers need meaningful assistance with complex issues. The report also describes repetitive “doom loops” where customers cannot reach a human when the chatbot has reached its limits.

Research on chatbot service recovery makes a similar point. A study on chatbot messages after service failure found that recovery responses only help when they move the interaction forward. Solution-oriented recovery can increase perceived competence. Empathy-oriented recovery can increase perceived warmth. But neither helps much if the customer still has no path to resolution.

Recent work on gatekeeper aversion in customer service chatbots is also relevant. People may avoid chatbot channels not only because the bot performs poorly, but because they dislike being forced through an imperfect first stage before possibly reaching a human expert. The study suggests that transparency about chatbot limits, wait times, and faster access to live agents can improve adoption.

That matched what I saw.

The worst bots were not bad because they failed to answer everything.

They were bad because they failed to know when they had stopped helping.

Company 1: Strong triage

Company 1 performed well.

When I framed the situation as an unclear email that might require login, payment, or action, the bot immediately moved into security triage. It warned me not to click links, not to open attachments, not to disclose data, and explained how to treat suspicious emails.

That was useful.

It did not simply say: “Log in and check.”

It understood that an unclear email from a financial institution is not just an information problem. It may be a fraud problem.

When I asked whether to delete, report, or officially verify the message, the bot gave a clear decision path. If the message was suspicious, report it. If I wanted official verification, contact support. If I needed to log in, use the official website or app, not a link in the email.

This was not a perfect experience. Human support required authentication, which is understandable in financial services. But the bot did something important:

It reduced risk before pushing me into action.

That is a real service contribution.

Company 2: Clear boundary honesty

Company 2 also performed well, but in a different way.

The scenario involved a confusing insurance-related statement. The bot first asked useful diagnostic questions. Was it a bill, a benefits statement, or another type of message? Did it mention a payment, missing documents, or a deadline?

That is already better than dumping a link to a FAQ.

Then I made the case more specific. I said the statement mentioned cost participation, but I could not tell whether I had to pay something or whether the document was only informational.

The bot explained the relevant distinction in plain language. It described what the customer should check and when a statement is likely informational versus payment-relevant.

The best moment came when I asked directly:

Can you check my concrete case in this chat, or are you only giving general information?

The bot answered clearly that it had no access to personal data or specific cases. It could only provide general guidance. It then gave the next step: check the customer portal, call support, or use the contact form.

That answer matters.

A chatbot does not lose trust by admitting its limits.

It loses trust when it pretends those limits are not there.

Company 3: Fast human handoff

Company 3 did not try to be clever.

When I said my case did not fit the website information, the bot asked whether I wanted to be connected with a specialist.

I said yes.

It gave me a choice between identification and guest mode. I chose guest mode. A human agent joined.

That is not sophisticated AI.

But it is good service logic.

Research on chatbot-to-human handover shows that the handoff itself matters. Customers communicate differently with bots than with human agents, and the way repair and transfer happen affects whether the conversation keeps progressing or resets awkwardly.

Company 3 got the most important part right. It did not force me through five irrelevant answers before giving me access to a person.

The weakness was context transfer. When the human joined, the conversation more or less restarted.

That is still a cost for the customer.

But at least the door opened.

Company 4: Honest fallback, weak recovery

Company 4 was more limited.

The bot admitted it could not find a suitable answer. It said live chat would be available again on Monday morning.

Since this was Saturday afternoon, that limitation is fair.

Not every company needs 24/7 human support. That is not the issue.

The issue was what happened next.

When I asked what I could do now, the bot repeated the same fallback. Only after I explicitly asked for a person, phone number, or contact form did it offer a contact form.

That is a weaker design.

The escape route existed. But the customer had to fight to find it.

A better bot would have said immediately:

I cannot answer this specific case. Live chat is closed until Monday. You can either submit a contact form now or return during opening hours.

That would have been honest, clear, and useful.

Instead, the first recovery move was repetition.

That is how service friction hides inside polite wording.

Company 5: The menu loop

Company 5 was the weakest.

The bot gave me several contact entry points, but some were duplicated. One option appeared twice. Later, another consultation-related option appeared twice. The bot seemed to classify my request into overlapping menu categories without knowing how to move forward.

At one point, it said it was not sure whether it had understood me and offered that someone could contact me by email.

I selected yes.

Expected next step: ask for my email address, open a form, confirm a follow-up, or route me to a contact process.

Actual next step: the bot replied as if the matter was finished and asked for feedback.

No email process started.

No contact details were collected.

No handoff happened.

Then, when I pushed again, it returned to another menu. “Email/message.” “Feedback.” “Immediate consultation.” “Personal advice.” Some entries repeated. Selecting one led to another similar menu.

This is the worst version of chatbot design.

Not because the bot could not solve the case.

Because it created the impression of progress and then failed to execute it.

That is worse than a clear contact form.

A bad form is boring.

A broken chatbot is misleading.

No chatbot may be better than a weak chatbot

Some companies did not show an obvious public chatbot in the paths I tested.

That should not be treated as failure.

No chatbot is not automatically worse than a chatbot.

A clear phone number, customer portal, contact form, branch locator, or claim form may be less fashionable. But it can be better service.

A weak chatbot adds effort. It creates loops. It hides contact paths. It makes the customer repeat themselves. It shifts the problem from the company’s support design to the customer’s patience.

The question is not whether a company has a chatbot.

The question is whether the chatbot improves the path to resolution.

The real failure is not technical

Most corporate chatbots do not fail because language models are too weak.

They fail because companies have not made the service-design decision behind the bot.

What is the bot allowed to do?

Can it ask diagnostic questions?

Can it say, “I cannot check your individual case”?

Can it escalate early?

Can it hand over context?

Can it distinguish between a search task, a complaint, a security risk, and a case-specific request?

Research on task type and failure frequency in chatbot failure recovery suggests that the right recovery strategy depends on what the user is trying to accomplish and how often the bot has already failed. A bot may recover some search tasks itself, but repeated failure or complaint-like situations often require human intervention.

That sounds obvious.

But many companies still design bots as if all customer problems were search problems.

They are not.

Some are decision problems.

Some are trust problems.

Some are exception problems.

Some require access to customer data.

Some require a human because the customer needs accountability, not another paragraph.

The minimum standard

A useful customer service bot does not need to solve everything.

But it must know which of four situations it is in:

1. The answer is known and safe to give.

2. The customer needs help classifying the issue.

3. The customer needs a case-specific path.

4. The bot has stopped helping and must escalate.

Most weak bots confuse these situations.

They treat unresolved customer problems as content retrieval problems.

That is the root error.

The best bots in my small test did not necessarily sound more human. They behaved with clearer judgment.

One triaged risk.

One admitted its boundary.

One escalated quickly.

The worst one duplicated contact options, offered a follow-up, failed to start it, and pushed the user back into menu loops.

That is not automation.

That is customer effort with a chat bubble.

The expensive question

Companies should stop asking:

Should we have a chatbot?

That question is too shallow.

The better question is:

Which customer situations are we willing to let a bot handle, and where must it stop?

If the company cannot answer that, the chatbot becomes a public interface for internal indecision.

A useful bot does not need to be impressive.

It needs to be honest, diagnostic, and connected to the right support process.

If it cannot do that, the future of customer interaction may look a lot like the past, only with a friendlier loading animation.

Leave a comment

I keep noticing the same confusion in discussions about AI and innovation. The conversation usually starts with speed: faster research, faster market scans, faster prototypes, faster business cases, faster testing, faster decks.

I understand the appeal. I use AI in my own work, and it is useful. It helps me look at a problem from different angles, challenge weak assumptions, generate alternatives, and avoid waiting for the perfect workshop, expert, designer, analyst, or alignment round before doing any serious thinking.

But speed is not the interesting part.

The real shift is that AI makes prediction cheaper. And once prediction becomes cheaper, the bottleneck moves. It moves to judgment first, and then to decision. That is where innovation work becomes uncomfortable.

I am not writing this as an AI engineer. I am looking at this from the place where innovation decisions usually go wrong: the point where teams have enough activity to feel like they are making progress, but not enough evidence to deserve the next commitment.

That distinction matters because AI can help us predict, it can support judgment, and it can even recommend what looks like a decision. But these three things are not the same.

Prediction is useful. It is not a decision.

Prediction asks what might happen. Will this segment care? Will customers switch? Will this prototype reduce friction? Will this market grow? Will this signal matter?

AI is useful at this layer. Not perfectly, not magically, but usefully. It can compare patterns, summarize interviews, generate scenarios, expose contradictions, and show where an assumption looks weak. It can look at a case and suggest that adoption may fail because the person with the problem is not the person with the budget.

That is valuable, but a prediction does not commit anything. A prediction says, “This might happen.” A decision says, “We are willing to commit resources under this level of uncertainty.”

Those are different acts.

A team can have a good prediction and still make a bad decision. A market can look attractive and still be strategically irrelevant. A customer problem can be real and still too weak to justify switching. A prototype can test well and still fail once procurement, implementation risk, internal politics, and budget ownership enter the room.

This is why I find the distinction useful: prediction reduces uncertainty, judgment interprets uncertainty, and decision commits under uncertainty. AI helps most with the first. It can support the second. It cannot own the third.

