Agentic AI
80% have embedded agents. 31% have deployed them. The gap is the whole story.
Survey figures showing most enterprises experimenting and a third in production get read as slow adoption. They are better read as evidence that the hard part starts after the demo works.
Enterprise survey data for 2026 puts roughly 80% of organisations embedding agents somewhere and around 31% actually deploying them. The usual commentary treats the gap as adoption friction — procurement, skills, change management.
Having watched a number of these programmes, we would locate it somewhere more specific.
The demo is not the hard part, and it never was
Getting an agent to do a task impressively once is now genuinely easy, which is why 80% have done it. The frameworks are good, the models are capable, and a motivated engineer can produce something that works on a Tuesday afternoon.
What stops the other half is everything between working once and being allowed to run unattended against production. That is not adoption friction. It is the absence of an answer to a question the demo never had to face: what happens when this is wrong, and who finds out?
A pilot proves the capability exists. Deployment requires proving the failure is survivable, and nobody demos that.
What the 31% built that the others did not
Consistently, four things. A bounded authority model, so the worst case is enumerable rather than imaginative. Observability into decisions rather than infrastructure, because the failures return HTTP 200. A rollback path that does not depend on identifying the problem first. And an owner — a named person accountable for the agent's behaviour, which sounds bureaucratic and is the single strongest predictor of whether a system survives its first incident.
None of that is AI work. It is the operational engineering that any consequential automation requires, and it is unglamorous enough that it rarely gets budgeted in a programme sold on capability.
The implication for planning
If you are in the 80%, the useful question is not which model or framework. It is what you would have to build to let this run for a quarter without supervision — and then whether the value justifies building it.
For a good number of pilots the honest answer is that it does not, and stopping there is a better outcome than joining the population Gartner expects to decommission in 2027.
Sources
Written by the Devence Lab research team.
