1. The use case is interesting, but not important

Pilots often begin with what is technically impressive rather than what removes a costly bottleneck. Start with a decision, delay or repeated task that already has an owner and a measurable outcome.

2. Nobody owns the operating result

A technical team can build the prototype, but a business owner must be accountable for quality, exceptions, adoption and value after launch.

3. Human review is undefined

“Human in the loop” is not a control until the organisation defines who reviews what, at which threshold, with which evidence and what happens when confidence is low.

4. Data and vendor risks arrive too late

Security, privacy, retention, model terms and supplier dependencies should be assessed while the workflow is designed—not after users have become dependent on it.

5. Success is measured by usage alone

Useful metrics include cycle time, error rate, rework, escalation volume, user confidence, cost and the percentage of outputs requiring correction.

Scale only what survives the test: demonstrated value, clear ownership, controlled data, reliable fallback and measurable quality.
← All insights