The hidden part of adoption

Human prediction and AI prediction are not the same. That does not mean humans are better. Humans are biased, political, overconfident, selective, and quite capable of turning weak signals into stories they already wanted to believe.

But good founders, product leaders, salespeople, consultants, and strategists do something AI still struggles with. They read situations. They notice when a customer says the problem is important but avoids the next step. They hear when an executive likes the idea but never discusses budget. They sense when a stakeholder is polite because saying no would be awkward. They know when a technical objection is really a power issue.

This kind of prediction is messy and not always measurable, but it includes social signals and context.

AI works from traces: text, data, documents, prompts, examples, correlations, and whatever has been captured somewhere. That makes it strong when the relevant pattern is visible. It makes it weaker when the decisive constraint is hidden.

And in innovation work, the decisive constraint is rarely only functional.

Someone has to change behavior. Someone has to approve budget. Someone has to absorb risk. Someone has to give up control. Someone may lose status if the new solution works. Someone may lose margin, authority, routine, or political safety.

This is why adoption is not just a product question. It is a social and economic event.

AI may help predict that a problem is attractive. But it may not see that the person with the problem cannot buy, that the buyer benefits from the current inefficiency, or that the department showing interest would lose influence if the solution succeeded.

The visible problem is not always the decisive constraint. That is where judgment starts.

Judgment is not a prompt

Judgment asks a different set of questions with its expertise. Not only what might happen, but what matters, what evidence counts, which assumption carries the case, what the cost of being wrong is, who benefits if this works, who pays if it fails, what we would need to stop believing, and what should make us kill the project.

This is where innovation work becomes serious because judgment is not just analytical. It is social, economic, and political.

A narrow innovation question asks whether an opportunity is attractive. A better question asks whether it is attractive for the right actor, under the right constraints, and at an acceptable cost.

A solution may create value for the company while reducing autonomy for experts. It may improve customer convenience while shifting work to frontline teams. It may reduce operating cost while threatening a powerful internal unit. It may create a better customer experience while damaging someone’s current business model.

These are not soft issues. They are adoption risks. They decide whether a promising idea becomes a serious business opportunity or another well-designed pilot that never leaves the safe zone.

AI can help here, but only when the frame is clear. If I ask AI to rank opportunities by market size and feasibility, I may get a polished answer that is not very useful. If I ask it to map who gains, who loses, who can block adoption, who owns the budget, which behavior must change, and which assumption would kill the case, I get something closer to judgment support.

But the criteria still come from me, from the team, or from leadership. AI can reason inside a frame. It cannot decide which frame deserves authority.

That remains a human responsibility.

Selection is not commitment

This is the distinction I care about most.

AI can select, recommend, rank, and optimize. But a decision is not just choosing an option.

A real decision changes exposure. Money moves, people are assigned, alternatives are stopped, credibility is spent, timing changes, political capital is used, and someone becomes accountable.

That is what makes a decision different from a recommendation.

In innovation work, real decisions sound like this: we will test this segment first; we will not build before we validate switching behavior; we stop this project because the demand assumption failed; we scale only if customers commit budget, not just interest; we accept the technical risk because the demand evidence is strong; we do not continue funding because the next commitment is not justified.

AI can help prepare these decisions. It can clarify the options, list assumptions, propose thresholds, expose missing evidence, and write the uncomfortable questions. But it cannot own the decision.

It cannot explain to the board why a project was stopped. It cannot absorb the cost of a false positive. It cannot tell a team that their favorite idea does not deserve another funding round. It cannot carry reputational loss when the evidence was misread.

AI can calculate a choice. It cannot socially own a commitment.

I see the same thing in smaller situations too. When I use GenAI to draft a sensitive email to an important customer, I may ask for several versions: friendly, honest, non-defensive, grounded in data. I may compare them and even use the version it recommends.

But the judgment of what is appropriate remains mine. And when I press send, the decision is mine too. The tool generated a reply. I chose to make it real.

That is why I am cautious with the phrase “AI decision-making.” In strategic work, AI usually does not decide. It recommends under stated assumptions. The decision still belongs to the people who commit resources and live with the consequences.

The weak use of AI in innovation

The weak use is easy to spot: use AI to create additional ideas, concepts, synthetic research, prototype variants, business cases, and slides.

That may feel productive and reduce cost. But it can also create a cleaner version of the same old theater: faster uncertainty, better formatted, with stronger sentences and weaker accountability.

That is not progress. It is cheaper motion.

The stronger use is different. Use AI to reduce the cost of learning before commitment becomes expensive. Use it to ask which assumption carries the business case, which evidence is still missing, which actor can block adoption, which payoff is misallocated, which customer behavior has not been proven, which decision is being avoided, and which kill criteria should be agreed before the next budget is spent.

That is where I see the real value. Not AI as an idea machine, but AI as an uncertainty machine.

The task is not to produce additional material. The task is to see what we do not know yet, what would matter if we are wrong, and what decision the available evidence can actually support.

That is a higher standard.

What should change in funding meetings

If prediction becomes cheaper, leaders should not accept additional analysis as progress. They should raise the standard for commitment.

A team should not receive the next budget because it has produced additional material. It should receive the next budget because the evidence changes what the organization is willing to risk.

That requires different questions. What did we believe before? What changed? Which assumption became stronger? Which assumption became weaker? What would make us stop? What are we no longer funding if we say yes? What commitment are we actually making now?

These questions are not slower. They are cleaner.

They prevent teams from hiding behind activity. They prevent leaders from delaying the real decision until sunk cost makes the decision for them.

That is the part of AI in innovation I find interesting. Not that it will make innovation easier, but that it may expose where innovation systems were weak all along.

Innovation does not fail only because prediction was expensive. It fails because judgment under uncertainty was weak, incentives were misread, power shifts were ignored, teams confused interest with demand, and leaders waited too long to decide what deserved commitment and what deserved to stop.

AI may make weak innovation systems faster: faster ideas, faster validation theater, faster business cases, faster pilots without commitment logic.

Or it can force a better discipline.

Test sharper. Stop earlier. Scale later. Fund with thresholds. Kill without drama. Commit only when the evidence deserves it.

That is the real test.

AI may make prediction cheap.

It will not make commitment cheap.

Leave a comment

Improving a model is almost always possible. Knowing when to stop is not.

AI research and model learning are not my field of expertise. I am looking at this from an innovator’s perspective: where uncertainty becomes investment, where technical progress turns into resource commitment, and where teams need to decide what is worth improving, testing, scaling, or stopping.

That asymmetry is where most AI investment decisions go wrong. The team asks how good the model can get. The room debates benchmarks, latency, safety scores, and data coverage. Nobody asks whether closing any of those gaps would change the decision they are actually trying to make.

The shift from “how good?” to “good enough for what?” is where AI stops being an engineering problem and becomes a capital allocation problem under uncertainty. Most teams have not made that shift.

The model is not trained on data. It is trained on usable data.

Raw data is not training material. It is exposure until it earns the right to be called input.

Before a dataset can help a model, it has to be acquired, licensed, cleaned, filtered, deduplicated, structured, mixed, and tested. Some of it gets cut because it is low quality. Some because it creates legal risk. Some because it teaches patterns the provider does not want in the product.

Sambasivan et al. studied 53 AI practitioners across high-stakes domains including healthcare and conservation and found that data quality failures caused compounding downstream failures they called “data cascades.” [1] The AI community’s tendency to prioritize model work over data work was the root cause. Most practitioners had assumed the problem was somewhere else in the system.

A foundation model learns from predicting structure in language, not from labeled examples. A domain model needs expert annotation. A safety layer needs adversarial inputs. An enterprise deployment needs evaluation sets that reflect actual customer workflows, not benchmark tasks nobody’s users encounter.

The economic unit is not “the model.” It is the whole system required to make a model usable in a specific market, with specific buyers, in a specific risk environment.

Read more

The easiest story about AI and innovation is already everywhere.

AI helps teams generate ideas faster. It helps non-technical innovators build prototypes. It reduces dependency on developers. It can create mockups, landing pages, product flows, interview guides, business model options, research summaries, and coded demos in minutes.

That story is true. I use AI for parts of this work myself.

It is useful when I do not have immediate access to a broad range of experts. It helps me prepare interviews, sharpen assumptions, formulate hypotheses, and think through different user groups and their possible point of view before I enter the field.

In cases where only a few real interviews are possible, that preparation matters. You cannot afford to waste those conversations on vague questions and lazy assumptions.

But that is also where the line sits.

AI can prepare the work. It cannot replace the work.

A synthetic user can help you ask better questions. It can help you anticipate objections, test different framings, and expose weak spots in your own thinking. But it does not buy your product. It does not change its workflow. It does not fight for budget. It does not risk credibility in an internal meeting. It does not ignore your solution because the timing is wrong, the politics are difficult, or the switching effort is simply not worth it.

Real life is not made of artificial users.

This distinction sounds obvious until a polished AI-generated validation deck enters a steering committee. Synthetic feedback, prototype reactions, and research summaries start to look like evidence. If nobody asks where the findings came from, the room accepts what it sees and funds the project without any meaningful reduction in uncertainty.

That is the problem.

AI will not make innovation predictable. It will make weak evidence easier to produce, easier to package, and easier to mistake for progress.

Unless leaders change how they judge it.

