A new account manager is preparing a renewal and asks the company assistant a simple question: can we offer quarterly billing to a customer on the enterprise plan?
The assistant does not know. It finds the published pricing page, which says enterprise plans are billed annually. It finds nothing else.
Ask anyone on the sales team and you get a different answer instantly. "Just ask Dana in finance." Dana knows. Quarterly billing has been approved for a handful of accounts under specific conditions, and Dana has explained those conditions in direct messages, on calls, and in quick replies more times than anyone has counted.
Every one of those answers was correct. None of them landed anywhere the assistant, or the next account manager, could find.
Dana is not the problem. Dana is the most helpful person in the building. That is exactly why the answer is invisible.
The Most Helpful People Leave the Fewest Traces
Every company has a few people like Dana. The senior engineer who knows why the payment service retries the way it does. The support lead who knows which customers have custom terms. The operations manager who knows which vendor contract has the unusual renewal clause.
They are generous with what they know. Someone asks, and they answer, fast and accurately. That generosity is a big part of why the company works.
It also has a side effect. When a question gets answered in a private message, on a call, or at someone's desk, the answer solves one person's problem once. Then it disappears. The next person with the same question asks again, and the expert answers again.
From the inside, this feels efficient. The expert is quick, the person asking is unblocked, and nobody had to write a document. From the outside, it means the company's most valuable knowledge lives in the one place no system can reach: a few people's heads, and the conversations they had one at a time.
Where Expertise Actually Goes
It is worth being specific about where knowledge goes when it does not go into documentation, because it is rarely nowhere.
Private conversations. Direct messages and small group chats are where a lot of real answers happen. They are fast and low pressure, and they are visible only to the people in them.
Calls and meetings. A decision gets made on a call. Everyone present understands it. The meeting notes, if there are any, record that a decision was made but not the reasoning behind it.
Outcomes without reasons. A ticket is closed as "won't fix." A pull request gets merged after a conversation that happened somewhere else. A contract term gets changed. The system records what happened. The why lives in someone's memory.
Corrections that never travel back. An expert notices that the wiki page is out of date, answers the question correctly in a message, and moves on. The page stays wrong.
None of this is negligence. It is how busy, capable people naturally work. But it means the written record of a company is consistently thinner than what the company actually knows, and the gap is widest exactly where the expertise is deepest.
Why AI Cannot Fill the Gap
An AI assistant working on company questions can only answer from what it can read. If the answer to the quarterly billing question exists only in Dana's direct messages and Dana's memory, the assistant has two honest options and one dishonest one.
It can say it does not know. It can point to what it did find, the published pricing page, and make clear that it is incomplete. Or it can produce something plausible from the general pattern of how billing usually works, and present it with the same confidence as a real answer.
The first two are useful. The third is how a team learns to stop trusting the assistant.
This is not a flaw in any particular model. The most capable model available still cannot use information that was never written down somewhere it can reach. Swapping models does not change that. Neither does connecting more systems, if the systems you connect do not contain the answer. As we argued in Why More Context Can Make Your AI Worse, coverage and context are different things. Here the problem comes before both: the knowledge never entered a system at all.
This Is Not a Documentation Project
The usual response is a documentation push. Ask the experts to write everything down. Stand up a new wiki. Schedule time for knowledge transfer.
These efforts tend to fade, and not because experts are unwilling. Writing a complete document about something you know well is slow, it is never quite finished, and it competes with the work that made the person an expert in the first place.
There is a smaller change that works with how experts already behave. Keep answering questions, but answer them in a shared place. A team channel instead of a direct message. A comment on the ticket instead of a call. A short note on the pull request about why the change was made.
An answer written in a shared channel is documentation. It is informal and it is short, but it is real, it carries the expert's name, and it can be found by the next person. If the AI is connected to that channel, it can be found by the AI too, with a citation pointing straight back to the expert who wrote it.
Dana does not need to write a billing policy manual. Dana needs to answer the next quarterly billing question in the finance channel instead of a direct message.
How Modly Handles It
Modly connects the systems a company already works in, such as Slack, Jira, Confluence, GitHub, Salesforce, and Google Drive. That matters here because the place where experts are most likely to leave a trace is not a formal document. It is a thread, a ticket comment, or a pull request discussion.
Answers cite their sources. When the account manager asks about quarterly billing, and Dana has answered that question in a connected channel, the answer links to that thread. The person asking can read Dana's actual words, check the conditions, and see who to follow up with. The expert's knowledge gets reused instead of re-asked, and the credit stays with the person who knew.
Modly says "I don't know" instead of guessing. When the sources do not support an answer, Modly says so rather than filling the gap with something that sounds right. That keeps the assistant trustworthy. It also makes the gaps visible. When the same kinds of questions repeatedly produce "I don't know," that gives the team something useful to investigate. The answer may exist somewhere Modly cannot reach, or it may only live in someone's head. Either way, the gap is visible instead of being covered by a plausible guess.
Permissions travel with the knowledge. Moving answers into shared places does not mean making everything public. An answer posted in a restricted channel is only used for people who are allowed to see that channel. Each person gets answers built from the sources they can already access, on every query.
Context Profiles put the right threads in front of the right teams. A sales profile can include the finance channel where billing questions get answered, so account managers benefit from Dana's answers without anyone copying them into a sales wiki. We covered how profiles work in Context Profiles: Why AI Shouldn't Treat Everyone the Same.
None of this extracts knowledge from anyone's head. What it does is make every answer an expert writes down in a connected place count for far more than one conversation.
What Changes
Go back to the account manager and the quarterly billing question.
Suppose the next time someone asks Dana, the answer goes in the finance channel. A few sentences: when quarterly billing has been approved, what conditions applied, who signs off.
The time after that, the account manager asks the assistant instead of messaging Dana. The answer comes back with the conditions and a citation to Dana's thread. The account manager reads it, sees that it applies, and moves on. Dana never gets the question.
If the situation is unusual, the citation also tells the account manager exactly who to ask. The expert is still the authority. They just stop being the only way in.
That is the shift. The knowledge your AI needs is mostly not missing. It is being given away generously, one private conversation at a time. Your best employees are not holding your AI back on purpose. They are doing what helpful people do. A small change in where they answer is often worth more than a model upgrade.
This post is about knowledge that stays in people's heads while they are still here. The next one is about what happens when they leave.
If you want to see how cited, permission-aware answers work on your own systems, we run a 30-minute walkthrough on real data: Book a demo.
Related reading:
