The dependable flow
The inputs are: issue, repository, rules, tests and acceptance criteria. Before AI processing, completeness, data class and permitted sources are checked. The planned output comprises plan, isolated change, tests and reviewable diff.
The output remains a proposal until the defined approval stage is passed. Missing data, contradictions or low confidence visibly route the process to manual handling.
What is not automated
The workflow does not delegate merge, deployment, data migration and production to a model for convenience. Werkverstand defines who decides, which evidence must remain visible and how the decision is recorded.
- Keep sources and assumptions visible
- Require domain approval before an output affects an external party
- Measure errors and exceptions as operational evidence
Pilot and measurement
A two-to-four-week pilot uses real, redacted cases and a fixed comparison sample. It measures cycle time, required corrections, error types, cost and adoption — not merely a subjective wow effect.