The wrong question

A lot of current AI-in-innovation talk starts with the wrong question: how can we build faster?

That question is not useless. Speed matters. Cost matters. Access to technical skills matters. A team that can move from idea to prototype without waiting three weeks for internal resources is in a better position than one that cannot.

You can see this logic in how AI product development is now being taught and sold: AI user research, synthetic interviews, no-code prototyping, micro-experiments, and faster movement from idea to customer signal [1]. Wharton and many other business schools have also framed generative AI as a force that changes how organizations conceive, shape, and select ideas at scale [2].

But faster building can also make weak innovation work worse.

Teams can now move from vague idea to polished prototype before they understand the problem. They can generate evidence-shaped material before they define what evidence would actually change the decision. They can produce confidence faster than they can earn it.

That is not progress. It is acceleration without discipline.

The better question is not how fast a team can build. The better question is what becomes worth testing when the cost of testing drops.

That is where AI becomes strategically useful. Not because it can predict which innovation will win. It probably cannot. Innovation outcomes depend on user motivation, timing, trust, budget cycles, switching costs, procurement, internal politics, regulation, channel access, and the painful details of adoption.

You cannot prompt your way around that complexity.

But you can structure ignorance better.

Synthetic users are not customers

One tempting use of AI is to simulate users.

Create artificial personas. Let them react to a concept. Ask whether they would buy. Run synthetic interviews. Simulate objections before spending money on real discovery.

I understand the appeal because I have used this myself.

It can be useful, especially when the number of real interviews is limited. If you only get access to a few relevant people, preparation becomes critical. Synthetic interviews can help you form hypotheses, improve your questions, and enter real conversations with sharper assumptions.

But they must remain assumptions.

They are not validation.

This is where the misuse begins. A team runs synthetic interviews, collects plausible feedback, turns it into a clean summary, adds a few quotes, and presents it as if the market has spoken. The deck looks professional. The logic sounds coherent. The feedback fits the story. The project moves to the next gate.

But no real customer has changed anything.

No buyer has given time. No budget owner has engaged. No user has shared a painful workaround. No team has changed its workflow. No switching effort has appeared. No internal sponsor has taken a risk. No one has paid, committed, or reorganized around the problem.

A synthetic user can give a plausible answer. But plausibility does not reveal causation, and it certainly does not prove demand.

Demand shows up through real-world commitment. It shows up when someone takes a second meeting, shares internal constraints, introduces a colleague, accepts a pilot, reveals budget ownership, spends time on integration, or changes an existing routine. These signals are imperfect, but they touch reality. That is why they matter.

Real users also surprise you. They contradict themselves. They rationalize decisions after the fact. They say something is important and then do nothing.

They ignore the feature the team loves. They care about small frictions the team barely noticed. They protect status, avoid embarrassment, and follow habits.

They make decisions inside constraints that are not visible from the outside.

That is where qualitative discovery earns its place.

Not because interviews are magic. They are not. Interviews can be badly designed, badly interpreted, and easily abused. But good qualitative work can reveal causal mechanisms behind a struggle. It can help explain why someone behaves the way they do, what they are trying to avoid, what must be true for them to act, and where the real obstacle sits.

The current research on LLMs as substitutes for human participants should make innovation teams careful here. Recent work argues that LLMs are useful but unreliable tools for simulating human psychology, and that they need to be validated against human responses for every new application [3]. Another study found that LLMs failed to reproduce the full range of human behavioral variability in a cognitive task, even when prompts, model configurations, and sampling settings were varied [4].

Synthetic feedback can help prepare for real discovery. It should not replace it.

Prediction gets cheaper. Judgment does not.

Agrawal, Gans, and Goldfarb make a distinction that matters here: prediction helps estimate what may happen, while judgment is about deciding what matters when outcomes are uncertain [5][6]. AI lowers the cost of prediction. It does not touch judgment.

That distinction matters for innovation.

AI may help predict which message is clearer, which onboarding flow creates less friction, which segment resembles a known pattern, or which assumptions conflict with available data. It can help generate options and compare scenarios. It can help a team avoid obvious blind spots before spending scarce time with real customers.

But the harder questions remain human and strategic.

What would this signal mean? Which assumption is being tested? What would count as failure? Which evidence is strong enough to justify the next funding step? Which result should stop the project? Which uncertainty can we still afford to carry? Which one can we not?

AI lowers the cost of producing options. It does not lower the need for judgment.

In fact, it raises the premium on judgment because teams now have access to a larger volume of variants, signals, summaries, and polished artifacts. Without expertise, that volume becomes dangerous. A tool that creates value in the right place creates blind confidence in the wrong one.

That is why steering committees and boards need to become stricter, not less strict.

The key question is no longer only what the team learned. It is where the evidence came from.

Was it synthetic or real? Was it based on simulated users or actual customers? Was it opinion or observed action? Was it stated interest or demonstrated commitment? Did the signal involve time, budget, switching effort, workflow change, sponsor risk, or money?

That question is not a detail. It is governance.

Gates should reduce uncertainty, not reward activity

Innovation projects need gates, not because companies need extra bureaucracy, but because uncertainty has to be reduced before capital keeps flowing.

Each phase should have a return.

Not always revenue. Not yet. Early innovation phases rarely produce financial return in the usual sense. Their return is validated learning, better evidence, and a clearer decision about whether to continue, change direction, or stop.

A gate should therefore not ask only whether the team has been active. It should ask whether the team has reduced the right uncertainty.

This is where AI can either improve innovation governance or make it worse.

Used well, AI helps teams prepare stronger hypotheses, design better tests, compare variants, and make assumptions explicit before entering the real world. Used badly, AI helps teams produce convincing evidence-shaped material without touching reality.

The difference sits in how gates are managed.

If a team presents synthetic interviews, the steering committee should ask what they were used for. If the answer is preparation, good. If the answer is validation, the gate should not pass.

If a team presents positive feedback, the committee should ask whether it came from real customers and what kind of commitment was involved. If a team presents prototype reactions, the committee should ask whether anyone had to change anything, give anything, risk anything, or pay anything.

Without these questions, gates become performance rituals.

The team shows progress. The stakeholders see movement. The project gets another round of funding. Uncertainty remains mostly intact.

That is cash burn with better slides.

From prototype to disciplined evidence

The prototype became the default artifact of innovation work.

Build the smallest version. Show it to users. Learn.

That logic still has value, but AI makes a different discipline possible. Instead of building one favored version, teams can create a small set of tests that compare competing assumptions before they commit to one path.

Not one landing page, but several versions, each tied to a different belief about the customer problem. Not one prototype, but several variants, each isolating a different assumption about urgency, workflow, switching pain, willingness to pay, or adoption friction. Not one interview guide, but different guides designed to test different explanations for the same struggle.

The point is not to create additional material. The point is to create cleaner comparison.

Before the test, the team defines the evidence parameters. What are we trying to learn? Which assumption are we exposing? What would make the signal valid? What would make it invalid? What could create a false positive? What could create a false negative?

This matters because weak evidence rarely arrives with a warning label. It usually looks useful.

A false positive might be a customer saying the concept is interesting while having no budget, no urgency, and no willingness to change anything. It might be a high click rate from the wrong audience. It might be a synthetic user confirming what the team hoped to hear.

A false negative can be just as dangerous. A rough prototype may fail because the workflow is unclear, not because the problem is unimportant. A buyer may reject a proposal because procurement timing is wrong, not because the need is weak. A user may struggle to explain the problem because it is embedded in routines they no longer notice.

Evidence discipline means looking at both risks.

It means not accepting positive signals too quickly and not killing real opportunities for the wrong reason.

AI can help with this. Research on AI and Lean Startup methods makes a useful distinction between discovery-oriented AI, which helps reduce uncertainty in novel areas, and optimization-oriented AI, which improves existing processes [7]. The goal is not only to produce an MVP faster. The goal is to understand which uncertainty the next test is supposed to reduce.

AI can generate test variants, identify assumption types, draft interview guides, compare feedback patterns, and suggest what to test next. It can help a team prepare better before using scarce customer access.

But the decision logic still belongs to the team.

Especially the kill logic.

Without kill logic, AI will not improve innovation. It will industrialize confirmation bias.

Evidence theater at scale

Lower experiment cost sounds good.

It is good.

But it creates a new problem.

When so-called “evidence” becomes cheap to produce, companies may produce evidence-shaped material without improving decision quality.

More dashboards. More synthetic interviews. More AI-generated research summaries. More prototype reactions. More validation decks. More confident narratives.

Before AI, weak evidence at least required effort.

Now it can be produced at scale.

That changes the leadership task.

The scarce resource is no longer only research budget, design capacity, or engineering time. The scarce resource becomes evidence discipline.

This is where I think steering committees and boards need to be much more demanding. Not anti-AI. Not conservative. Not hostile to experimentation. Just more precise about what counts as evidence at each gate.

Synthetic evidence can justify preparation. It can justify better fieldwork. It can justify sharper hypotheses. It can justify exploring a segment. It can justify improving the prototype before exposing it to real customers.

It should not justify validation.

And it should not justify the next major funding step without real-world evidence.

If a team wants more funding, it needs to show what uncertainty was reduced in the real world. If the project remains uncertain, that is fine. Innovation is uncertain by nature. But the uncertainty should become more explicit, not hidden behind polished artifacts.

