The number on the board slide is the wrong number
Most enterprise AI adoption reporting counts seats issued and weekly active users. Both go up. Neither tells you whether a single process now runs faster, cheaper, or with fewer handoffs than it did last year. So you get the plateau. Everybody has a copilot, the tool sprawl chart looks impressive, and the operations team will tell you privately that nothing about how work moves through the company has changed. Assistants got adopted. Delivery didn't.
What the plateau actually looks like from the inside
It has a recognisable shape. Usage is broad and shallow: thousands of people using AI for drafting and summarising, which are the two tasks that need no integration and change no system of record. Nothing is wired into the workflow, so every output needs a human to carry it to the next step by hand. Nobody owns an outcome, because the rollout was owned by a platform team whose success metric was activation. And the finance question lands hard about nine months in. Someone asks what the spend bought, and the honest answer is a lot of faster first drafts. That's real value, and it's nowhere near the business case that funded the programme. The uncomfortable part: the teams furthest along on seat count are often furthest behind on production. They optimised for the metric that moved with the least effort.
Count these instead
Cycle time on a named process, before and after, measured the same way both times. Not perceived time saved from a survey. The clock. Steps removed from a workflow. If the AI produces a draft and a human still does every downstream step, you added a step rather than removing one. Cost per completed outcome, with the model spend, the review time, and the rework all in the numerator. A workflow where AI handles 80 percent of volume but the 20 percent it gets wrong costs two hours each to unpick is not a saving. Exceptions and rework rate, tracked over months. This is the number that decides whether a pilot survives its first quarter in production, and almost nobody instruments it before launch. Run those four against one workflow and you'll learn more in six weeks than another year of activation dashboards will tell you.
The clause in the AI Act that speaks to adoption, not risk
Compliance conversations about the EU AI Act default to risk classification, which is why Article 4 gets missed. It has applied since 2 February 2025 and it obliges providers and deployers to ensure a sufficient level of AI literacy among the staff and others operating AI systems on their behalf, taking account of their technical knowledge and the context of use. Read as an adoption instrument rather than a compliance chore, Article 4 is the only part of the Act that tells you to invest in the people using the system. Which happens to be the constraint. A team that doesn't know where a model's output is unreliable will either trust it everywhere or trust it nowhere, and both produce a plateau. ISO/IEC 42001 gets at the same thing through Annex A control A.9 on the use of AI systems: define how these systems are meant to be used, by whom, with what oversight. Teams treat that as documentation. It works better as an enablement plan.
Finish two workflows before you licence anyone else
Here's the recommendation that gets argued with in steering committees, and we'd make it anyway. Freeze net new licences. Pick two workflows that a named executive already wants to be faster. Finish them end to end, into production, with the controls in place and the numbers instrumented from day one. End to end means the AI writes into the system of record, not into a chat window somebody copies from. It means the exception path is designed, not improvised. It means someone owns the outcome, and that person sits in the business, not in the platform team. Two finished workflows do something a broad rollout can't. They produce a real cost per outcome, a real rework rate, and a delivery pattern the next ten workflows can reuse. They also settle the internal argument about whether this works here, which is usually what's actually blocking the budget. The reason enterprise AI adoption plateaus isn't that the models aren't good enough. It's that adoption was scoped as a distribution problem and it's a delivery problem. Operators who have taken AI to production at 300,000 organization scale see the same sequence every time: value from the first governed workflow, then permission to do ten more. Difinity partners with regulated teams to find the use cases worth doing, ship them with the team, and put the control and audit in place so they hold up once they're carrying real volume.