Insights
Notes from the work
Written by the people doing the engagements. Specific enough to disagree with.
- AI systems
- Evaluation
Jul 22, 2026
AI systems
2 min read
Most failed AI pilots are data failures with better branding
Six pilots that stalled, and what was actually wrong in each. In five of them the model was fine.
Jun 18, 2026
Evaluation
2 min read
Write the evaluation set before you write the prompt
Forty graded cases, drawn from real inputs, cost about a day. Skipping them costs about six weeks.