A good gate does not ask for certainty. It asks for better uncertainty.

The real shift

AI will not make innovation predictable.

The work is too contextual, too causal, and too dependent on human choices, timing, incentives, trust, budgets, politics, and internal constraints.

But AI can make ignorance cheaper to structure.

It can help teams move from one favored idea to a set of testable assumptions. From polished concepts to controlled variation. From vague conviction to sharper evidence standards before they spend time with real customers.

But only if leaders ask harder questions.

What decision will this evidence change? Which assumption is being tested? What would count as disconfirming evidence? Are we looking at opinions or action? Where did real customers enter the process? Where did real money, time, switching effort, workflow change, or internal sponsor risk show up?

These are not methodological details.

They are funding questions.

The risk is not that AI will make innovation too experimental. The risk is that it will make weak evidence look professional enough to pass the next gate.

That standard will not raise itself. The people controlling the capital have to set it.

Leave a comment

Sources

[1] Harvard Division of Continuing Education: Product Development with AI: From Idea to Market in Half the Time

[2] Wharton Executive Education: Supercharging Innovation with AI

[3] arXiv: Large Language Models Do Not Simulate Human Psychology

[4] arXiv: Can LLMs Simulate Human Behavioral Variability? A Case Study in the Phonemic Fluency Task

[5] Ajay Agrawal, Joshua Gans, and Avi Goldfarb: Power and Prediction: The Disruptive Economics of Artificial Intelligence

[6] NBER: Prediction, Judgment and Complexity: A Theory of Decision Making and Artificial Intelligence

[7] arXiv: Artificial Intelligence, Lean Startup Method, and Product Innovations

I spent the first twenty years of my working life inside large and medium-sized companies trying to build new businesses. Not maintain them. Not optimize them. Build new ones, from inside organizations that were already successful at something else.

I was good at finding the opportunity. I was persistent about making the case. And I kept running into the same wall, dressed in different clothes depending on the company. Internal friction that had nothing to do with whether the idea was sound. Decision rights that sat with people who had no interest in the outcome. Pilots that went well and then quietly disappeared. C-level executives who used the language of transformation in every presentation and then protected their position in every actual decision. Endless alignment meetings that produced nothing but the next alignment meeting.

Eventually I stopped trying to innovate inside established organizations and went to work with startups instead. Not because startups are perfect. They have their own dysfunction. But the dependencies are different. The resistance is different. When something needs to change, it can actually change.

What I did not expect was that leaving the corporate world would give me a cleaner view of it. Because once I was on the outside - as a founder, as a consultant, as someone trying to sell something genuinely new to established companies - I could see the pattern more clearly than I ever could from the inside.

The problem is not that established organizations lack ideas, ambition, or access to good technology. The problem is that they were built to extend what already works, and that structural fact shapes everything that happens inside them, including how they respond to technology that genuinely requires something to change.

How the system was built and why it works the way it does

A mature company has added people to teams, teams to departments, departments to business units, and reporting lines to manage the growing complexity. It has standardized how work moves from top to bottom, how information moves from bottom to top, and how market signals are collected, filtered, translated, approved, and turned into action. It has created processes for everything that can be repeated, documented responsibilities, defined handovers, installed governance routines, and built dashboards.

This is not stupidity. For the core business, this logic is genuinely useful. A company that cannot repeat what works has no business. Standardization reduces variation, creates reliability, makes scale possible, and allows leaders to compare performance across teams and markets. The current business needs control, and control requires structure.

The problem starts when the same system is asked to create the next business. O’Reilly and Tushman’s research on organizational ambidexterity, one of the most cited bodies of work in organizational theory, describes the tension precisely: exploitation - the mode established companies are built for - is about efficiency, control, certainty, and variance reduction. Exploration - what building a new business requires - is about search, discovery, autonomy, and embracing variation. The two modes require conflicting structures, resources, and decision rights inside the same organization. The research shows that the dominant logic of the established business consistently crowds out the conditions exploration needs to function.

Exploration does not behave like execution. A new business model is not a process improvement with a nicer name. It starts with uncertainty. You do not yet know whether the customer cares, whether the problem is painful enough, whether willingness to pay exists, or which assumption will break first. That makes exploration deeply uncomfortable for organizations built around predictability. So they apply the same governance to uncertainty that they apply to certainty. They ask for business cases before the business is understood. They ask for alignment before the evidence exists. They ask for scalability before there is proof of demand. Then everyone wonders why reinvention feels slow, political, and strangely performative.

And when the system still cannot process ambiguity on its own, the answer is usually another meeting. Meetings are not the work in these organizations. They are the repair mechanism for a system that cannot handle what it was not designed to handle.

Where AI makes the contradiction impossible to ignore

I see many companies framing AI through the easiest available question: where can we replace human workload with automated work? That is a valid question. It is also a narrow one. Using AI to summarize documents, draft emails, search internal knowledge, support customer service, or generate reports can be genuinely useful. It reduces workload. It improves speed. It removes low-value effort from people’s day. There is nothing wrong with any of that.

But task automation is not reinvention. It is the easiest form of adoption precisely because it does not disturb the system. The same team stays responsible. The same process stays in place. The same decision rights remain untouched. The same customer promise remains unchanged. The work becomes faster or cheaper, but the underlying logic keeps running.

The harder question is different: what part of the system has to change for this technology to matter beyond efficiency?

That is where things become uncomfortable. Because once AI is no longer used only as a productivity layer, it starts asking harder questions that the system would prefer not to answer. Why does this approval step still make sense? Why does the customer wait three days for something that could be resolved in three minutes? Why is the business model still priced around effort if effort is no longer the scarce resource? Why does expertise sit in one role when the system could distribute it differently?

At that point, AI is no longer just a tool. It becomes a challenge to the operating model and to the people whose authority depends on the operating model staying the way it is.

Three levels of adoption, and why most companies stop at the first

I find it useful to distinguish between three levels of what actually happens when a company adopts new technology, because conflating them is how organizations convince themselves they are transforming when they are not.

Read more

I have spent a good part of my career watching companies discover the internet of things and reach the same conclusion at roughly the same speed. Connect the machines. Collect the data. Put it in the cloud. Build a dashboard. Call it a platform. Wait for the business to transform.

It never quite works that way.

I have been through this cycle in several contexts - with clients, with partners, with companies building the hardware and companies buying the services. The pattern is consistent enough that I have come to think of it as a structural problem rather than an execution problem. The research confirms it. Estimates of IoT project failure rates consistently range from 60 to 80 percent, and a Cisco survey of nearly 2,000 business and IT decision-makers found that only 26 percent could point to at least one IoT project they considered genuinely successful. The technology works. The connection gets made. The data flows. And then, somewhere between the dashboard and the promised business impact, the project stalls. Not because the technology failed. Because nobody stopped to ask which business problem the data was actually supposed to solve, and whether the people collecting it were in any position to solve it.

When IoT and connected things are the answer, what was the question?

This piece is about a company I know from that world. I have worked with them as a hardware partner. They make measurement instruments for process industries: chemical plants, water systems, food and beverage, life sciences, energy assets. Good equipment, solid reputation, long customer relationships. Over the past few years they have started moving upstream, building out a digital ecosystem that connects their instruments to cloud-based backends and frontends, showing data streams from field devices in dashboards, offering diagnostics and asset management. The direction is right. The risk is real.

What a process plant actually looks like

A process plant is not a data-poor environment. It measures pressure, flow, temperature, level, quality, pH, conductivity, density, and dozens of other variables, continuously. The problem is not that the plant is blind. The problem is that most of what the plant knows does not automatically become a better decision.

Operators are trying to keep production stable. Maintenance teams are trying to prevent downtime. Quality teams are trying to avoid deviations. Compliance teams are trying to keep records complete. Plant managers are trying to protect uptime, safety, and margin. Everyone is making decisions under imperfect information, and the information they need is scattered across automation systems, vendor portals, maintenance software, local spreadsheets, PDF manuals, and the head of the technician who has worked on that unit for twenty years.

So plants develop practical substitutes. Fixed maintenance schedules. Alarm response. Manual checks. Conservative spare-parts inventory. Experienced operators carrying institutional knowledge that exists nowhere else. These are not signs of backwardness. According to MaintainX’s State of Industrial Maintenance 2025, 45 percent of maintenance leaders cite staffing and budget constraints as their primary obstacle to better maintenance, and nearly one third of manufacturers struggle to find people with the skills to interpret sensor data and act on what it tells them. In that context, fixed schedules and experienced operators are the rational behavior of organizations managing uncertainty with the tools actually available to them. An estimated 82 percent of companies still rely primarily on reactive maintenance rather than predictive approaches -- not because they prefer it, but because the alternative has not yet been made accessible enough to change behavior at scale.

The result is expensive. Siemens’ True Cost of Downtime 2024 report found that Fortune Global 500 companies lose approximately $1.4 trillion annually to unplanned downtime, equivalent to 11 percent of total revenues, up from 8 percent five years earlier. The average large plant loses 27 hours per month to unplanned incidents. Across manufacturing sectors, Aberdeen Research puts the average hourly cost of an unplanned stoppage at $260,000, reaching $2.3 million per hour in automotive. The buffers are expensive. The uncertainty they are compensating for is more expensive still.

