AI in Knowledge Management

In 2012 the McKinsey Global Institute estimated that the average office worker whose job is mostly interaction with other people spent nearly a fifth of the working week looking for internal information or tracking down colleagues who could help¹. Eleven years later, Microsoft surveyed 31,000 people in 31 countries. Sixty-two percent said they struggled with too much time spent searching for information in their workday².

The two figures measure different things. The first is a share of working time, estimated for one group of workers; the second is a share of survey respondents who reported a struggle. They are not a trend line, but they describe the same complaint. Between those two surveys, enterprise search improved considerably. It learned to rank, to read synonyms and to search across many systems at once. The complaint stayed the same. That should make any leader cautious about the promise now being made for generative AI: that employees will simply ask, and the organization’s knowledge will answer. The promise is partly true. AI really does change the economics of knowledge. But the decade of better search suggests the hard part was never only finding things.

Knowledge has to move

Data Strategy for AI made the point that data which exists is not the same as data you can use. Knowledge has the same problem and one more: a great deal of it is not written down at all. It lives in the heads of experienced people, in the reasons behind past decisions and in the habits of teams that do something well.

Knowledge management, stripped of its jargon, is the work of moving know-how from the people and places that have it to the moment someone needs it. Generative AI makes that movement much cheaper. It can turn an interview into a structured note, combine a dozen documents into one answer and deliver it in plain language at the moment of work. Those are capabilities; the clearest evidence of results so far, discussed below, comes from customer support. What it cannot do is decide which of those documents is right, who is accountable for keeping it right, or what to do when two of them disagree.

That split is the argument of this chapter. The rest explains why knowledge is hard to move, what AI can and cannot carry, and the decisions that make an AI knowledge system trustworthy.

Why knowledge sticks

Many leaders assume that when a good practice fails to spread inside a company, the cause is motivation: the not-invented-here reflex, or teams guarding their turf. Gabriel Szulanski tested that assumption. He studied 122 transfers of best practice inside eight companies and found that motivation was not the main barrier. Knowledge itself was³.

The three barriers he identified map closely onto what AI can and cannot do. An arduous relationship, the expert who is in another country, on another shift or simply too busy, is where an assistant helps most: it is always available and never tired of the same question. Low absorptive capacity, a receiver who lacks the background to use what they are given, is also partly addressable, because an assistant can explain the same material at the level of the person asking. Causal ambiguity is different. If nobody can say why a practice works, an assistant will repeat it confidently, including in situations where it no longer applies.

There is already good evidence that AI can carry know-how across the first two barriers. As AI Is Changing Everything described, a support assistant trained on the work of a firm’s best agents raised productivity by 15 percent on average, with much larger gains for the least experienced⁴. The assistant was, in effect, a knowledge-transfer system.

What AI can carry, and what it cannot

As AI vs Automation noted, the philosopher Michael Polanyi put the deepest problem in one line: “we can know more than we can tell”⁵. An experienced formulator knows when a batch looks wrong before any test says so. A veteran planner knows which supplier’s promised date to discount. Much of that knowledge is tacit: real, valuable and hard to write down.

Ikujiro Nonaka and Hirotaka Takeuchi built one of the best-known models of how organizations convert knowledge between its tacit and explicit forms⁶. Their four conversions are a useful map of where AI fits.

Combination, explicit to explicit, is where AI is strongest: summarizing, comparing and merging written sources. Internalization, explicit to tacit, benefits when an assistant explains a procedure at the moment someone performs it. Externalization, tacit to explicit, is where AI is newly useful: it can interview an expert, transcribe the discussion of a hard case and draft the note, though the expert still has to check that the note says what they meant. Socialization, tacit to tacit, still happens mostly between people, through shadowing and apprenticeship. Where AI sits in each quadrant is this course’s judgment of capability, not a measured result.

The practical lesson is that AI pulls knowledge management toward what is written. That is valuable, but it can quietly starve the other quadrants. The risk of losing the expertise that AI cannot carry is the subject of Operational and Workforce Risk, in Module 06.

The knowledge loop

A knowledge base is often imagined as a library: build it, fill it, let people borrow. In practice knowledge behaves more like a loop, and a system that only stores will decay.

AI helps at every step. It drafts the record of an expert interview, suggests tags and classifications, finds the passage that answers a question and shows it at the moment of work. Its least appreciated contribution is in the last step. Every question the system could not answer, or answered and then had corrected, is a signal of a gap. A knowledge team that reviews those questions every month knows exactly what to capture next.

One distinction deserves care in the capture step. A meeting summary is a convenience; a decision record is an authority. AI now produces the first in seconds. If a summary is going to be treated as the record of what was decided, a person has to check it and the right owner has to file it. Without that step, organizations accumulate confident accounts of decisions that were never quite made.

Which source may answer

How a grounded assistant works, retrieving relevant and permitted passages at question time and answering from them with sources, is explained in RAG and Enterprise Knowledge — Executive Mental Model. The knowledge-management question sits one level above the technology: which source is allowed to answer a given kind of question?

The answer is an organizational decision, not a feature of the model. For each domain that matters, someone must name the authoritative system, the person accountable for keeping it correct and current, how often it is reviewed and how old versions are retired. The rhythms in the table are examples; each organization sets its own. None of this is new. Data management has long treated ownership and stewardship as core disciplines⁷, and knowledge management has had an international management-system standard since 2018, which treats roles, responsibilities and review as management duties⁸.

