Writing · essay
What goes into an AI knowledge base for employees
An AI assistant's answers are only as good as what it reads. How I build an AI knowledge base for employees that stays true after launch, with an owner for every page and a way for staff corrections to stick.
An AI knowledge base for employees is the set of pages an internal AI assistant reads before it answers anyone. A lot of my time this week went into exactly that, because I've found an assistant's answers are only ever as good as what it's been given to read.
People tend to focus on the model. The model matters less than they think. Put a strong model on top of a messy, contradictory pile of documents and you get confident answers that are wrong in ways nobody catches.
What an owner decides before launch
Before anything gets written, a couple of calls belong to the owner. Which document wins when two disagree, and what the assistant should never answer, like pay or anything personnel related. Those decisions shape every page after them, and nobody below the owner can make them for the business.
Written from how things actually run
Old procedures often describe how a process was supposed to work years ago. I write the knowledge base from the running systems and from the people doing the work today, and I explained that approach in write the owner's manual from the running system.
It also captures the steps everyone does but no document mentions, which are exactly what new employees ask about.
Every page has an owner
Each entry has a named owner. When a process changes, updating the page is part of that person's job, and the assistant's answers follow automatically.
It's worth a simple check that compares the knowledge base against the live systems where it can. When the two drift apart, somebody hears about it before an employee does.
When staff correct it
After launch, people will correct the assistant. That's the best thing that can happen, as long as the correction goes somewhere. In the systems I build, a correction goes to the owner of that page first. One person's fix might be right, or it might be how their team does it differently, and the owner is the one who can tell. Once they confirm it, it becomes a permanent part of the knowledge base, and the assistant answers it right from then on. I wrote about that habit in every correction should become a permanent rule.
Over time its answers start to sound like the person everyone already asks.
Where an AI knowledge base for employees fits
The knowledge base sits underneath everything else. It's what an internal AI help desk answers from, and it's what keeps a reviewer's time on new questions.
If you're planning an AI assistant for your team, I'd spend the first weeks on the knowledge, well before you pick a model. The model will change a few times over the next few years. The knowledge base carries over every time. If you'd like to see how I approach it, my work shows the kinds of systems it sits under.