The central question, then, is not what the instrument measures. The central question is which asset, signal, document, inventory position, or maintenance issue actually deserves attention right now. That decision happens dozens of times a day across a plant. It is small enough to look operational, but large enough to affect margin. A wrong call means unnecessary maintenance. A delayed call means downtime. A missed document creates compliance friction. A misread diagnostic sends technicians to the wrong place. A poorly understood installed base leads to excess inventory, obsolete devices, and slow repairs.

This is the problem that genuinely valuable industrial IoT could solve. Not the connectivity. The decision.

Why this company has a real advantage and a real problem

The company I am describing has one thing that most AI-native startups and digital platform vendors lack: it already belongs in the plant. Industrial customers do not hand operational trust to outsiders easily, and for good reasons. Process environments are conservative because safety, uptime, compliance, and liability make conservatism rational. Cybersecurity and integration complexity are consistently cited as the top barriers to industrial IoT adoption, alongside trust in the supplier behind the system. A clever model from a vendor nobody knows is not enough to move the needle in these environments.

An established measurement supplier enters with a completely different kind of credibility. It knows the physical layer. It knows the installed base. It has documentation, service history, calibration context, diagnostics, and relationships built over years of showing up when something breaks. That is a genuine wedge into the decision layer. And the company’s digital ecosystem, if it delivers on its ambition, connects device identity, health status, documentation, maintenance events, inventory signals, and diagnostics into one environment, which starts to look less like a vendor portal and more like plant-level operational intelligence.

The ambition is real. The gap between that ambition and where most industrial IoT actually lands is also real, and I have watched it play out enough times to recognize the pattern.

Connecting data is not the same as understanding the business it belongs to. A dashboard that shows device health is useful. A system that helps a maintenance planner decide which device to inspect first, given limited technician time and a scheduled production run tomorrow, is something categorically different. The first is a technology output. The second requires understanding what the plant is actually trying to protect, what the cost of different failure modes looks like, and how maintenance decisions interact with production schedules, compliance deadlines, and procurement cycles. Research on predictive maintenance consistently makes the same point: the value is not in transforming sensor streams into a dashboard. It is in transforming sensor streams into actionable maintenance decisions that change what people do. That understanding does not come from the data. It comes from being genuinely embedded in the customer’s operational reality.

The business model problem that technology cannot solve

I have argued for a long time, in various forms and with various clients, that IoT is never really about technology. It is about whether the data enables something that was previously a barrier or a struggle worth paying to remove. The technology is only the means. The question that matters is: what does the customer struggle with today that this data could actually fix, and is the fix worth the investment?

That question sounds simple. It is not, because it requires the supplier to understand the customer’s business at a level most hardware companies never reach. They understand how the instrument works. They understand how to install it, calibrate it, maintain it. They may even understand the process it measures. But do they understand what it costs the customer when a device behaves unexpectedly? Do they know how maintenance planning actually happens inside that plant, who makes the decisions, what information they have access to, and what would genuinely change their behavior? Do they understand the compliance burden well enough to make documentation faster rather than just more connected?

These are not engineering questions. They are business model questions. And the company that can answer them - credibly, specifically, for the customers it already serves - will build something that changes behavior. The company that cannot will build a dashboard that sits beside the real workflow rather than inside it. The distinction matters more than it looks from the outside. Data outside the workflow is information. Data inside the workflow changes what people do. The gap between those two states is not a technical integration problem. It is a question of whether the supplier understands the customer’s job well enough to redesign the work around better information.

What the honest test looks like

There are two easy mistakes when looking at a company like this from the outside.

The first is to assume the platform transition is a natural extension of the hardware business. It is not. An installed base is a data acquisition opportunity, but turning that opportunity into recurring decision support requires capabilities that hardware companies rarely develop organically: service design, workflow integration, customer success, and a genuine willingness to be measured on business outcomes rather than technology features.

The second mistake is to assume that because the technology works, the business model follows. It does not. Predictive maintenance is not created by labeling diagnostics as AI. When properly integrated into operational workflows, predictive approaches have reduced monthly downtime incidents by roughly 40 percent compared to five years ago according to Siemens’ own longitudinal data. But the operative phrase is “properly integrated into workflows.” Most deployments stop short of that. They create visibility without changing accountability. They produce analytics that sit beside the real maintenance process rather than inside it.

Vendor neutrality is probably the hardest strategic test. A process plant is a multi-vendor environment. A system that only serves one supplier’s installed base is a vendor tool, useful but limited. A system that works across the messy multi-brand reality of an actual plant floor becomes something the plant owns rather than something the supplier provides. That transition requires the company to invest in becoming genuinely useful to the plant, even when that usefulness does not directly sell more of its own hardware. That is a harder internal commitment than it sounds.

What I think is actually at stake

The strongest version of this company is not an instrument supplier with a digital layer attached. It is a decision-infrastructure company for process industries, one that uses its installed base and operational credibility to become the trusted home for the asset intelligence that drives maintenance, compliance, procurement, and reliability planning across the plant.

That is a genuinely different business. It has recurring revenue characteristics, customer dependency, and defensibility that hardware sales alone do not. It is also a much harder business to build, because it requires the company to develop a depth of customer understanding that goes far beyond knowing how the instrument works.

The companies that get this right will be the ones that stop asking what their technology can do and start asking what their customers cannot solve without it. That requires sitting in the plant, in the maintenance meeting, in the compliance review, in the procurement discussion, and understanding the decisions those people make every day with incomplete information.

The data is not the product. The better decision is the product. The companies that understand that distinction early enough to build around it will be the ones worth watching.

Everyone else will have connected a lot of instruments to a dashboard that nobody changes their behavior because of.

Leave a comment

I have been spending time lately with AI agents. Reading about them, following the discussions, and running some myself in the background to see what they actually do. The honest verdict after a few weeks of this: the automation feels impressive on first contact. You set something in motion, walk away, and come back to find work completed. That is genuinely new in terms of feel.

But once you look under the hood, most of what gets called an agent today is structurally simpler than the name implies. A markup file that orchestrates a sequence of API calls to a large language model, with some tool access layered on top. Not rocket science. In many cases, a thoughtful set of chained prompts would produce something similar. The impression of autonomous intelligence is real. The underlying architecture is considerably more modest.

I am not saying this to dismiss agents. I am saying it because the gap between how agents feel and what they actually are matters for the question I want to ask. Because whether the technology is genuinely autonomous or just cleverly automated, the organizational consequence is the same: something is now acting inside your business without a human making each individual decision. And that changes something important.

It changes who is responsible when the action goes wrong.

The difference that actually matters

A chatbot suggests. A copilot assists. A dashboard informs. An agent acts. It does not generate language for a human to evaluate and then decide. It pursues a goal, uses tools, interacts with systems, makes intermediate choices, and changes the state of the business. That is not a marginal technical improvement. It is a shift in delegation. And delegation always moves accountability somewhere.

The comforting story organizations tell themselves about agents is that they will handle routine tasks, coordinate between systems, and free people for higher-value judgment. Some of that is true and the value is real. But the productivity story skims over the deeper question. Once software starts acting inside the organization, work no longer moves only through people. Decisions no longer sit only in meetings, approvals, and managerial routines. Authority starts migrating into systems. At first this feels harmless - the agent schedules something, drafts something, summarizes something. Then it prepares actions. Then it executes within rules. Then it escalates only exceptions. Then it coordinates across systems. The change does not arrive all at once. It arrives through convenience, and because convenience feels like progress, very few people stop to ask what has actually moved.

What has moved is judgment. And with judgment, accountability becomes harder to locate.

The data is already telling the story

Deloitte’s 2026 State of AI in the Enterprise research, based on a survey of 3,235 IT and business leaders across 24 countries, found that by 2027, 74 percent of companies expect to use AI agents at least moderately, with 23 percent expecting extensive use and 5 percent planning full integration into core operations. That is a dramatic acceleration from where most organizations sit today. What makes the finding striking is what sits alongside it: only 21 percent of those same organizations report having a mature governance model for agentic AI in place right now.

That gap between deployment intent and governance readiness is the actual story. The organization discovers what the technology can do before it decides who is responsible for what the technology does. That sequence is dangerous not because agents are inherently dangerous, but because organizations were already struggling with accountability before agents arrived.

McKinsey’s 2026 AI Trust research found progress in AI trust maturity overall, but persistent gaps specifically in strategy, governance, and agentic AI controls. Only about a third of organizations report meaningful maturity in those dimensions. McKinsey Partner Rich Isenberg put the core shift cleanly: “Agency isn’t a feature -- it’s a transfer of decision rights. The question shifts from ‘Is the model accurate?’ to ‘Who’s accountable when the system acts?’”

That is the right question. Most organizations are not yet answering it.

Accountability was already blurred before agents arrived

Read more

I read a lot. Books, reports, research papers, economic analyses - anything I can find on what AI is actually doing to organizations, labor markets, and the underlying logic of how companies create value. Not the hype pieces. The ones that sit with the uncomfortable data.

What keeps surfacing across all of it is a gap that should bother leaders more than it does. Organizations are adopting AI at a significant pace. Individual employees report real productivity gains. And yet the evidence that any of this is changing how companies actually work -- how decisions get made, how value gets created, how the organization relates to its market -- is remarkably thin.

