AI Governance Policy and Workforce Capability Retention

AI governance documents are built around risk. Where does the data go? What can the model access? Can its output be trusted? Who approves its use, and what happens when it gets something wrong?

These are all valid questions. There is another one, and it sits outside the frame.

What happens to the capability of the people using AI when the technology works exactly as intended?

That is a very different kind of risk. There is no security incident, no hallucination to report, no failed control, no obvious warning sign. The work gets done, perhaps faster than before. But somewhere along the way, the people doing it are getting less practice at what makes them human. And good at their job. Thinking.

Skill is an asset, and assets can be depleted quietly.

The question is what gets handed over

Cognitive offloading has been with us for a long time. Calculators removed the need to do arithmetic by hand, satnav removed the need to remember every turn, calendars removed the need to hold every appointment in our heads.

Generative AI reaches further than the mechanical part of a task. It can recommend how to start. How to structure the problem. What questions to ask. What the first draft should say. How the database should be designed. What the commercial argument should be.

That distinction is the whole thing. Using AI to accelerate work that has already been thought through is one activity. Asking AI to do the thinking that creates the solution in the first place is another. The first makes people faster. The second, done habitually, can make them dependent.

Reviewing work is a different job from creating it

Governance frameworks rightly insist on human oversight. The EU AI Act goes as far as naming the psychological pull directly. Article 14 requires that anyone assigned oversight of a high-risk system be enabled to remain aware of the tendency to over-rely on its output.

Reviewing work and creating work are separate activities. Reviewing an answer is a different mental exercise from creating it. Checking a proposal is not the same as developing the win strategy behind it. Approving a database structure is not the same as deciding how the data should relate in the first place.

A Microsoft Research and Carnegie Mellon study of 319 knowledge workers found that greater confidence in generative AI was associated with less reported critical-thinking effort. The nature of the work shifted too, moving away from producing the work and towards supervising what the AI produced.

That falls short of proving AI causes people to lose critical-thinking ability. But it does point at something organisations should take seriously. If people spend less time exercising a capability, then it is a fair question to ask, "what happens to that capability over time?".

The problem has a long history. In 1983, Lisanne Bainbridge described what she called the ironies of automation. Automate the routine work and the human operator becomes a monitor. Then something unusual happens, and that same operator is expected to step back in and exercise skills they have had little occasion to use. The technology has moved on considerably. However the irony still holds.

AI can build capability too

There is an important counterpoint. Research covering 5,172 customer-service agents, published in the Quarterly Journal of Economics, found that access to an AI assistant lifted productivity by 15% on average, with the largest gains among the least experienced staff. The AI was effectively passing on the working methods of the strongest performers.

The detail that matters most is, during software outages, when the AI offered nothing at all, those workers still performed above their pre-AI baseline. Some of the gain had become theirs.

So the honest conclusion is that AI builds capability or erodes it depending on what it is asked to do. That distinction belongs in AI governance.

The compounding judgement test

There is a practical way to decide what stays human. Before automating a task, ask one question: does the judgement involved in doing this work improve through repetition?

Where the answer is yes, think hard before handing the point of origination to AI.

Consider an RFP response. AI can extract the requirements from hundreds of pages, identify unanswered questions, retrieve approved company material, assemble a first draft, and compare the response against the scoring criteria to flag gaps. Those are measurable productivity gains and they are worth having.

The win strategy is a different matter. What the customer really values. How the company differentiates itself. What commercial position the business is prepared to take. Those are judgement calls, and they belong with the commercial team, because every time that team makes those judgements, they get better at making the next one.

The same applies elsewhere. AI can build a database quickly, once someone has understood how the data needs to relate. AI can generate an interface, once someone has decided what the user needs to accomplish. AI can draft a negotiating position, once the commercial team understands the interests, the leverage and the trade-offs.

The question has stopped being whether AI can do the task. Increasingly, it can. The question is whether doing the task is part of how the organisation builds expertise it will still need tomorrow.

This fits in a paragraph. It is a principle rather than a framework, and it reads something like this:

Where performing a task develops judgement, expertise or commercial capability that the organisation considers strategically important, AI should augment, rather than replace, human origination.

That principle then gets applied whenever a process is redesigned or automated. It applies selectively, and that is the point of having it. Some work stays repetitive however many times somebody performs it. Extracting tender requirements. Reconciling records between systems. Classifying hundreds of customer messages and routing them to the right person. Moving approved information from one place to another. Automating that work is straightforwardly useful.

Other work compounds. Commercial judgement. Problem solving. Strategic positioning. Architecture. Negotiation. Writing an argument rather than polishing one. Those capabilities become more valuable through use.

When an organisation gradually stops exercising them, the loss stays invisible for a while. Then the AI fails, or the market turns, or a genuinely new problem lands on the desk, or the business needs somebody to form a position from a blank page. At that point the capability matters again, immediately and expensively.

Governance can protect more than systems

Governance asks what AI might do to an organisation. Bias. Security. Privacy. Hallucination. Compliance. Poor decisions. Those risks are real and they deserve the attention they receive.

The companion question is what an organisation might stop knowing how to do, because AI made it convenient to stop doing it.

The answer to that is deliberate use, decided in advance and written down.

Use AI to remove work that adds little human value. Use it to accelerate execution. Use it to expose people to expertise they have yet to build. Use it to critique, test, assemble and automate. And watch closely at the point where it begins to take over the work through which judgement itself develops.

Speed is the easy thing to measure. Capability is the thing you will want back.