Five patterns I keep seeing in large AI rollouts

A note on what this is

These are observations rather than case studies: patterns I've seen come up again and again in how large organisations roll out AI, often enough to be worth writing down. Nothing here identifies a specific organisation.

The pilot always looks better than the rollout

Pilots consistently produce better results than full rollouts. A pilot runs on examples someone picked because they'd show the tool off well. A rollout runs on whatever material the organisation actually has, unfiltered. The drop isn't the tool getting worse. It's the first honest measure of how much hand-picking the pilot was doing. Why your AI rollout doesn't match the pilot unpacks this pattern on its own.

The most senior person in the room usually has the least accurate model of what junior staff do all day

Leadership tends to underestimate how much implicit judgement sits inside roles that look, from a distance, like they follow a fixed process. That miscalculation shows up directly in rollout plans: tasks get slated for full automation because they look procedural from an org chart's height, and turn out to require exactly the kind of situational judgement that was invisible from that vantage point.

The team that documents the least resists AI the hardest, and it's rarely stubbornness

Teams with the least documentation push back hardest on AI, and it's rarely stubbornness, even if it looks like resistance to progress from the outside. Usually their processes exist only in people's heads, and an AI tool exposes that straight away: every attempt to automate a task shows how much of the work was never written down. The resistance is a symptom of context debt, not a problem with attitude.

Nobody budgets for maintaining the context, only for building it

Rollout plans reliably include a line for the initial setup: the material a model needs, curated once, at launch. Almost none include a line for keeping that material current as policies, products, and exceptions change. It's the same blind spot that used to catch software teams off guard before "maintenance" became a normal line item of its own, showing up again for a different kind of asset.

The rollout that succeeds quietly is never the one anyone tours

Nobody shows off a rollout that just works. The high-profile initiatives with executive sponsors get the attention, the case studies and the recognition. The unglamorous deployments that quietly fix a back-office bottleneck rarely get noticed, precisely because there's no drama: they work without anyone having to step in.

What to take from this

None of these are universal laws, and there are exceptions to every one of them. What's notable is how consistently they recur, across companies with nothing else in common (different sectors, different sizes, different starting points on AI). That consistency is itself information: it suggests these are structural patterns in how organisations meet AI, not quirks of any one company's culture.

The same underlying variable shows up one level down, inside a single company rather than across many: see why two teams get different results from the same AI.