When the custom GPT stops feeling magical
You set up a custom GPT (or similar) for your team. You upload the policies, the procedures, the org chart, the playbook, the onboarding deck, whatever seems like useful context. You ask it a few test questions. It answers reasonably well. A few people are impressed. Some early wins get shared in Teams.
Then someone asks it a question that actually matters.
“Why did we stop selling to that segment?”
“How do we usually handle it when a client asks for X?”
It fumbles. It gives a confident, plausible answer that’s slightly wrong, or generic, or missing the thing everyone in the room knows but nobody ever wrote down.
The instinct is to fix it. Better prompts. More documents. Tweak the instructions. Maybe a different model. But the gap isn’t a settings problem.
What the documents don’t hold
Most of an organization’s knowledge lives in people’s heads, not in any repository. The exact share is hard to measure, but it is large. The documents don’t hold the answers because most of the answers were never in any document.
They live in the decision your leadership made three years ago that quietly reshaped how the sales team qualifies. They live in the unwritten rule that you don’t escalate to that client’s CFO without a heads-up to her assistant first. They live in the meeting where someone said “we’re not doing that anymore” and everyone just stopped doing it. No memo. Just a shift in how the work gets done, carried in the heads of the people who were there.
That’s institutional knowledge. It is the residue of old decisions, relationships, and hard calls. Most of it is practically useful and genuinely hard to put into words, which is exactly why nobody has.
And even the documents you do have are thinner than they look. A policy says what we do. It rarely says why we do it that way and not the other way. A procedure tells you the steps. It doesn’t tell you which step is the one people skip, or which step is most important when something goes wrong. The custom GPT reads it all literally. It doesn’t know what’s stale, which paragraph everyone treats as gospel, or that the “approval process” on page four is what legal wishes happened, not what actually happens.
People in the company know. They work around the gap between the written version and the real version every day, mostly without noticing. That gap is where the institutional knowledge sits. The custom GPT has no access to it.
This is why “just upload more documents” doesn’t fix the problem. You can double the corpus and the same gap remains. More volume of the wrong input doesn’t get you to the right output. The reasons behind the work mostly aren’t on paper.
Gartner expects organisations to abandon 60% of AI projects through 2026 for lack of AI-ready data. In operations, “AI-ready data” mostly means documented reasoning, not the undocumented know-how stuck in employees’ heads. The projects don’t fail because the model was wrong. They fail because the inputs were never built.
Where the answers actually live
If you want AI to be useful past the demo phase, the real work is harder and slower than uploading files.
It looks like sitting with the people who do the work and capturing the reasoning behind the choices. Asking your best account manager what she does when a renewal goes sideways: how she chooses one option, and why she avoids another option the playbook technically allows. Writing down the things that feel too obvious to write down, because “too obvious to write down” is exactly where the institutional knowledge hides.
This is also where AI governance leadership actually shows up. Most of what gets filed under an AI acceptable use policy is a list of what employees can’t paste into a chatbot. Useful, but it’s the smaller half of the job. The bigger half is owning the question of what the company actually knows, where that knowledge lives, and how it gets into the tools people are already using. That’s the part leaders can’t delegate to IT or legal.
The awkward part: the people who hold the most institutional knowledge are usually the same people already drowning in everything else. The tacit knowledge lives with your most overloaded people, which is part of why it stays tacit. Nobody has time to sit with them and pull it out. So it stays in their heads, gets a little more out of date every quarter, and walks out the door when they leave.
The custom GPT didn’t stop feeling magical. It just ran out of the only thing it had to work with.
The better question is not whether to change models, switch platforms, or build an agent. It’s whether the actual operating knowledge of your company exists anywhere outside the heads of the people running it. For most organizations we talk to, the honest answer is: barely. Not in any form an AI tool, a new hire, or a successor could pick up and use.
That’s the project. The custom GPT just made it visible.
Categories
Most Recent Posts
To get SharePoint ready for Microsoft 365 Copilot, fix who can see what before you do anything else. Copilot shows…
Bring AI into your business and you’ve hired a stranger. A bright one, fast and eager to help with whatever…