That gap is not an implementation problem. It is a strategic one. And most companies are not asking the right question to see it.

The question most companies are asking

The question is usually some version of: how much faster can AI make us? How many reports can we automate? How many emails can we draft? How many tasks can we complete with fewer people and less delay?

It sounds reasonable because speed is visible. Cost reduction shows up in quarterly numbers. A shorter task feels like progress. A dashboard full of AI usage statistics creates the impression that something important is happening.

But speed is not the same as movement. A company can move faster and still not move forward.

What I keep seeing is AI entering organizations as a tool for acceleration - used to improve what already exists. Existing processes become faster. Existing roles become more efficient. Existing documents are easier to produce. There is real value in that. But there is also a danger that is harder to name: AI can make an outdated company feel modern again. It can make an old operating model look energetic. It can give leaders the feeling of transformation while the organization remains fundamentally the same.

The tire metaphor

I have been thinking about this as a tire problem.

Many companies are using AI the way you patch an old car with new tires. Marketing gets AI. Sales gets AI. Customer service, HR, legal, finance, product - every function finds a place where the new technology reduces friction. For a while it feels like progress. The tire leaks less air. The vehicle moves a little better. Leaders can point to adoption. Employees can show efficiency gains. The organization feels less exposed.

But it is still the same old car.

Sometimes the real question is not how to patch the car. The real question is whether the vehicle is still the right format for the terrain. Maybe the road has changed. Maybe the ground has become unstable. Maybe old model was built for a world of paved roads, predictable routes, and known destinations, while the next environment is mud, fragmentation, and genuine uncertainty. Or maybe the next advantage is no longer on the ground at all.

That is what AI forces leaders to confront. Not only how to use the technology, but what the technology makes obsolete about the company’s current strategy. Most companies avoid that question. So they patch and patch and...

Why productivity is the safe story

Productivity is the easiest AI story to tell because it does not threaten anyone. It does not question the business model. It does not challenge the operating logic. It does not ask whether the organization is still built for the right environment. It simply says: let us do what we already do, but faster. That is why it is so attractive. It gives the company movement without demanding renewal.

The data makes the gap visible. Gallup’s February 2026 survey of 23,717 US employees found that 65 percent of workers in AI-adopting organizations say AI has improved their individual productivity and efficiency. That is a real finding. But only 12 percent strongly agree that AI has transformed how work gets done in their organization. One in ten. Despite billions spent, despite widespread adoption, despite genuine individual gains -- the organizational level is barely moving.

This finding is not isolated. An NBER survey of nearly 6,000 global executives found that 89 percent see no effect on labor productivity at the firm level. An MIT study found that despite roughly $40 billion in enterprise investment, 95 percent of organizations have seen zero measurable impact on profits. Individuals feel faster. The company may not have moved.

That is the uncomfortable truth sitting underneath the adoption numbers: AI can improve the performance of work that should no longer exist. It can make yesterday’s logic more efficient. And because the improvement is real and visible, it becomes harder to see the deeper problem. The company feels better before it becomes better. Pain decreases. Urgency disappears. A slow organization becomes a slightly faster slow organization.

When everyone optimizes, no one differentiates

There is another reason the productivity frame is too small.

Everyone can use it. Your competitors can summarize faster too. They can generate content, automate internal analysis, equip their teams with the same tools, buy from the same vendors, and follow the same use-case libraries. What feels like an advantage early quickly becomes the new baseline. The first mover feels clever. The second feels responsible. The rest eventually feel behind. But once the technology diffuses, the advantage does not come from using it. The advantage comes from changing the system around it.

PwC’s 2026 AI Performance Study makes this divide visible: 74 percent of AI’s economic value is captured by just 20 percent of organizations, while the majority remain stuck in pilot mode. That finding separates AI activity from AI value. Many companies are adopting. Fewer are genuinely changing.

This is where most organizations stall. They want the benefits of AI without the discomfort of strategic change. They want speed without asking whether they are moving in the right direction. They want efficiency without asking what should stop. They want transformation without disturbing the model that still pays the bills. So they optimize. And because everyone else optimizes too, the whole market accelerates without necessarily changing. More activity. More output. More automation. More internal excitement. More AI language in strategy decks. No new strategic position. No new source of advantage. No new logic.

The most dangerous AI failure

A failed pilot is visible. A bad tool gets rejected. A poor use case dies. The company learns and moves on.

The most dangerous AI failure is a successful optimization of the wrong thing. That is harder to see. The company becomes faster at producing reports that should no longer guide decisions. Faster at preparing meetings that should not happen. Faster at serving a customer journey that should be redesigned. Faster at protecting margins in a business model whose relevance is quietly weakening.

The tire keeps rolling. The ride feels smoother. The problem is that the road has changed.

This is the real seduction of productivity. It reduces pain without forcing diagnosis. And when pain decreases, urgency disappears. A confused organization becomes a more productive confused organization. A legacy business becomes a better-defended legacy business. New technology extends the life of old assumptions. That is not transformation. It is a delay mechanism.

What the technology is actually changing

A business model is not strong in absolute terms. It is strong relative to the environment in which it operates. When the environment changes, yesterday’s strengths can quietly become constraints. A distribution advantage weakens. A knowledge advantage becomes widely available. A trusted process becomes friction.

AI changes the terrain because it changes what is scarce. When knowledge becomes easier to access, judgment becomes more important. When content becomes abundant, relevance becomes more important. When analysis becomes cheaper, decision quality becomes more important. When automation becomes common, choosing the right work matters more than doing all work efficiently.

MIT Sloan research argues that AI’s largest impact may not come from isolated task gains, but from reshaping workflows: how tasks are sequenced, connected, handed off, and recombined between humans and machines. That is true as far as it goes. But the strategic implication goes further. If workflows change, operating models change. If operating models change, business models can change. And if business models can change, the question is no longer whether AI helps the current company. The question is whether the current company is still the right answer.

The question that changes the conversation

Most AI programs begin with use cases: where can we use AI? That sounds practical. It gives teams something concrete to do. It fills a roadmap. But it contains a hidden assumption -- that the current organization is the right starting point.

The better question is this: what kind of company would be built today if AI were already normal? Would you still organize the same functions? Sell the same bundle? Price the same way? Protect the same assets? Define expertise the same way? Call the same activities core?

The World Economic Forum’s 2026 report on organizational transformation makes the same distinction: AI has moved beyond early experimentation, and the opportunity now is to rethink how work is performed, how decisions are made, and how operating models are designed. Adoption inserts AI into the current company. Transformation asks what the company should become because AI now exists. Most companies are doing the first. The second is where the real decisions sit.

The leadership question

The easiest way to weaken AI is to make it an IT project.

The AI team owns it. The digital team owns it. The transformation office owns it. This helps with coordination but creates distance from the real issue. AI becomes a portfolio of initiatives -- visible but not decisive.

The real question belongs to leadership. What parts of our strategy become stronger because of AI? What parts become weaker? What parts become obsolete? And what would a new entrant build today if it had no legacy, no internal politics, no historic revenue to defend, and the same access to AI that we have?

That last question matters most. Because the most dangerous competitor is not the one using AI to improve the old model. It is the one using AI to ignore the old model entirely. They are not patching the tire. They are asking why everyone is still driving.

Every leadership team should sit with one honest question: are we patching the car, changing the vehicle, or reconsidering whether the vehicle still belongs on this terrain? Patching has value. It buys time and reduces waste. Changing the vehicle is harder but necessary. Reconsidering the terrain is where strategy actually lives.

Most companies will patch. Some will redesign. Few will rethink the game.

AI is not mainly a productivity tool. It is a strategy test. It tests whether leaders can see beyond efficiency, question the model that made them successful, and distinguish motion from movement. A smoother ride does not mean you are going in the right direction. A faster vehicle does not matter if the road no longer leads anywhere worth going.

The terrain is changing. The question is whether you are measuring the right things to notice.

Leave a comment

Something has been bothering me for a while, and a presentation I sat through recently brought it into focus.

The topic was AI-powered innovation workflows: agentic tools, automated pipelines, synthetic personas, AI-generated user research. The pitch was familiar. Move faster, reduce cost, generate more concepts, test more variants, validate earlier. The automation was real. The speed gains were real. About halfway through, I noticed something missing.

There were no actual users in the process. Somewhere along the way, the conversation with a real person had been replaced by a synthetic abstraction of one. AI personas built from demographic assumptions. Behavioral models constructed from historical patterns. Simulated responses from people who do not exist.

I understand the appeal. Real users are hard to reach, slow to schedule, and inconsistent in ways that make analysis uncomfortable. Synthetic stand-ins are faster, cheaper, and available at two in the morning. The research on this is also real: a 2024 Stanford and Google DeepMind study found that AI agents built from two-hour interviews with 1,052 people replicated their subjects’ social survey responses with 85 percent accuracy. That is genuinely useful at the hypothesis-generation stage.

But there is a hard boundary. ACM Interactions research published in late 2025 puts it clearly: synthetic personas produce confident but inaccurate direction. They validate bad assumptions, confirm biases, and create blind spots. A comparative study of B2B research found that AI-generated personas showed strong positive bias compared to real respondents and followed a herd mentality that real buyers do not. The practical rule that emerges from this body of work is straightforward: synthetic research is useful for the first 80 percent of discovery. The remaining 20 percent -- the deep, situational, emotionally textured part of a decision -- still requires a real person.

