An AI agent that answers questions about your company will sound confident whether it is right or wrong. When it does not know a price, a policy or a service detail, it fills the gap with something plausible. That is the most common failure I see, and it has a practical fix.

It is also widespread. In a VentureBeat survey of 101 enterprises in June 2026, 57% had traced an incorrect agent answer to missing or inconsistent business context. Only 25% had a governed context layer in production.

What the agent is missing

An agent needs five kinds of company knowledge to answer well.

  • Facts. Services, products, locations and approved figures, each with a source and a date.
  • Voice. How the company writes, with words to use and avoid.
  • Limits. What the agent must never claim, share or promise.
  • Process. How work gets done and who approves what.
  • Sources. An index that tells the agent which file answers which question.

Most teams give an agent the first item and hope for the best. The other four are what make its answers reliable.

Why a pile of documents is not enough

The common first move is to upload every document the company has. The result is old and new versions side by side, long files that cover many topics and nobody responsible for keeping them current. The agent finds something, and it may well be the wrong version.

A knowledge base works differently. It starts small, keeps one current version of each topic in its own file and gives every file an owner. Answers can name the file they came from, so a person can check them.

What I built

I co-built a shared AI knowledge base for a company's agents and Claude projects. It is 86 interlinked Markdown files, one topic each, with query routing that points each question to the right file. It includes a tone-of-voice standard every agent follows. Projects connect with defined permissions, so some can only read and some can write back approved updates.

The same base supports several agents and daily AI projects. When the base improves, every connected system benefits.

Where to start

List the 20 questions your staff and clients ask most often. Write one short file per topic that answers them, add an index and test the agent on all 20 questions. Every wrong answer points to a file that needs fixing. Fix the file, not the prompt, and test again.

You can build a useful first version in two weeks. Start with the questions people ask most, and tell me what you find.

Download the two-page guide: Give the crew the script (PDF), Guide 5 in the Director's Chair series. It is free to share with your team.