What AI changes is the cost of skipping the decision. A search engine that returned five versions of a procedure left the reader to choose. An assistant that blends them into one fluent answer hides the choice. An assistant must also never become a new, unofficial source of truth: if it gives an answer that no authoritative source contains, that answer has no owner. Who may see what is the other half of the decision; permission-aware retrieval belongs to the RAG chapter, and the oversharing risk to Privacy and Confidential Data.

When sources disagree

Some answers do not sit in any one document. The assistant has to combine several, which is more useful than search and also riskier, because sources disagree. Suppose one procedure says a deviation is approved by the plant manager and another says it goes to quality.

Among the most dangerous outputs a knowledge system can produce is a silent choice: one version picked, presented fluently, with no sign that another existed. The person asking cannot know that a decision was made on their behalf. A well designed system does the opposite. It shows both sources, says that the answer is uncertain, and sends a review task to whoever owns the domain. High-risk topics, such as safety, legal or medical questions, can follow the same path even when the sources agree.

Each reported conflict is a piece of what might be called knowledge debt, inconsistency the organization always had but could not see. Grounding reduces wrong answers without removing them, as Accuracy, Hallucination and Reliability explains. Conflict handling is how a knowledge system turns the remaining errors into work someone can do.

Measure the task, not the search count

Knowledge programs are prone to vanity metrics: documents connected, searches run, questions asked. Those numbers show activity. They say nothing about whether anyone did better work.

Better measures follow the task: time to a correct answer, the share of answers accepted without correction, the number of conflicts closed by owners and the share of authoritative sources reviewed on schedule. Suppose an employee spends thirty minutes finding what a task needs and the assistant cuts that to three. The twenty-seven minutes are a gain only if they go to better work, which is the argument of Productivity vs Realized Capacity.

Where to start follows from the same logic. Frequent, frustrating, lower-risk questions in a domain that already has an authoritative source make the best first case. Be wary of letting the system create policy, decide which source wins, publish knowledge nobody checked or change who may see what. An agent can find every document affected by a change and open review tasks; a person approves. AI Agents and Intelligent Workflows, later in this module, takes that idea further.

Story: the probate assistant that took fifteen months

When someone dies in Alaska, settling their estate goes through a court process called probate, with its own forms, deadlines and procedures. The Alaska Court System already had human facilitators who explained that process to the public. With a grant from the National Center for State Courts, it set out to build an AI assistant that could answer probate questions at any hour: the Alaska Virtual Assistant, or AVA. It was meant to be a three-month project⁹.

The scope was narrow from the start: one domain. The team restricted the assistant to the court’s own probate documents and self-help materials, rather than letting it search the wider web. Even so, it invented things. Asked where to find legal help, an early version suggested the alumni network of a law school in Alaska. The state has no law school⁹.

Testing turned out to be the expensive part. The team wrote 91 test questions, then found that running them and having a person judge every answer took too long for the stakes involved. It settled on a short list of about fifteen, mixing questions the assistant had answered wrongly, hard questions and basic ones [@nbc-alaska-ava-2026; @courtleader-ava-2026]. Users shaped it too. People in mourning said they were tired of being told how sorry everyone was for their loss, so the team took the condolences out⁹.

The goal itself moved. The court’s administrative director said the team had wanted the assistant to replicate what its human facilitators share with people, but was not confident a bot could work that way, because of inaccuracies and incompleteness⁹. AVA was released in January 2026¹⁰. Its developer expects regular checks after launch, and changes to prompts or models as new ones arrive, rather than a hands-off system⁹.

Set the case against this chapter. The technology was cheap: one configuration cost about 11 cents for 20 questions⁹. The months went on the decisions the chapter describes: one domain, one authoritative source, people who judge whether answers are right, and someone who keeps checking after launch. What the facilitators know from years of conversations with families, the tacit part, did not move into the assistant. A leader who budgets three months for the tool and nothing for those decisions has budgeted for the wrong thing.

What this means for leaders

Four lessons follow. First, decide authority before scale: no assistant should go live in a domain until someone has named the authoritative source and its owner. Second, design for disagreement: require the system to surface conflicts and route them, and treat the resulting queue as an asset. Third, protect what AI cannot carry: pair capture with apprenticeship for the expertise that matters most, and start it long before people leave. Fourth, measure trusted outcomes: report time to a correct answer and conflicts closed, not questions asked.

Check yourself

1. Better enterprise search has largely solved the problem of finding internal information.MythFact
2. The main barrier to spreading a good practice inside a company is usually motivation.MythFact
3. AI is most capable at combining knowledge that is already written down.MythFact
4. An AI-generated meeting summary can serve as the record of a decision without review.MythFact
5. When two sources conflict, a good assistant should pick the more recent one and answer confidently.MythFact
6. Questions asked is a weak measure of a knowledge assistant’s value.MythFact

Reflection: where does your knowledge stick?

What comes next

Much of what organizations know arrives as documents and data: contracts, invoices, forms, reports and records. The next chapter, AI in Document and Data Processing, looks at how AI can read, extract, classify and check that flow at scale, and where people still need to stay in the loop.