That distinction is not a footnote. It sits at the center of what Jobs to Be Done is actually for.

What JTBD is really about

Jobs to Be Done is usually introduced through three dimensions, and most teams stop there.

The functional job is the practical task someone is trying to complete: file a claim, compare options, write a report, diagnose a problem. The social job is about how someone wants to be seen by others: competent, prepared, credible, not the person who missed something obvious. The emotional job is about how someone wants to feel, or avoid feeling: confident, in control, not exposed to regret, not left holding a decision they cannot defend.

These three still matter. But in the age of AI they are no longer sufficient to explain where human value remains or where competitive advantage actually sits. Two additional dimensions are becoming more strategically important, and they are the ones that automation handles worst.

The relational job is about how someone wants to be treated by another human being. Not just served - treated. Understood. Taken seriously. Not processed. When a customer reaches a genuinely difficult moment in a decision, what they often need is not a faster answer. They need to feel that the person or organization on the other side of the transaction actually sees their situation.

The situational job is about fit to a specific context. Help me navigate my case, not the average case. Help me adapt this to my constraints, my timing, my trade-offs, my history, my risks. The average case is a statistical construct. No buyer lives there. They live in one specific company, one specific team, one specific set of pressures that the standard solution was not designed around.

These five dimensions together give a much more complete map of what progress actually means to a customer - and they reveal something structurally important: AI is increasingly capable on the functional layer and progressively weaker as you move toward the relational and situational ones.

Where value moves when the functional layer gets cheaper

Read more

A few weeks ago I had dinner with a friend who has spent the better part of the last decade trying to do something genuinely useful inside the education system. By the end of the evening I was sitting with a feeling I did not expect. Not sadness exactly. More like the particular frustration you get when a system fails someone in a way that is completely avoidable and entirely predictable at the same time.

Let me explain what I mean.

He came to education through an unusual path. He studied first history at the local university and then later in his life at Hyper Island, which describes itself as a global platform for lifelong learning focused on helping individuals and organizations meet the challenges of a changing world through transformative education. (hyperisland.com)

After that, he became part of an edulab focused on helping children and young adults build exactly those capabilities. The schools he worked with were glad to have him. That part matters. Because it means the gap he was filling was real enough to be felt and valuable enough that schools actively welcomed outside help to close it.

And yet the arrangement was always structurally fragile. The schools benefited. The value was visible. But the funding from the government and schools were weak, temporary, or absent. So the gap got filled informally. The work was recognized but not secured. This is what many systems do when they cannot genuinely reform: they improvise around the problem instead of rebuilding around it. That can look like progress from the outside. Until the person doing the bridging needs to make a living.

The credential question

When the informal arrangement ran its course, he ran into the next wall and got finally fired because of missing financials. Outside the school environment, the market did not know what to do with what he had built with others. Inside the school system, the formal requirement returned with full force: if you want to belong here permanently, you need recognized pedagogical training. In Switzerland, that means a degree-based path combining disciplinary study, educational science, and supervised teaching practice. That standard makes sense. Institutions need baselines. But standards solve one problem while sometimes deepening another, and the system was not asking what capability it lacked. It was asking what credential counts. Those are not the same question.

A Person Can Solve a Real Problem and Still Not Fit the System

So he did what serious people do when an institution sets a gate. He respected it. He invested years of his time and a significant amount of his own money to complete the formal path into teaching, with the reasonable hope that he could finally bring his earlier background into the system with legitimacy rather than just goodwill. On paper, the logic is sound: if the problem is that schools need future-oriented capability, and the system requires formal pedagogy to let you in, then acquire the qualification and return stronger.

But this is where the story turned, and where I found myself setting down my glass at dinner.

What he actually found inside

Once inside the system properly, he discovered that it was still not genuinely organized around the capability gap that had drawn him there in the first place. Instead of entering a profession oriented around digital-age readiness and the kind of practical, entrepreneurial thinking he had spent years developing, he largely entered a system still centered on curriculum delivery and conventional teacher roles. The gap had not disappeared. It had been normalized. His background was acknowledged. It was not made central.

This is what institutional neutralization actually looks like, and it is worth naming clearly because it is far more common than outright rejection. Systems rarely refuse new capability directly. They do something more subtle. They absorb it, dilute it, and assign it to the margins. A person enters because they can help solve a known deficiency, and then they are folded into a structure whose main routines were not built for that deficiency to matter. The original reason they were valuable gets downgraded to a side topic, an add-on, a special session. The system keeps the person. It protects itself from what the person actually represents. That is how institutions can acknowledge the future while continuing to operate from the past.

OECD’s 2025 work on the digital transformation of school education says access to high-quality digital technologies remains uneven and that their use often falls short of genuinely transforming teaching and learning practices. A related OECD paper on preparing teachers for digital education makes the point even more directly: teachers are being asked to handle new digital demands, but systems face persistent barriers in turning those demands into routine classroom practice. That is a careful way of saying what my friend experienced firsthand. The problem is not hardware. It is not software. It is whether schools have the people, incentives, and structures to turn digital possibility into changed practice. That is a much harder problem, and it is exactly where people with his background should matter most.

The double waste

What struck me most at dinner was not the injustice of it, though that is real. It was the strategic clumsiness. Because the system is not just failing one person. It is failing itself twice over.

The first failure is the one already described: a person invests years building relevant capability, then invests again to meet the formal criteria of the system, then ends up largely absorbed into routines that treat his original capability as secondary. That is already an inefficient use of human potential and private investment.

The second failure is harder to excuse. Because the tools now exist to spread exactly the kind of expertise my friend carries across many more classrooms than any single person could reach. UNESCO’s work on AI in education points to broader access and more personalized learning as real possibilities, with the important caveat that institutions need strong safeguards around equity and rights to make it work. The OECD’s 2026 Digital Education Outlook goes further and argues that generative AI tools need to be co-created with teachers so they can actively guide student learning rather than simply bolt new technology onto old routines.

The implication is direct. If every school had to hire a rare hybrid profile -- part educator, part digital practitioner, part practical guide to a changing world -- the model stays expensive and fragile. But if that expertise can be codified into tools, guidance, and support systems that ordinary teachers can use inside ordinary classrooms, the problem shifts from heroic hiring to scalable enablement. That is a fundamentally different proposition. The irony is that a system saying it cannot fully afford people who bring the missing capability is simultaneously underusing the tools that could spread that capability more cheaply across the people it already has. My friend’s knowledge, instead of being pushed to the margins of one school, could in principle be traveling across dozens. The system is choosing, structurally if not consciously, not to ask that question.

What schools actually need to change

The hard question is not whether pedagogy matters. It does, and the formal requirements exist for good reasons. The hard question is whether schools mean it when they say they want to prepare children for a world that looks nothing like the one their institutions were built for.

If the answer is serious, then hiring logic needs to ask not only whether someone fits the existing mold, but whether they bring a capability the system has already demonstrated it needs. Funding needs to stop depending on fragile informal arrangements for work that has already proven its value. Teacher formation needs to become a route for bringing new capabilities into the center of the profession, not a process that sands them down on the way in. And schools need to use technology to spread scarce expertise across the teachers they already have, not as a gesture toward modernity, but as a deliberate act of leverage in a budget-constrained system.

None of that is radical. All of it requires a decision that most systems keep deferring.

What stays with me

My friend is not a disappointed idealist. He is a serious professional who made a rational series of decisions, respected the system’s gates, paid the costs, and still ended up on the wrong side of a structural problem that the system itself has never had the honesty to name directly.

What stays with me from that dinner is not the personal frustration, though I understand it. It is the recognition that the future is not absent from education. It is present in people like him, sitting inside institutions that welcomed them for what they could offer, then organized themselves to need it as little as possible.

Schools keep saying they want to prepare children for what is coming. Then they hire for what was, credential for what has always counted, and call the gap a reform agenda.

The future is not missing. It has just learned to stop expecting a proper welcome.

Leave a comment

In February, Anthropic published a short statement that stopped me mid-read. The opening line: “There are many good places for advertising. A conversation with Claude is not one of them.”

That sentence did not land as marketing. It landed as a deliberate positioning decision with real commercial consequences. Because at almost exactly the same moment, OpenAI was moving in the opposite direction, testing ads inside ChatGPT for logged-in users on its free and Go tiers, with sponsored content clearly labeled and separated from answers, privacy protections around chat data, and restrictions around sensitive categories like health, politics, and legal or financial questions.

Two of the most significant AI companies in the world, looking at the same surface, reaching opposite conclusions about what it should become. That contrast is worth thinking through carefully.

This is not a product comparison

The easy read is that one company is keeping it clean while the other is compromising for revenue. That framing is too simple and too comfortable to be useful.

Anthropic’s argument is structural. A conversation with an AI assistant is meaningfully different from a search result or a social media feed. People share more. The format is open-ended. An appreciable share of conversations involve topics that are sensitive or deeply personal, the kinds of things you might say to a trusted advisor rather than type into a search box. Anthropic argues, and I think correctly, that introducing advertising incentives into that context would shift what the model is optimizing for, even if the ads themselves appear separately from the answers. The risk is not only manipulation. It is the slow drift of the whole system toward engagement metrics that have nothing to do with being genuinely useful. As Anthropic puts it, the most useful AI interaction might be a short one, or one that resolves a question without prompting further conversation. Ad-optimized systems are not built to want that outcome.

