Generative AI for knowledge management: a plain guide

What knowledge management means

• Written

  Policies, manuals, proposals and tickets

• Recorded

  Calls, meetings and training videos

• Tacit

  What experienced staff know but did not write down

Knowledge management is how a company stores, finds and shares what it knows. That includes policies, past projects, customer notes, procedures and the know-how in people's heads.

Most of it sits in shared drives, email threads, chat history and wikis nobody updates. The problem is rarely a lack of knowledge. It is that people cannot find it when they need it.

What generative AI changes

Traditional search matches keywords and returns a list of files. You open each one and hunt for the answer yourself.

Generative AI for knowledge management works differently. An employee asks a question in normal words, and the system reads the relevant documents and writes a direct answer, pointing to the source it used.

The value is not clever writing. It is that the answer comes from your own material, so it reflects how your company actually works.

The cost of searching for information

Searching is a quiet drain. McKinsey research, as cited by Agility Portal, puts it at around 20% of a knowledge worker's workweek.

A separate McKinsey figure, cited by Cottrill Research, says employees spend 1.8 hours a day, or 9.3 hours a week, searching and gathering information.

A Pryon report, cited by Cake, found that 47% of professionals spend 1 to 5 hours a day looking for specific information.

These are vendor and secondary sources, so treat them as a sense of scale, not a precise measure for your company.

Common ways teams use it

• Policy answers

  Procedures explained in plain words

• Past work

  Find earlier proposals and project notes

• Summaries

  Long documents and meetings made short

• Onboarding

  New staff learn without interrupting others

• Drafts

  Replies built from approved content

The easiest first use is an internal question-and-answer assistant.

Staff ask how to handle a refund, what the travel policy says or how a past project was priced, and get an answer with the source attached.

The same approach supports new hires, who can ask questions without interrupting a colleague, and customer-facing teams who need accurate answers during a call.

Capturing what retiring staff know

The Retirement Income Institute, as cited by APQC, says more than 11,000 Americans will turn 65 every day from 2024 through 2027. That is roughly 4.1 million people a year reaching retirement age.

APQC's research found that 84% of senior leaders worry about knowledge loss from retiring employees, yet only 8% of organizations consistently capture it.

AI can help close that gap. It can interview experts with guided questions, turn recordings and notes into written procedures, and file them where others can find them.

The expert still reviews the result, because tacit knowledge is easy to get subtly wrong.

Keeping content current

An assistant is only as good as the documents behind it. Old policies produce confident, outdated answers.

Good practice is to give every document an owner and a review date.

AI can help by flagging pages that contradict each other, noting files nobody has opened in a long time and suggesting updates when a process changes.

APQC's 2025 benchmarks found that 38% of KM programs assign a dollar value to each success story. Tracking value this way helps justify the effort of upkeep.

• Name an owner for each document set
• Set a review date and stick to it
• Flag conflicting or stale pages
• Retire old versions so they stop appearing in answers

Risks to plan for

• Wrong answers

  Require sources and an honest I don't know

• Access control

  Answers must follow existing permissions

• Messy sources

  Clean the core documents first

The first risk is wrong answers. Generative AI can state something false with confidence. Requiring the assistant to cite its source, and to say when it cannot find an answer, reduces this.

The second is access control. If the assistant can read every file, it can repeat salary data or legal notes to the wrong person. Answers must respect the same permissions staff already have.

The third is quality of the source material. Messy, duplicated or contradictory documents produce messy answers. Cleaning up the core set first is part of the work.

How to start

Do not try to connect every file at once. Pick one team and one document set, such as HR policies or support procedures, and gather the ten to twenty questions people ask most.

Build a small assistant over that set, test it against the real questions and have the people who know the answers check the results. If it is accurate and people use it, widen the scope.

If you want help scoping a first build, our custom AI development work covers assistants built on your own documents.

For the wider picture, read the guide to AI for business or see how AI for operations fits in.