In our article “Building Faster Isn’t the Same as Building Trust Faster,” we mentioned a spreadsheet with a cell that no one dared touch anymore. No one knew exactly why it was there, but as soon as it disappeared, the numbers no longer added up. That pattern deserves a deeper discussion.
Every organization has spots like this. A report that’s been structured the same way for years. An exception rule that was once added for one specific client. An assumption that made sense when it was made but was never written down. As long as the person who came up with it is still on staff, it goes unnoticed. Only when that person leaves does it become clear just how vulnerable the situation is. You can train someone new relatively quickly, but being capable isn’t the same as being in the know. The new person learns what the rule does, but rarely why it was ever devised.
With AI systems, we see the same pattern repeating itself—only faster. There’s a specific reason for this: whereas a human had years to pass on knowledge informally by working alongside others and tackling problems together, an AI system is set up in days or weeks. There is simply less time for knowledge to be transferred organically.
A common objection we hear is: aren’t prompts and chat histories automatically saved, and isn’t that already documented reasoning? To some extent, yes, but a chat log records what was asked, not why that question was the right one or which alternatives were deliberately rejected. Without someone to organize that information afterward, a pile of prompts is just as opaque as a pile of old emails.
Standalone documentation isn’t enough on its own. It becomes outdated and gets overlooked when time is tight. More effective is ownership at the level of definitions, not of individuals. A business rule or KPI definition should have an owner who is responsible for its meaning—for example, a data steward per domain, or a glossary with a single definition and a single name for each term. If the owner changes, the responsibility changes with them—not the knowledge, which is lost.
In addition, it helps to have people who aren’t immersed in your systems on a daily basis take a critical look at them regularly. Those who are right in the thick of things no longer ask the questions that have become too familiar.
Finally: build for portability from the start, not as a cleanup effort afterward. Consider lineage information or a model card that describes assumptions and limitations, linked to the artifact itself and updated with each version. That doesn’t become outdated in the same way as a separate document.
People leave—that’s inevitable. The question is whether that knowledge leaves with them, or whether the organization itself is structured to retain that knowledge, regardless of who comes and goes.
Curious about how vulnerable your organization is in this regard? We help organizations not only build a solid data platform but also properly establish ownership and transferability, so that knowledge about your data and tools—including AI—doesn’t disappear the moment someone leaves. Schedule a 30-minute, no-obligation intake session, and we’ll work with you to find the best solution.