Escaping the pilot graveyard

A pilot is a bet that a workflow can be improved enough to justify changing how people work. Most pilots we are asked to rescue failed for reasons that had nothing to do with model quality.

Pick a workflow with a number attached

'Improve knowledge sharing' cannot succeed. 'Reduce average intake handling time from 14 minutes to under 5, without increasing rework' can. The number should already exist in a system you can query, and someone should be accountable for moving it.

Avoid the highest-stakes workflow first

The most valuable workflow is often the worst first project: it carries the most scrutiny, the longest approval chain and the least tolerance for iteration. Start one layer beneath it — the preparation work, the first draft, the exception triage — and earn your way up.

Name the owner before you write code

Every successful deployment we have seen had a named operational owner who used the system daily. Pilots owned by a committee become demos. Find the person whose Tuesday gets better, and build for them.

The goal of a pilot is not to prove AI is impressive. It is to prove one workflow is now cheaper, faster or safer — and to leave behind the instrumentation that keeps proving it.