OpenAI’s counter is also structural. Its ad design is built around the observation that people come to ChatGPT when they are actively exploring options, comparing ideas, or working toward a decision. That is a commercially valuable surface, and OpenAI has chosen to monetize it explicitly rather than through subscriptions alone. Both positions are internally consistent. What makes the contrast interesting is what it reveals about where value will actually accumulate in the AI economy, and who captures it.

Digital markets concentrate. Then the winner monetizes the position.

The underlying pattern is familiar enough to name quickly.

Digital markets rarely settle into healthy pluralism. They tend to concentrate around the strongest product or the best distribution, and once a platform becomes the place where people search, compare, or ask for help, monetization stops being a side activity. It becomes the operating logic of the system. Alphabet still breaks out Search, YouTube Ads, and Google Network as its major advertising revenue lines, and in early 2026 it reported annual revenue exceeding $400 billion for the first time, with Search and YouTube still growing. Google did not just win search. It won a privileged position between human attention and commercial intent. The gap between winning the product and winning the monetization layer is what made that position durable for two decades.

Conversational AI is starting to look structurally similar, and the scale of what is at stake is worth naming directly. Search captured what people typed into a box. Conversational AI can capture what people are actually trying to do, where their confidence is shaky, and what they still need before they are ready to act. That is a more granular and more actionable position than keywords alone. When a conversational AI becomes the first place people go for research, planning, professional orientation, or a second opinion on a decision, it does not just intermediate information. It intermediates intent. And unresolved intent is commercially valuable in a way that a completed search query is not.

Both Anthropic and OpenAI understand this. They are simply betting on different ways to sit inside it.

The gap between useful and trustworthy is where the real market forms

Read more

In February, the OECD published a short report I have not been able to stop thinking about. The headline number: AI firms captured 61 percent of all global venture capital in 2025. That is $258.7 billion out of a total $427.1 billion, more than double AI’s share from just three years earlier. The report is measured and descriptive. It does not draw the conclusion I am about to draw. But when you read it alongside everything published on demographic stress, climate adaptation, care systems, and institutional trust, the picture that emerges is uncomfortable.

We are not building the future. We are building one part of it, very fast, and leaving the harder parts largely to chance.

This is not an argument against AI

The deeper issue is that AI productivity gains are not enough if companies do not also redesign where value is created, captured, and defended.

Let me be direct about what this is not. I use AI tools. The productivity gains are real and the downstream applications are significant. Some of what is being funded will genuinely matter, both economically and for human welfare.

But 61 percent is not a portfolio. It is a concentration. And concentrations have a habit of revealing priorities that no one ever explicitly decided on. That raises a harder question: who should fund corporate innovation when capital is chasing the most fashionable category instead of the most exposed problem?

An economy that allocates capital also allocates attention, talent, and problem-solving effort. When three-fifths of global venture money flows to one category (even it becomes a general purpose technology), the implicit message is that this is where the important problems are. Everything else competes for the remaining 39 percent. Elder care, flood resilience, antimicrobial resistance, public health infrastructure, institutional competence -- all of it, every other domain, shares what is left. That is worth sitting with for a moment before concluding that the market has this right.

The next 20 years will not be a computation problem

Here is what I think the next decade and a half will actually test.

Populations across the developed world are aging rapidly, and the pace is no longer a projection. It is already visible in labor markets, care systems, and pension finances. The UN’s 2024 population outlook is explicit: decades of low fertility combined with longer life expectancy are driving rapid aging in many countries, with some already seeing population decline. This was not a surprise. It has been in the data for a long time, which makes the absence of a serious innovation response all the more striking.

Climate adaptation is a different story from the one that gets told most often. Public discussion still centers on mitigation: carbon reduction, energy transition, new technology to lower emissions. Those things matter. But for the next 10 to 20 years, much of the lived experience of climate change will be adaptation to conditions already locked in. Heat stress, flooding, water scarcity, the retreat of insurance from entire regions, the redesign of urban infrastructure for a world that is already warmer. The IPCC states with very high confidence that risks and damages escalate with every increment of additional warming. That means the human problem is not only how to stop future warming, but how to make societies physically livable and economically viable under the warming already underway. That is a genuinely different innovation agenda from the one that captures most of the capital.

Then there is institutional trust, which tends to get treated as a soft concern, something governments should fix with better messaging. It is not soft. A society with low institutional trust cannot execute difficult transitions cheaply. Every reform becomes more contested. Every necessary sacrifice requires more coercion or more subsidy. OECD’s Government at a Glance 2025 puts aging, the green and digital transitions, low trust, stagnating productivity, and constrained fiscal space into the same frame, as combined pressures on public governance. That framing matters because it puts institutional competence where it belongs: at the center of the problem, not at the edge of it.

Why capital keeps going where it goes

The market is not irrational. It is solving for its own objective function, and AI fits that function extremely well.

Software scales. Once built, it can reach a billion users at near-zero marginal cost. AI sits at the intersection of software economics, infrastructure scarcity, and what looks like a genuine platform shift. The companies that own the foundational models and the compute beneath them hold leverage that investors can see and value clearly. Elder care does not work that way. Flood resilience does not work that way. Antimicrobial resistance research does not work that way. These domains require patient capital, public coordination, long time horizons, and returns that are distributed across society rather than captured by shareholders.

That is precisely why they are easy to underfund. It is not malice. It is structure. But what makes sense for capital allocation does not automatically make sense for civilization. We are financing the machinery of cognition faster than we are financing the human systems required to absorb its effects. That gap is the actual risk, and it is growing.

Productivity is not the same as progress

This is where the standard innovation narrative becomes evasive, and it is worth naming that directly.

Productivity growth is still treated as though it were automatically social progress. It is not. Productivity growth does not answer the central political economy question: who captures the gains, and what happens to the people whose roles, assets, or bargaining power weaken in the process? A society can become technologically stronger while becoming socially more brittle. It can automate work and still fail to create security. It can lower friction and still deepen distrust. It can raise GDP and still erode the conditions that make growth politically tolerable over time.

The language we reach for -- transformation, disruption, the future of work -- tends to obscure this. It implies that gains eventually spill over, that if the technology is powerful enough, everyone benefits eventually. History is less charitable. Gains spread when institutions, bargaining structures, public investment, and asset access force or enable diffusion. They do not spread simply because the technology is impressive.

The problem health resilience reveals

Take one example that rarely appears in the innovation conversation at all.

One of the next major health threats is not a surprise pandemic but the slower erosion of medicine’s effectiveness through antimicrobial resistance. WHO’s 2025 surveillance report analyzed more than 23 million bacteriologically confirmed infections across 104 countries, covering bloodstream infections, urinary tract infections, gastrointestinal infections, and urogenital gonorrhoea. One in six of those infections involved bacteria no longer responding to standard antibiotics -- rising to one in three for urinary tract infections. WHO frames antimicrobial resistance as a serious, growing threat that is undermining the foundations of modern medicine. This is exactly the kind of long-burn, high-consequence problem that attracts nowhere near the cultural excitement of a frontier model release, and nowhere near the capital.

That contrast is instructive. The problem is real, well-documented, and not going away. The market simply does not find it as monetizable as the next infrastructure layer for AI inference.

The reactive logic and its limits

Some will argue that markets eventually redirect capital when real pain becomes impossible to ignore. When adaptation costs rise enough, they become investable. When care shortages intensify, labor-saving redesign becomes unavoidable. When public systems crack, governments pay attention.

There is some truth in that. But it is a reactive logic. It waits for stress to become expensive enough for capital to care, which means it consistently arrives late, after preventable damage has accumulated. That is a poor operating model for structural transitions that are already underway. Many of the most consequential investments have public-good characteristics that cannot be justified through venture math alone. They create stability rather than hype, they compound differently from software, and they benefit people who are not the ones writing the checks. That is exactly why they are chronically easy to deprioritize.

The benchmark is wrong

This is also why innovation fails when strategy avoids hard choices.

The strongest critique here is not a moral one. It is strategic.

We are overfunding the acceleration layer and underfunding the absorption capacity. We are scaling computational power faster than social resilience. The longer that gap persists, the more likely we are to confuse technical progress with civilizational progress. They are not the same thing.

A serious innovation economy should assess investment less by novelty and more by whether it addresses high-consequence bottlenecks in welfare and societal stability. It should be honest that markets systematically underfund domains with public-good characteristics, and design deliberate reweighting -- through public investment, procurement, and institutional reform -- to compensate. It should retire the language that treats all innovation as equivalent, because it is not. Some innovation expands human room to maneuver under pressure. Some mainly intensifies competition inside already overcapitalized domains.

The OECD report is not an alarm. It is a ledger. It tells you where the bets are going. Reading it carefully, the open question is not whether AI will be important. Of course it will. The question is what we will be able to do with it in a society that failed to invest adequately in the human systems that make any technology livable.

We have become extremely good at building what scales. The harder question is whether we can still build what holds.

Leave a comment