Insights · AI delivery · 10 November 2026

Why AI pilots stall, and how to reach production in 30 to 90 days

Most AI pilots never reach production. The reason is rarely the model. It is the order in which decisions are made.

McKinsey's 2026 State of AI research found that 63% of organisations report no EBIT impact from their AI use. On first calls we hear the same sentence again and again: "Our AI pilots never reach production."

The pattern is familiar. A pilot is built quickly to show what is possible. It impresses in a demo. Then production requirements surface: security review, integration, monitoring, data access, a named owner, a way to roll back. The pilot was not built for any of them, so it is rebuilt, and the business loses patience.

Ask four questions before you build anything

We only build where all four answers are yes. If a use case fails one, we say so.

  • Adoption: will your people use it? Test it against the workflow it lands in, with the people who will run it every day.
  • Feasibility: can it be built well, inside your infrastructure? Check data quality, access and controls before a line of code is written.
  • Need: does the business need it now? Rank it against every other use case on manual density, data availability, compliance risk, labour exposure and speed to value.
  • Value: will it show on the P&L? Agree a baseline, a target and a named P&L line before the build starts.

Set production requirements in week one

The single biggest change is to write production requirements as acceptance criteria at the start. The proof of value then becomes a rehearsal on a small scope, not a throwaway demo. Nothing has to be rebuilt, because it was built for production from the first sprint.

Week one should agree the workflow and its volumes, the accuracy and cost thresholds, the data in scope and a named owner who will sign off each stage.

Use decision gates, not momentum

Pilots drift when nobody is asked to decide. Three signed gates keep the work honest:

  1. Brief signed. Baseline, thresholds, data in scope and named owner agreed.
  2. Accept, iterate or stop. Measured at real volume against the week-one thresholds.
  3. Cleared to go live. Integration, monitoring, review workflow and rollback signed off.

Stopping at gate two is a good outcome when the evidence says so. It costs weeks, not a year.

When accuracy is poor, look at the architecture

Many production problems look like tuning faults and turn out to be architectural. At one client, a compliance review pipeline told reviewers that answers were missing when suppliers had provided them. The pipeline read each page in isolation, so a question at the foot of one page arrived separately from its answer at the top of the next.

The fix was not a better prompt. The submission is now read as a whole structured document, and the pipeline tracks which section and question it is inside. Measured against previous production behaviour, AI calls to read a 48-page document set fell from 46 to 9, and cost per submission fell by 18%.

Plan for the day you leave

A use case is only finished when your own team can run it. Code, configuration, the evaluation set and the runbook should transfer at handover, with training for the people who operate it. That is what turns one successful use case into a capability the business can repeat.

Next step

Scope your first use case

One value stream, one pod, success metric agreed up front.