Context
Someone arrives with a process, a hoped-for saving, and a request to automate it. The damage starts when scoring lives in a spreadsheet nobody can replay: the assumptions that funded the work disappear.
The Automation CoE Toolkit exists so those assumptions stay inspectable.
Decision
A bounded JSON intake goes through schema validation, then value, feasibility, risk, and priority scores. Annual benefit, risk-adjusted cost, payback, and a three-year ROI are calculated from the numbers you supplied. Evidence, security, privacy, and continuity gates can still hold a candidate when the score looks attractive. The output is decision support.
What I rejected
A model that invents a business case from a paragraph of process description, and an autonomous “fund this” action. Both hide the arithmetic and move money or identity without a person on the record.
Failure scenario
A high-ROI access-review proposal is missing attestation. The toolkit still produces a plan: retain, revoke, or send to manual review. Hash-chained audit records and a visible side_effects_executed: false flag keep privilege changes with a person.
Evidence
Python CLI, Pydantic intake, explainable scores, and the access-review example that refuses to execute side effects.
Inspect: Repository · Run locally