The memory the company never wrote down
Why AI stalls in organizations before the model and before the tool: the company cannot say what it knows.
If we think of an organization as a living organism, what happens when it needs to reflect on its own memory?
Most companies make a very fragile attempt at formalizing the knowledge that circulates inside them, when they do anything at all. That is not necessarily carelessness, because doing this work is expensive, requires technology few can sustain, and only a small slice of companies saw a return in investing in that kind of maturity. Many organizations know it makes sense, but almost none do it.
What has changed is that the research literature of recent months keeps pointing to the same place when it asks why AI does not take hold inside organizations, and the obstacle it runs into appears before the model and before the tool. The blockage lies in the absence of formalized knowledge, in the fact that the company cannot say what it knows.
This is a subject software engineering has discussed for decades, under the name of requirements engineering, and the order there was always to obtain the knowledge of the people who operate, understand what those people need to do day to day, and only then build the system that will support that work.
At some point companies stopped caring about that part, and what remained were two similar patterns: those who modeled the knowledge once, on the day they bought the system, and never revisited it, and those who forced a generic system, designed from other companies, to fit an organism that was not that one, contorting themselves to accommodate a process imported from outside.
Today's promise is that adapting has become cheap and fast, and it is indeed possible to customize each person's relationship with the company's systems in a way that was unthinkable until recently. But without someone organizing this, the result turns into cacophony, because at the end of the day the work of an organization is to coordinate action, and coordinating requires agreeing on things explicitly, with each person knowing what they do and where the two meet. Customizing everything for everyone without making that agreement visible produces a lot of people using chat and improving their own performance a little, without any of it becoming a capability of the company.
When you externalize the knowledge that lived in someone's body, it looks as if that person has become dispensable, and that would hold if memory were something static. But intelligence is born from a memory being continuously reassessed, and the criteria for that reassessment are human, made of judgment and of projecting the future.
AIs project the past very well and invent futures that are linguistically coherent without those futures being truly coherent, which should leave to someone the task of looking at the company's body of knowledge and deciding what still holds, what has aged, and what was never quite like that.
Formalizing an organization's memory, then, places people inside a permanent cycle of reflection on what the company knows, and whoever carried that knowledge in their body begins to discuss, review and update it together with others, instead of keeping it alone until the day they leave the company and take everything with them.
If you are trying to make AI work inside your organization, I would very much like to know where it gets stuck for you: whether your company can say what it knows, and who, in practice, revisits that knowledge.
This article was inspired by the second episode of Life After AI, with Diogo Dutra, available on YouTube and Spotify. The episode is in Portuguese.