Welcome to A(I)verage Land
What happens when companies use AI to detach work from expertise instead of expanding what expertise can do?
A colleague whose work I have followed for years recently described how he used AI to build a small interface for a task that an existing software product handled poorly. He did not wait for the vendor to improve its product or accept the friction as unavoidable. He described what he needed, worked with the model, evaluated the result and created something that made the task easier.
It was an impressive example of how generative AI can lower the cost and speed of turning an idea into a working tool. It would also be easy to draw the wrong conclusion from it.
The result was not useful because AI had made expertise unnecessary. It was useful because an expert was using AI.
He understood the work before he began. He could explain the problem, distinguish a relevant feature from an attractive distraction and recognize when the generated output failed to support the outcome he wanted. The visible production work became easier, but the judgement did not disappear. It moved into framing, directing, testing and deciding what was good enough.
This is the distinction many organizations are in danger of missing.
They see AI producing a report, a prototype, a summary, a strategy memo or a software interface and conclude that the task has been detached from the person who previously performed it. The output appears, often faster and at lower cost, so the expensive expertise surrounding it begins to look optional.
What they may actually have detached is the visible artifact from the invisible judgement that made it trustworthy.
AI is reaching work that was difficult to codify
Earlier waves of workplace automation were most effective when processes could be expressed through stable rules. A transaction met predefined conditions or it did not. A field was complete or missing. A machine followed a sequence that engineers had specified in advance.
Generative AI operates differently. It can draft, classify, synthesize, translate, code and propose options even when nobody has written an explicit rule for every possible input. That is why it is reaching occupations that previous automation affected less directly.
Research published in Science estimated that approximately 80% of the US workforce could have at least 10% of its tasks affected by large language models, while around 19% could see exposure across at least half of their tasks. Higher-income knowledge work is not protected; in many cases, it is more exposed.[1]
Exposure, however, is not the same as replacement. The same research evaluates tasks, not complete occupations. David Autor has made this distinction across several generations of automation: technologies substitute for some tasks while complementing others, changing the composition of jobs rather than simply eliminating an occupation in one move.[2]
This is where the managerial logic can become misleading. A company decomposes a role into visible tasks, tests which of them an AI system can perform and then calculates how many people it may no longer require. The analysis appears precise because tasks are easier to count than judgement.
A task list can show who drafts the document. It rarely shows who notices that the question behind the document is wrong.
The floor rises faster than the ceiling
One of the most important field studies of generative AI examined more than 5,000 customer-support agents. Access to an AI assistant increased productivity by 15% on average, but the benefits were distributed unevenly. Lower-skilled and less experienced employees improved substantially, while the most experienced and highest-performing agents saw little productivity improvement and a small decline in conversation quality.[3]
The study was conducted in one company and one comparatively structured occupation, so it should not be generalized to every form of knowledge work. Its pattern is nevertheless significant.
AI captured and distributed some of the practices associated with stronger performers. It helped less experienced employees move faster along the learning curve. That is valuable. It raises the floor of organizational performance.
Raising the ceiling is a different problem.
In a field experiment involving 758 consultants, participants using GPT-4 completed more tasks, worked faster and produced higher-quality results on assignments that fell within the model’s capability frontier. On a complex task outside that frontier, however, AI users were 19% less likely to reach the correct answer.[4]
AI did not become universally helpful because the users were intelligent or professionally trained. Its contribution depended on the task, the system’s capability and the user’s ability to recognize where that capability ended.
This creates an uncomfortable possibility for organizations. AI can make average work better while also making weak judgement harder to detect. The output becomes more polished, coherent and confident even when the reasoning beneath it remains incomplete.
The floor rises. The ceiling does not move automatically.
A(I)verage Land looks productive
A(I)verage Land does not look like technological failure. It looks efficient.
Reports are completed faster. Customer interviews are summarized within minutes. Presentations become more coherent. Campaign concepts appear in large numbers. Strategy documents use the correct language and follow a convincing structure. People who previously struggled to produce acceptable work can now produce it quickly.
This is real progress when the organization’s problem is inconsistent basic execution.
The strategic risk appears when similar companies use similar models to produce similar outputs from similar data. The efficiency advantage becomes a new operational baseline rather than a durable source of differentiation.
There is already evidence of this convergence in creative work. In an experiment published in Science Advances, access to AI-generated ideas improved the average quality and creativity of short stories, particularly for less creative writers. At the same time, the resulting stories became more similar to one another. Individual performance increased while collective diversity declined.[5]
This does not prove that every organization using AI will become strategically identical. A short-story experiment is not a corporate strategy process. It does show the mechanism behind the average trap: a tool can improve each individual output while narrowing the variation across the system.
That trade-off matters because organizations do not innovate by producing the largest number of acceptable answers. They need variation, dissent, contextual knowledge and unusual combinations from which stronger possibilities can emerge.
If AI helps everyone produce the most statistically plausible response, the organization may become more articulate without becoming more original.
Expertise does not disappear. It changes location.
The visible output is only one part of professional work.
A researcher does not merely produce a synthesis. The researcher decides which question matters, which evidence belongs in the analysis, which sources are credible and which anomalies should not be averaged away.
A strategist does not merely produce options. The strategist diagnoses the situation, recognizes trade-offs, understands organizational constraints and determines which uncertainty must be resolved before commitment.
A designer does not merely produce an interface. The designer understands the user’s job, the consequences of friction and the difference between a usable screen and a useful experience.
When AI performs part of the visible task, expertise moves into the surrounding system. It becomes more important in problem framing, context selection, exception handling, quality assessment and decision ownership.
The danger is that organizations see less visible expert labour and conclude that less expertise is required.
Research on automation bias has shown that people can accept incorrect automated advice, omit their own checks or stop searching for contradicting information. These effects occur among both inexperienced and expert users and are not reliably eliminated through simple instructions.[6] A 2024 behavioral experiment also found that participants followed AI advice even when it conflicted with available contextual information and their own interests.[7]
Over time, the risk extends beyond isolated errors. A longitudinal case study of an accounting organization found that reliance on cognitive automation weakened activity awareness, competence maintenance and the ability to assess outputs. The skill loss remained partly hidden until employees needed to operate without the system.[8]
The organization had not merely automated a task. It had allowed the capability to understand and verify the task to erode while human employees remained accountable for the result.
That is the road into A(I)verage Land: expert-shaped output expands while the organization’s ability to recognize when it is wrong begins to contract.
The job behind an AI investment
Leadership teams often frame AI adoption through the question:
Which tasks can we automate?
That question is useful for identifying efficiency opportunities, but it is too small for deciding how AI should change an organization.
The more consequential job is:
Help us increase the quality and reach of organizational judgement without losing the expertise required to recognize errors, exceptions and strategic alternatives.
Once the job is framed this way, the decision changes.
The objective is no longer to produce the same work with fewer experts. It is to remove the clerical and analytical drag that prevents expertise from being applied where it matters most.
An AI initiative remains operational when it only reduces the time or cost of producing an existing output. It becomes strategic when it allows the organisation to examine more evidence, consider more credible alternatives, detect weak assumptions earlier or make decisions that were previously too slow, fragmented or expensive.
Both forms of value are legitimate. Confusing them is not.
A company can become cheaper without becoming better. It can produce more without learning more. It can distribute expert-shaped language without distributing the judgement that gives that language meaning.
Before expanding an AI initiative, leaders therefore need to decide whether they are raising the floor, extending the ceiling or quietly removing the structure that holds the ceiling up.
