Most Agent Pilots Never Reach Production. The Reasons Are Not Technical.

The most quoted number in enterprise AI this year is that roughly 88% of agent pilots never reach production. It originated in Anaconda and Forrester research, has been replicated in independent surveys by a16z and the MIT Sloan CIO panel, and Gartner's version of the figure lands at 89%. When four sources with different methods converge, the number is usually describing something real.

Gartner also expects more than 40% of agentic projects to be cancelled by 2027, and 22% of deployments that did reach production report negative return at twelve months.

The instinct on reading this is to conclude the technology is oversold. The root cause data says something more specific and more useful.

Where the failures actually come from

Forrester attributes 41% of failures to unclear success criteria, 33% to insufficient tool or data access, and 26% to drift in evaluation coverage.

Read that list again with an eye on what is absent. Not model quality. Not capability. Not cost. The three named causes are, respectively, a definition problem, an access problem, and a measurement problem. All three are decided before anyone builds anything, and all three are the sort of thing an organisation defers because they are unglamorous and nobody owns them.

Unclear success criteria means the pilot began without anyone writing down what a correct output is. Insufficient tool or data access means the agent was asked to do a job while cut off from half the information a human doing that job would use. Drift in evaluation coverage means the tests written at the start stopped resembling what the system encounters now, and nobody noticed.

Why this is a consulting market rather than a technology story

“Not model quality. Not capability. Not cost. The three named causes are a definition problem, an access problem, and a measurement problem, and all three are decided before anyone builds anything.”

Each of those failure modes is a piece of work. Not vague advisory work either. Each has a deliverable, a clear end state, and a buyer who has already been burned once.

Defining success criteria for a specific workflow produces a written standard for what correct means, including the edge cases and who adjudicates them. That is a two to four week engagement and it is the highest-leverage document in the whole programme, because 41% of failures trace back to its absence.

Mapping data and tool access produces an inventory of what the agent needs to see, what it currently sees, and what it is blocked from and why. Half of those blocks turn out to be policy rather than technical, which means they are negotiable.

Not model quality. Not capability.

Building an evaluation set that keeps pace produces a regression suite grounded in real cases plus a cadence for refreshing it. This is the one clients understand least and need most, because drift is invisible until it is expensive.

Who is positioned to sell this

Not the model vendors, whose incentive is to demonstrate capability rather than to constrain it. Not the large integrators in most cases, whose economics favour longer programmes than these.

The person best placed is someone who knows the workflow well enough to say what correct looks like without a discovery phase. That is a domain expert, and the pilot failure data is the reason their knowledge is now billable in a form it previously was not.

The old version of this expertise was sold as opinion, in a room, by the hour. The version that the current failure data creates demand for is a specification: a written artefact that a technical team builds against. That is a productisable deliverable with a fixed price and a repeatable shape.

The uncomfortable part

This work is most valuable before a pilot starts and most purchased after one has failed. Selling it upfront means arguing against optimism, and the client who has not yet failed does not believe the 88% applies to them.

The practical answer is to lead with the failure data rather than with your service. The number does the persuading. Your offer is what somebody does about it, and it lands far better as a small fixed-scope engagement attached to a project already in flight than as a strategy conversation ahead of one.

See the operator's case for building rather than optimising for the adjacent product opportunity, and productising your consulting for turning a repeatable specification into a fixed offer. The consultant track covers how this fits an existing practice.