The pilot is not the product

88% use AI. 12% see a return. The seven-fold gap is not a technology story, and the sector breakdown says exactly what it is.

Eighty-eight percent of enterprises use AI, twenty-three percent scale agents, twelve percent report a return. Eighty percent of enterprise applications embed an agent; thirty-one percent run one in production; eighty-eight percent of pilots never ship.

Read the sector split

Banking and insurance lead production deployment at 47%. Healthcare and government trail at 18% and 14%.

That ordering is the answer to the whole puzzle, and it is not the ordering you would predict from engineering capability. It is the ordering you get from asking a different question: who already has an apparatus for being accountable for an automated decision?

Banks did not win at agents. They won at audit trails, twenty years ago, and are now collecting.

Payback tracks blame, not difficulty

Median time-to-value is 5.1 months. SDR agents pay back in 3.4; finance and operations agents in 8.9. Nobody thinks prospecting is technically harder than reconciliation.

The difference is what a wrong answer costs. A bad outbound email costs an email. A bad journal entry costs a restatement. The fast deployments are the ones where the failure mode is embarrassment, and they are fast because nobody had to build a control apparatus first.

The forecast follows mechanically

More than 40% of agentic projects are forecast to be cancelled by end of 2027, on escalating costs, unclear business value, and inadequate risk controls. Only the third is a thing the field could build and mostly has not.

Set against 88% of pilots never shipping, a 40% cancellation rate among survivors is not a crash. It is a market finding out which applications were ever real.

What this asks of you

Stop treating the control layer as the thing you add before go-live. Provenance, a bounded authority envelope, and a stop button are not compliance overhead — they are the difference between a 3.4-month payback and a cancelled programme. The sectors that already had them are the ones in production.

The pilot proves the model can do the task. Nothing about the pilot proves you can be responsible for it, and that second proof is the product.

Digital Applied — AI Agent Adoption 2026: 120+ Enterprise Data Points → · Joget — AI Agent Adoption 2026: What the Analysts' Data Shows →