The Knowledge Problem AI Cannot Solve Alone

Most discussions about AI at work stop at the moment the answer arrives. My concern starts there.

In 2008 I finished a doctoral study on how companies create knowledge and turn it into new products. One result has stayed with me. In that data, formal rules were associated with better use of what a company had already written down, and with greater difficulty in getting people to put into words what they only knew from experience. The same structure looked helpful for one part of the process and unhelpful for another.

I return to that finding whenever I think about AI. A company can become very good at working with what it already has on record while still learning very little from what its people know.

Large language models make this more pressing, not less. They help us prepare an analysis, compare options and produce a solid first draft. None of that tells us whether the organization is learning.

From Data to Knowledge

A simple example, to be clear about the terms.

That last step needs evidence the numbers cannot supply — exit conversations, promotion histories, what managers have noticed. Once we have it, we can decide what to change, and then find out whether the change worked.

This is where organizational learning happens: people interpret, argue, act, and see the result. Some of that can be written down. Some of it stays in the experience of the people who were there.

What the Data Showed

My study used survey responses from 294 professionals in 93 IT companies in Turkey. It built on Ikujiro Nonaka's model of knowledge creation and on the framework of Heeseok Lee and Byounggu Choi, and it examined practical conditions — trust, learning, formal rules, information technology — alongside new product development capability and perceived company performance.

Three results matter here.

First, learning and IT were both positively associated with all four knowledge creation processes: sharing experience, putting experience into words, combining what is documented, and learning through use.

Second, in the final model only one of those four processes had a significant direct path to new product development capability: combination, the organizing and joining of knowledge that is already written down (0.34). Socialization showed a separate, direct association with qualitative performance instead. The two sides of knowledge creation appeared to pay off through different channels.

Third, formalization was associated with two of those processes in opposite directions: positively with combination (0.28) and negatively with putting experience into words (−0.12). That does not tell us whether the overall effect was beneficial. It tells us to look at what each procedure helps people do, and what it makes harder.

Some limits. This was a cross-sectional survey, in one sector, in one country, with performance reported by the respondents themselves. The accepted model was reached by revising an initial specification that did not fit, so it is better read as exploratory than as a confirmed causal structure. And it was done long before any of the tools we are now arguing about existed.

It is still a useful place to stand. Technology mattered. So did the conditions in which people used it. Both belong in a conversation about AI.

What AI Makes Easier

Long before I did that research, I was working at a bank in its early years, growing fast and meeting suppliers constantly. One of them showed us a platform. We looked at it closely, decided the fit was wrong for us, and wrote up why.

Some months later the executive responsible for operations moved on. His successor met the same supplier, was impressed, and asked me to evaluate a possible partnership and report back to him.

Nobody had done anything wrong. The note existed. It sat with the people who had been in the room. The company had reached a judgement and then had no way to bring it to the next decision.

This is the kind of problem AI is genuinely good at. Had that note been in a shared, searchable system rather than in the participants' own files, the second evaluation could have started from what we already knew. I want to concede that clearly before I argue the other side.

The productivity evidence points the same way. In the 2023 working paper Generative AI at Work, researchers studied a tool used by customer support agents and reported an average productivity gain of 14%, with larger gains among less experienced and lower-skilled workers.1 AI can spread practices that used to depend on years of exposure.

In the language of my own field, this is combination getting cheaper and faster — and combination was the one process in my data with a direct link to product development capability. That is worth sitting with. So is the fact that formalization was associated with the same process in the same direction, while being negatively associated with the other.

What AI Does Not Reach on Its Own

Years later I joined a payments company where several of the teams reporting to me had been almost entirely rebuilt. On the POS side, the longest-serving person had been there eight months. The card director left shortly after I started.

We had the co-branded card contracts in the system. What we did not have was the negotiation behind them: what our partners had asked for, what we had refused, and why. During those talks our IT and fraud colleagues had raised concerns out loud in the room. Some of those concerns were never documented.

One person was left who carried all of it. When a similar project came in, he was the reason we did not reopen a question the company had already settled, and the reason we saw certain risks coming before they reached us. Working from the contracts and the product specifications alone, we would have had the terms without the reasoning. That understanding existed nowhere except with him.

