Decisions about where to use AI in organisations often happen some distance from the work itself, and the people who know where the work gets complex are not always in the room.
I have spent a lot of time mapping how work happens in organisations, and two things are fairly consistent. The person doing the job can usually tell you where the time goes and where the workarounds sit, and the problems that do not happen often are not documented.
When AI arrives under the heading of efficiency, many people hear it as a question about headcount, and few will describe in detail which parts of their own job could be done without them.
None of this is new. I saw the same pattern during business services transitions, where people were asked to document their work so it could be handed over to another team. Some did that thoroughly, but others left out the parts that only they knew. It is just what happens when people are asked to make their own work easier to transfer.
Organisations then end up with the version of the work people are willing to describe, which is usually simpler than the real thing. If AI is designed around that simplified version, it will miss some of the judgement and exceptions that make the work function in practice.
I do not think better workshops or process-mapping questions will solve that on their own. What people are willing to say about their work depends partly on what they believe will happen afterwards. If they think the real purpose is to reduce headcount, or make their expertise less valuable, they are unlikely to give the full picture. That is a trust issue as much as a process issue.
There is an argument that AI will work some of this out over time. Given enough information about how a process runs, it may identify patterns and hold-ups that people have stopped noticing.
But it also carries a risk that AI follows the process as written, including the parts people stopped following because they no longer worked. Or it may find its own way around the gaps, without the human judgement that previously sat inside the workaround.
Most organisations have far more knowledge about their own work than they can currently document.
That may turn out to be the real test of AI implementation: whether the organisation can create enough trust to understand the work properly before it redesigns it.