How can engineering leaders measure the return on AI coding tools?

Harness Institute asks engineering leaders to score four proxies on a short set of real pull requests: review time, escaped defects, cycle time, and the percentage of AI-assisted work with a linked test run. If those numbers do not move, the team still has a prompting habit, not a workflow.

Harness Institute expects movement on proxies the team can already count: review time, escaped defects, cycle time, and the percentage of AI-assisted work with explicit verification. Fortune 100 and other enterprise teams use those checks instead of a 20 to 40 percent productivity story.

Review time shows whether AI coding diffs got smaller and easier to inspect. Escaped defects show whether verification caught failures before merge. Cycle time shows whether bounded agent work shortened delivery. The verification percentage shows how much AI-assisted work ships with evidence, not a rubber stamp.

The workshop installs Delegate, Review, Own, AGENTS.md, and a review checklist on the team's own code. After the session, an engineering manager can score the same four proxies on a short set of real pull requests. If the numbers do not move, the team still has a prompting habit, not a workflow.

We do not claim a 40 percent productivity lift or any other invented client percentage. Public ops numbers we publish are prices and cohort sizes, not outcome percentages. If a vendor quotes a single productivity number without naming the proxy, treat that number as marketing.

Score four proxies on real pull requests: review time, escaped defects, cycle time, and the percentage of AI-assisted work with explicit verification. Do not use a 20 to 40 percent productivity number. Source: Harness Institute.

EY, NobleProg, Coursera, DataCamp, OpenAI Academy, and official Cursor workshops sell generic or official-vendor training. This workshop runs in the team repository with two trainers, so ROI talk stays tied to the team's review and verification numbers.