The two episodes exposed different gaps. At the bank, we had recorded our reasons, but they did not reach the next decision. At the payments company, important reasons had never been recorded at all. In both cases the next team needed more than the final decision. It needed to understand why.

AI could help us search and compare the contracts. But the contracts alone could not reveal what we had refused, or why our colleagues had raised concerns. Those reasons had been discussed and never written down. AI could even help capture them, by supporting a conversation with the person who still remembered. Someone had to recognize that the conversation was needed.

I would be careful about turning any of this into an argument against procedure. Procedures let people find things, fix responsibility and avoid repeating mistakes. The difficulty is narrower: a procedure built for one purpose is often poorly suited to another. A standard project template makes reports comparable and leaves no line for the observation that does not fit the headings. An approval step establishes accountability and makes people less willing to voice a concern they cannot yet prove.

So the question for managers is not how much structure to have. It is which structures help people use what is already known, and which ones leave room for something not yet documented, agreed, or fully understood.

The Risk of Doing Less Thinking

I also wonder how people will build judgement when more of the preparation is done for them.

Working through a problem takes time, and some of that time is wasted. Some of it is how we learn to notice weak evidence, ask a better question, or see why an attractive answer will not survive contact with a client.

A 2025 study by Lee and colleagues surveyed 319 knowledge workers. Higher confidence in generative AI was associated with less reported critical thinking, and critical thinking itself shifted toward verifying information, integrating responses and overseeing the task.2 That is not proof of lasting damage to anyone's judgement. It is a reason to look closely at how the work is changing.

What concerns me is a routine in which a reasonable-looking answer is accepted with little examination, because nothing in the week rewards examining it. If that becomes normal, where do less experienced colleagues practise? How does anyone find out that an accepted explanation was wrong?

AI can support the opposite habit. A team can ask it to challenge an assumption, argue the alternative, or name the evidence that is missing. Whether people do that depends on what the organization expects of them. If the only visible target is speed, reflection is the easiest thing to skip.

Decision Architecture

What I am describing is a company's decision architecture: how information reaches a judgement, who is able to object before it hardens, and whether the result comes back to the people who decided.

AI changes the first part of that quickly. It does not change the rest automatically.

What I Would Ask Leaders to Change

Look at decisions, not adoption. The number of people using AI tells you something about deployment and almost nothing about the quality of the work. For a small number of important decisions, record the assumptions, the expected result and the main uncertainties, and set a date to look back. A good outcome does not always mean the reasoning was sound, and a poor one does not always mean it was not. Review both. Over time this becomes a record of how the organization thinks.

Write down the refusals — and make sure they are found. Before an AI-assisted recommendation is approved, bring in the people closest to the work and ask what the analysis has missed and what they find hard to explain. Then capture what was rejected and why, not only what was agreed. Recording it is half the job; the other half is making sure it surfaces the next time a similar question comes up, rather than sitting in the files of whoever attended. There also has to be room for the incomplete observation — I cannot prove this yet, but something is different. It may be wrong. It may also be the start of something worth knowing. Keep the controls that serve a purpose and look hard at the ones that produce paper without producing understanding.

Protect the practice of judgement. Ask people to say why they support a recommendation, which alternative they rejected, and what evidence would change their mind. This matters most for less experienced colleagues: if AI produces the first draft, they still need somewhere to examine reasoning and make a call. And managers have to make time to review outcomes, or people simply move from one finished task to the next without ever learning whether they were right.

What Stays With the Organization?

My research began with a plain management question: where should a company put limited resources to improve how it creates and uses knowledge? AI has made that question more urgent without answering it. It offers real capability. Trust, structure, learning and accountability remain the work of the people running the place.

I would judge an AI initiative by what stays with the team after the task is done. Can they explain the decision? Do they know its limits? Have they learned something they will use next time?

Return to the turnover example. Producing a better retention plan is useful. Understanding why people leave, testing a response, and passing on what happened gives the company something it can use again.

That is the outcome I would want an AI investment to produce.