Most stalled pilots did not fail on the model. They stalled because nobody owned the result, the data or access was never sorted out, the review step was unclear, or there was no agreed way to judge success. These four steps find the real blocker, fix it in a bounded way, and end with a clear decision.

Step 1 — Diagnose the stall

Goal: Find out what actually stopped progress.

  1. Write down what the pilot was meant to change and for whom.
  2. Ask the people involved where it stopped: ownership, data, access, review, cost or trust.
  3. Check whether a baseline and a success measure were ever agreed.
  4. Separate technical gaps from decision and ownership gaps.

You finish with: A short diagnosis naming the main blockers.

Step 2 — Re-scope to one workflow

Goal: Shrink the pilot to something that can go live.

  1. Pick the single workflow where the pilot showed the most value.
  2. Name a business owner who will run it day to day.
  3. Define approval points, stop rules and what the system may not do.
  4. Agree the measure that will decide whether it stays.

You finish with: A one-page scope with an owner, controls and a measure.

Step 3 — Fix the blockers and run in parallel

Goal: Remove what blocked go-live and test on real work.

  1. Sort out data access and permissions with the narrowest scope that works.
  2. Integrate with the tools the team already uses instead of a separate demo app.
  3. Run alongside the current process on real work for a few weeks.
  4. Log every exception and correction.

You finish with: A working version in parallel run with an exception log.

Step 4 — Decide: launch, adjust or stop

Goal: End with a decision rather than another extension.

  1. Compare results against the agreed measure and baseline.
  2. Launch with a runbook and owner if it meets the bar.
  3. Adjust once with a fixed date if it is close.
  4. Stop and record the lessons if it is not worth running.

You finish with: A recorded decision with an owner, or a documented stop.

Common pitfalls

  • Treating the stall as a model problem when it is an ownership problem.
  • Extending the pilot indefinitely instead of deciding.
  • Keeping the pilot in a separate tool nobody opens.
  • No baseline, so success cannot be shown.

See it as a workflow

Leadership

From stalled AI pilot to production

Take a pilot that impressed in a demo but never shipped, decide whether it deserves production, and get it there with owners and controls.

See the workflow

More guides like this

Related answers

Read as Markdown Fractional Chief AI Officer