Why AI pilots fail to reach production.

01What a stalled pilot looks like

An AI pilot proves its value and never reaches production. The prototype works, the business case is clear, and the project stalls because governance, security, cost control and ownership were never built in. It is a delivery problem, and rarely a model problem.

The pattern has a shape. A team builds an agent that works. Leadership approves a pilot. Six months later the pilot still runs, nobody can say what it costs or what it returned, and the budget review arrives.

02How common it is

In June 2026 research for Elastic among more than 500 senior decision-makers at Australian organisations, 32 percent had paused, cancelled or wound back AI deployments because the output could not justify the cost. Only 8 percent were tracking whether AI delivered revenue or savings.

McKinsey's State of AI survey, published in November 2025 from 1,993 respondents in 105 countries, found that 88 percent of organisations use AI in at least one function, about one third have begun scaling it across the enterprise, and 7 percent describe it as fully scaled. Only 39 percent attribute any enterprise-level profit impact to AI at all.

03Why pilots stall: four reasons, in the order they bite

1. The cost was never recorded. The pilot ran on a shared credential. When finance asks what it cost per run, the answer is an estimate. In 2026, KPMG found that 26 percent of large US enterprises could see AI running cost in real time.
2. The approval was never scoped. Someone said yes to a demo. Nobody wrote down what the agent could touch, what data it could reach, or what would need a second look. When production asks for that scope, it has to be invented after the fact.
3. Nobody owns the scale decision. The builder owns the pilot. The business owner owns the outcome. The decision to scale or stop sits between them, and decisions with no owner wait.
4. The path is different. Production means identity, approval, cost and audit on every action. The pilot skipped all four to move quickly. Each skip is a migration already scheduled, whether or not anyone knows it yet.

Deloitte's State of AI in the Enterprise 2026, from 3,235 leaders in 24 countries, found that 74 percent expect to be using AI agents by 2027 and 21 percent have a mature governance model for them. The gap between those two numbers is where the next two years of stalled pilots will come from.

04The pilot-to-production checklist

Run this before the pilot is approved, and it will not need a rebuild later.

1. The measure of success is written down, with the number that means scale and the number that means stop.
2. A named person owns the scale-or-stop decision, and a date is set for it.
3. The pilot runs on the same action path production will use, with the scope turned down.
4. Every run records its cost, attributed to the team and project.
5. The approval states what the agent can touch, what data it can reach, and what needs a human.
6. When the agent's model or tools change, the approval comes up for renewal.
7. The audit trail can answer what ran, what it cost and who approved it, in one query.
8. The budget is a cap that holds before spend is committed, with a named person who can raise it.

05What the ones that ship do differently

They treat the pilot as the production system with the scope turned down, and nothing else. Same path, same record, fewer users and a smaller blast radius. The governance is built once, in the path, and every new agent inherits it. The tenth pilot costs a day to stand up because the first one paid for the path.

That also changes the scale conversation. When the pilot runs on the production path, its numbers are real. You know what it cost per run and what it touched, so the decision to scale is a query instead of an argument between a champion and a sceptic.