The pilot-to-production gap.

The short answer

Roughly four out of five enterprises report having adopted AI agents, and roughly one in ten runs them in production. That gap is not a maturity curve that time will close on its own. The 2026 studies converge on three specific causes — inference cost underestimated at production volume, agents attached to processes nobody owns, and failures that stay silent until trust is gone — and each one is decided before a line of code is written.

How wide is the gap, exactly?

The numbers differ by survey population, but never by direction. A March 2026 enterprise survey found 78% with agent pilots underway and fewer than 15% reaching production. Analyst tracking through 2026 puts adoption near 79% against roughly 11% in production. MIT's State of AI in Business work found that 95% of generative AI pilots delivered no measurable impact on the P&L.

Gartner expects more than 40% of agentic AI projects to be cancelled before the end of 2027, naming escalating costs, unclear business value, and inadequate risk controls. Read those together and the picture is consistent: the constraint is not model capability. It is everything around the model.

Why does inference cost kill projects specifically at cutover?

Because pilots are measured at a volume where cost is invisible. A few hundred calls a day makes per-call cost a rounding error, so nobody instruments it. Production multiplies the volume, and the things that were tolerable in testing — retries on malformed output, context windows carrying far more than the task needs, a second model call to check the first — become the bill.

The overrun that kills the cutover is commonly two to three times the projected run cost, and it is rarely a surprise about price per token. It is infrastructure debt: caching, batching, routing cheap work to cheap models, and truncation policies that a pilot assumed would be added later. Later is cutover, and by then the budget was approved on the pilot's numbers.

This is measurable before you commit. Our own LLM cost calculator exists for that: run the projected volume, not the pilot volume.

What does “nobody owns the process” actually look like?

An agent is attached to a workflow that crosses three teams. It works. Then a question arrives that it handles badly, and there is no single person whose job it is to decide whether the answer was acceptable. Escalation has no destination. The agent is not switched off — it is simply never trusted with the volume it was built for, and it stays in the pilot column forever.

The deployments that ship share a shape: one high-volume, well-defined task, with one owner who already answered for that outcome before the agent existed. Document processing, support triage, invoice handling, clinical notes. Not “customer experience” — a queue with a name and a manager.

Why is silent failure worse than loud failure?

A crashed service pages someone. An agent that quietly answers 4% of cases wrong does not, and by the time the pattern surfaces it has been wrong for weeks in front of customers. Most pilots ship with application monitoring and no output monitoring: uptime is watched, answer quality is not.

The fix is unglamorous and belongs in the build, not after it. A held-out set that runs on every prompt change. Sampling of live outputs with a human verdict. An audit log that records what the agent saw and why it acted — which is also what an auditor will ask for under the EU AI Act. We published our eval harness and audit-log spec rather than describe them.

Does the market slowdown change the picture?

No — spending is still climbing. The agentic AI market moved from roughly $7.6B in 2025 to around $10.8B in 2026, and the enterprise coding-agent segment alone is tracked at $9.8–11.0B annualized. Gartner expects that by 2027, more than 65% of engineering teams using agentic coding will treat the IDE as optional, with control and validation moving to platforms.

What is changing is how it is bought. Vendors are moving from seat-based subscriptions to usage-based pricing, which pushes the inference-cost question from a footnote to the centre of the business case — exactly where the pilots that failed had left it.

What this means if you are scoping an engagement

• Price the production volume, not the demo. If the projected run cost is not in the proposal, the proposal is incomplete.
• Name the owner before the use case. If no single person answers for the outcome today, automating it will not create one.
• Budget the eval harness as build, not polish. Output monitoring is what separates a system you can defend from one you hope about.
• Take the boring queue first. High volume, narrow definition, existing owner. The ambitious one goes second, once the pattern is proven.

Sources

Figures in this article come from published 2026 analyst and academic work: Gartner's enterprise AI coding agent market analysis and its agentic-project cancellation forecast, MIT's State of AI in Business pilot-impact finding, and enterprise adoption surveys published through Q1–Q2 2026. Where surveys disagree on the exact production rate, we have given the range rather than the flattering end of it.