Why Does My AI Miss the Way We Actually Do Things?
Because the way you actually do things was never written down. What I keep finding is that the most valuable context in any company, the know-how that separates a good outcome from a bad one, lives only in people's heads. It is the unwritten judgment: how you really handle the difficult customer, why you make this exception but not that one, the shortcut everyone knows but no document mentions. Your AI misses it for a simple reason. You never gave it to anyone, including the new humans you hire. To make agents genuinely useful, you have to capture the knowledge nobody wrote down, and that is a solvable problem most companies have simply never tackled.
I am Mahmoud Zalt, an AI architect running Sistava, where autonomous agents do real business work in production. The single biggest difference between an agent that feels competent and one that feels clueless is whether it has this tribal knowledge. Here is how I think about capturing it.
What Tribal Knowledge Actually Is
Tribal knowledge is the operating wisdom of a company that lives in practice, not in documents. Every experienced team has an enormous store of it, and almost none of it is written anywhere. It shows up as the things people just know.
- The exception you always make for a certain kind of case, that no policy captures.
- The reason a rule exists, so you know when it is safe to break.
- The tone that works with your customers and the one that backfires.
- The step everyone skips because it never mattered, and the one nobody skips because it once caused a disaster.
- Who to ask, what to check, what usually goes wrong at this point.
This is the difference between someone in their first week and someone in their fifth year. Both can read the same documentation. Only one has the tribal knowledge, and that is why only one is trusted with the hard cases. When you connect an AI to your documented knowledge alone, you have effectively hired a very fast, very confident first-week employee. The documents tell it what you are supposed to do. The tribal knowledge tells it what you actually do, and it does not have that yet.
Why It Never Got Written Down
It helps to understand why this knowledge is missing, because the reasons tell you how to capture it. Tribal knowledge stays unwritten for three honest reasons.
First, it is invisible to the people who have it. When you have done something a thousand times, the judgment feels like common sense, not knowledge worth recording. Experts routinely cannot list what they know because they are not aware they know it. Second, it is contextual, not general. It is a hundred specific little rules tied to specific situations, not a clean principle you would put in a handbook. It resists being written as policy because it is not policy, it is accumulated pattern-matching. Third, nobody was ever asked. Documentation captures the official process because that is what documentation is for. The unofficial reality, the part that actually makes things work, was never the target of any writing effort.
None of these are laziness. They are structural. Which means capturing tribal knowledge is not about telling people to document better. It requires a different method, one designed to pull out knowledge the holder does not know they have.
How To Actually Capture It
The method that works is not a documentation drive. It is closer to an interview, and it targets the gap between the official process and the real one. Here is the approach I use.
- Sit with the person who does it best. Not the newest, not the manager, the one everyone quietly routes the hard cases to. Their head is where the tribal knowledge concentrates.
- Walk through real cases, especially the weird ones. Do not ask them to describe the process in general, they will just recite the official version. Ask them to walk through specific past cases, particularly the exceptions and the ones that went wrong. The knowledge lives in the specifics.
- Chase every it depends. Whenever they say it depends, stop and dig. On what does it depend? That dependency is a piece of tribal knowledge becoming explicit. Most of the value is hiding behind those two words.
- Capture the why, not just the what. Record why the exception exists, why the rule can bend here, what the failure looked like that created this caution. The why is what lets an AI generalize instead of just memorizing.
- Feed it in as context, then test on hard cases. Add the captured knowledge to what the AI can draw on, then run it against the same difficult cases the expert walked you through. The gap between its answers and theirs shows you exactly what tribal knowledge is still missing.
Done well, this single exercise often does more for AI quality than any model upgrade, because it hands the machine the exact context that separated your best people from your newest ones.
The Side Benefit: You Own Your Own Knowledge
There is a payoff here beyond making AI work, and it is worth naming because it changes how seriously leaders take this. When you capture tribal knowledge to feed your agents, you are also, for the first time, writing down the operating wisdom of your company. That knowledge stops being trapped in a few people's heads and starts being an asset the company actually holds.
Think about what that protects against. Today, when your best person leaves, their tribal knowledge walks out with them, and you rebuild it slowly and painfully in whoever replaces them. Capturing it for AI means you capture it, period. It survives departures. It onboards new humans faster, not just agents. The AI project becomes the forcing function that finally gets your company's real know-how out of individual heads and into something durable. That is a strategic win that outlasts any particular model you connect it to.
Frequently Asked Questions
Why does my AI not know how my company really operates?
Because the way you really operate is tribal knowledge that was never written down. Your documentation captures the official process, but the judgment that makes things work, the exceptions, the unwritten rules, the hard-won cautions, lives only in experienced people's heads. Connecting AI to documents alone gives you a fast, confident first-week employee. To make it competent, you have to capture and feed it the unwritten know-how.
How do I capture tribal knowledge for an AI system?
Treat it as an interview, not a documentation drive. Sit with the person everyone routes hard cases to, walk through specific real cases including the exceptions and the failures, and chase every it depends until the hidden dependency becomes explicit. Capture the why behind each rule, not just the what. Then feed it to the AI as context and test it on the same hard cases to find what is still missing.
Why was this knowledge never documented in the first place?
Three structural reasons. It is invisible to experts because long practice makes judgment feel like common sense. It is contextual, a hundred situation-specific rules rather than a clean principle that fits a handbook. And nobody was ever asked, because documentation targets the official process, not the unofficial reality that actually makes things work. That is why capturing it needs a different method than just telling people to write more.
Is capturing tribal knowledge worth it if we might change AI models later?
Yes, and this is the point. The captured knowledge is independent of any model. It is your company's operating wisdom, finally written down. It improves whatever AI you connect it to, and it also survives employee departures and speeds up onboarding for new humans. The model you use will change. The captured know-how is a durable asset that keeps paying off regardless.
Give the Machine What Only Your People Know
The market keeps framing better AI as a matter of better models. But in practice, the gap between an agent that feels competent and one that feels clueless is almost never the model. It is the tribal knowledge, the operating wisdom your best people carry and never wrote down. Capture that, and an ordinary setup starts producing work that feels like it came from your most experienced hand. Skip it, and the smartest model available still answers like it is on its first day.
Two takeaways. First, treat tribal knowledge capture as core to any serious AI effort, and do it with interviews of your best people, chasing the exceptions and the it-depends, not with a generic documentation push. Second, recognize the double payoff: you make your AI genuinely useful and you turn your company's know-how into an asset that outlives both the people who hold it and the model you happen to use.
If you want help pulling the knowledge out of your team's heads and into a form your AI and your future hires can actually use, that is exactly the work I do. Let us capture what makes your business work. More at Sistava.
