Practical AI delivery

Five reasons an internal AI pilot never reaches production

A prototype proves that a model can produce an interesting response. Production requires a valuable workflow, representative evidence, controls, integration and somebody prepared to operate it.

By Chris Domville · 2 September 2026

1. Nobody defined the business baseline

“The answer looks good” is not a commercial measure. A pilot needs to compare time, cost, quality, throughput or risk with the current process. Without that baseline, decision-makers cannot justify further investment or decide what level of performance is sufficient.

2. The test data was too convenient

Demonstrations tend to use clean, familiar examples. Production receives missing pages, unusual formats, conflicting instructions and requests that were never anticipated. Representative evaluation cases should include normal work, difficult edge cases and inputs the system must reject safely.

3. Human control was added as an afterthought

A vague promise that “a person will check it” does not define who reviews what, which evidence they see, how exceptions are prioritised or who is accountable for the final action. Human review must be designed as part of the workflow, not used to excuse unreliable automation.

4. Integration and security arrived too late

A standalone chat interface avoids the hard work of identity, permissions, approved data, existing applications and audit trails. Those constraints often determine whether the idea can create value. They should shape the architecture before the prototype becomes difficult to change.

5. There was no operating model

Models, prompts, data and costs change. Production needs monitoring, ownership, incident handling, evaluation and a controlled process for improvement. If nobody owns those responsibilities, the pilot cannot become a dependable service.

A minimum production-readiness checklist

ValueA baseline, measurable outcome and named business owner.
QualityRepresentative tests, thresholds and safe behaviour when confidence is low.
ControlPermissions, privacy, human review, escalation and traceable evidence.
OperationMonitoring, cost limits, support ownership and a managed change process.

The next step should still be bounded

Do not respond to a stuck pilot by funding a large production build. Assess the gaps, decide which risks can be remediated and prove the most important assumptions in a controlled stage.

Domville Tech can provide an AI production-readiness review. If the organisation is still selecting an opportunity, start with AI consulting for business instead.

Have a promising pilot?

Find out what production really requires.

Discuss the pilot