Lire en français

Clean the source. The formula fits in three words, and we often set it as a condition for deploying agentic AI. Here is what it cost when we had to live up to it.

Obsolete versions never really disappeared. New folder structures appeared on the network without the previous ones being deleted. Where naming conventions existed, they varied from one team to another. Some folders still carried the names of people who had left long ago. Equivalent documents were archived in different locations depending on the country or the team.

Yet the system worked, because someone always knew where to look.

There was the person who remembered that this type of document had never been stored in the official repository. The one who could trace the origin of a discrepancy between two sources meant to be authoritative. The one who knew that a file bore a misleading name because it had been moved without ever being renamed.

The consistency was not in the corpus. It was in people’s heads.

When information was missing, it meant going back through the digital archives, then the paper archives. And when the data was too important to ignore, on occasion it had to be requested from the competent authority, sometimes the only place where it still existed.

Document management was nobody’s deliverable. It appeared in no objective, blocked no deadline, and its absence was never visible. As long as the same people stayed, the system kept working.

It was a debt without a creditor. AI was the first creditor to show up.

The first year was spent defining a target structure and common conventions. In the years that followed, this tedious discipline had to be applied, at the expense of more rewarding objectives. The goal was renewed, again and again, like Sisyphus pushing his boulder to the top of the mountain.

To free up the resources, the optimisation of certain processes was postponed. In the meantime, they ran unchanged.

By the end of the exercise, discrepancies remained. On certain document types, teams doing the same work had continued to diverge. They shared the same objective, but nothing in their processes required them to coordinate, and a shared objective can be met perfectly well by each team on its own.

Cleaning the source corrects the state of the corpus. It does not change the processes that keep producing it.

Years of effort to make a corpus usable. How long will it stay that way?