Inside the working chart, two patterns do most of the actual work. One is the help graph: the unwritten map of who asks whom when they need an answer quickly. The other is the trust graph: the ambient answer, inside the company, to the question ‘whose word do I take on this kind of thing.’ Both are load-bearing. Neither appears on the formal chart. AI is changing both, in different directions.
Take the help graph first. A year ago, if an analyst needed a quick answer on a regex or a SQL join, they’d ask the person on the team who was known to be good at that kind of thing. That’s what the help graph is: a network of small dependencies. Most of it is unwritten. All of it was built one useful moment at a time.
Now the analyst asks a model and gets an answer that’s often good enough. The colleague who used to field that question never hears it.
Multiply that across a team of thirty over a quarter, and the help graph loses a visible fraction of its traffic. Multiply across a company of three hundred, and the graph looks thinner by the end of a year.
The answer still gets delivered, by the model now instead of the colleague. What goes away is everything around the answer: the five-minute chat where the colleague mentioned something tangentially useful, and the small signal that you and the colleague are still in the habit of talking at all.
The trust graph is the other half. Before AI, when a VP wanted to know whether a number could be trusted, they’d ask a specific person, because that person was known to have good judgment on that kind of number. The trust graph is harder to see than the help graph because it rarely gets articulated, but every company has one, whether or not anyone has drawn it.
AI pulls trust toward itself for certain kinds of questions. The thing that moves the trust is speed and the absence of a side request, which people route around even when they know the model is less reliable than the person they used to ask.
Over time, the trust graph reshapes. Certain questions that used to route to certain people start routing to models first, and only come back to the people when the model answer doesn’t quite fit. The person who used to be the primary trust node for that kind of question now operates as an escalation path. That’s a different role, and most companies haven’t noticed that their trust graph just moved it.
The person who was the trust node for a certain kind of question was usually doing more than answering questions. They were carrying context that didn’t travel well, and teaching the less experienced people by example. Some of that survives in their new escalation role. Most of it doesn’t, because the teaching happened through the volume of small routine questions that now go to models.
Put the two changes together. The help graph thins, and the trust graph redirects. Nobody drew a new chart and nobody moved anyone’s box. Everyone who reads the company by looking at boxes and arrows misses both movements entirely.
The two graphs are the cheapest place to start looking. Before any of the usual responses, try to draw what the help graph and the trust graph looked like two years ago and what they look like now. The gap between those two pictures is the thing most companies aren’t reading.
From there, two questions matter. Which of the help loops that disappeared were doing useful work, and do any of them need to be rebuilt on purpose? Which of the trust nodes that moved were doing work beyond answering questions, and who is doing that other work now?
These are the two specific moving parts that leave the clearest trace. The reorganisation as a whole is hard to see. The thinning of the help graph and the redirection of the trust graph are not. Watching what has happened to each is the shortest route to what has already moved.