What Happens to Knowledge Work When AI Does the Knowing

Knowledge work is work whose output is information and judgment rather than physical goods: advising, analysing, writing, designing, planning and deciding. Peter Drucker gave it the name in 1959, and for most of the time since, the scarce thing in it was knowing: having the information, the expertise and the time to turn one into the other. AI has made knowing cheap. What happens to knowledge work now depends on what was actually valuable in it all along.

The short answer is that the work moves up a level. Less of the day goes on producing the answer, more on deciding what the question is, whether the answer is right, how it fits what the organisation already knows, and who stands behind it. That is not a smaller job. For many people it is a harder one.

What the evidence says is changing

Three findings, from three different kinds of study, point in the same direction.

People are becoming supervisors of their own work. BCG’s AI at Work survey of 11,749 people, published in June 2026, found nearly half of respondents say their roles have shifted towards managing and directing AI rather than doing the work themselves, and 72% say the skills expected in their role have changed.

Thinking is shifting towards verification. A Microsoft Research study presented at CHI 2025 surveyed 319 knowledge workers about 936 real examples of using generative AI at work. It found AI shifts critical thinking towards verifying information, integrating responses and stewarding the task. It also found something less comfortable: the more confidence people had in the AI, the less critical thinking they applied, while the more confidence they had in their own ability, the more they applied.

Output can look better while being wrong. In the experiment with 758 Boston Consulting Group consultants, recommendations produced with AI on a task outside its capabilities were rated more persuasive even when they reached the wrong conclusion, as set out in will AI replace consultants.

Put together: the work is moving from making to checking, the checking is at risk exactly when the tool seems most reliable, and the finished product no longer reveals whether the checking happened.

The four jobs left when AI does the knowing

If AI produces much of the first draft, four things remain for the person, and each is a skill in its own right.

Framing. Deciding what question is worth asking, what a good answer would need to include, and what the answer is for. A well-framed request to a capable model produces useful work. A vague one produces a confident answer to the wrong question.

Verifying. Checking the output against sources, against the numbers, against what the organisation already knows to be true. Not reading it for tone, but testing whether it holds. This is where the Microsoft finding bites: the reflex weakens exactly when it is most needed.

Integrating. Fitting the answer into the context only the organisation has: this client’s history, last year’s decision, the constraint nobody wrote down. AI can only use the context it is given, which is why that context becomes an asset.

Deciding and owning. Choosing what to do and being answerable for it. No model carries accountability, and clients, regulators and colleagues all know it.

These four were always part of senior knowledge work. What has changed is that they are now most of it, and they are arriving earlier in people’s careers.

What it looks like in practice

Take a routine piece of advisory work: a client memo on a regulatory change.

The old version took a junior most of a day to research and draft, and a senior an hour to correct. In the new version, the junior spends twenty minutes framing the request: which client, which of their operations, what they already know, what the memo must decide. The model drafts in minutes. The junior then spends the next hour verifying: opening each cited provision, checking every figure, marking what they could not confirm. They integrate what only the firm knows, the client’s appetite for risk and last year’s advice on a related point. The senior then decides what the firm recommends and signs it.

The total time falls. The junior’s day is spent on the parts that teach judgment. And the file shows what was checked, by whom, which is what a client or a regulator will ask for if the advice is ever questioned.

The confidence trap

The most practical implication of the research is also the least intuitive. The danger is not the tool that is obviously unreliable. It is the one that is right often enough to stop being checked.

Organisations can design against this. Name a reviewer for anything that leaves the building. Make verification visible, with sources linked and figures traced, rather than implicit. Keep people practising the underlying skill, so their confidence in their own judgment stays earned, because that confidence is what the Microsoft study found keeps critical thinking switched on. And treat a polished answer as a claim to be tested, never as evidence that the testing was done.

Context becomes the valuable asset

When everyone has access to the same models, the difference in output comes from what each organisation gives them: its instructions, its examples of good work, its client knowledge, its standards. That body of context is what turns a general-purpose model into something that produces work only this firm could produce.

It is also the part most firms leave scattered across inboxes, personal chat histories and people’s heads. Writing it down, keeping it current and keeping it under the firm’s own control is one of the highest-value pieces of knowledge work there now is. The data side of that is covered in data sovereignty, and the operational side in the AI operating system.

How to organise knowledge work now

A few principles hold across professions.

• Redesign roles around the four jobs of framing, verifying, integrating and deciding, rather than around the production tasks that used to fill them.
• Measure the checking, not just the output. Error rates found at review, sources traced, decisions reversed. Output volume will rise on its own.
• Protect the apprenticeship. Juniors learn judgment by working through real problems. Give them the verification and the reasoning, not just a queue of AI drafts to tidy. The wider jobs picture is in the jobs AI will replace.
• Treat context as infrastructure, owned, maintained and portable, not as a by-product of whoever used the tool last.
• Bring people with you. None of this happens by rolling out a tool; it happens by changing how the work is done, with the people who do it.

If your firm’s context still lives in inboxes and people’s heads, tell Daniel what you would write down first, what you have already built, and what the business needs it to do. Book 30 minutes with Daniel.

Frequently Asked Questions

What is knowledge work?

Knowledge work is work whose output is information and judgment rather than physical goods, such as advising, analysing, writing, designing, planning and deciding. The term was coined by Peter Drucker in 1959. Lawyers, accountants, consultants, marketers, analysts and managers are all knowledge workers.

How is AI changing knowledge work?

It is moving the work from producing answers to framing questions, verifying output, integrating it with organisational context and taking decisions. BCG’s 2026 AI at Work survey found nearly half of respondents say their roles have shifted towards managing and directing AI rather than doing the work themselves.

Does AI reduce critical thinking at work?

It can. A Microsoft Research study of 319 knowledge workers found that higher confidence in AI was associated with less critical thinking, while higher confidence in one’s own ability was associated with more. AI also shifts critical thinking towards verification and integration of its output.

What skills matter most for knowledge workers in the AI era?

Four: framing the right question, verifying output against sources and facts, integrating answers with the organisation’s own context, and deciding and taking responsibility for the result. These were always senior skills; they are now needed earlier in careers.

Why does organisational context matter so much with AI?

Because everyone has access to the same models. The difference in output comes from the instructions, examples, client knowledge and standards an organisation gives them. That context is what makes the work specific to the firm, so it should be written down, maintained and kept under the firm’s control.

How should firms organise knowledge work around AI?

Redesign roles around framing, verifying, integrating and deciding; measure the quality of checking rather than output volume; protect how juniors learn judgment; treat organisational context as owned infrastructure; and involve the people who do the work in redesigning it.