Is this AI agent worthbuilding — or will the controls eat the ROI?

Because the fit score is context, not the decision. The Agent Use-Case Fit & Proof-of-Value Gate runs two worsen-only structural gates on top of the score, and the first one is dispositive: if the projected return doesn't survive the mitigation the agent needs — the controls, validation layers, and human oversight — net return is zero or negative and the verdict is DON'T BUILD no matter how strong the fit looks. The refund-agent example scores 94 and still reads DON'T BUILD because its $140k of required mitigation outruns its $120k return. A high score can't rescue money that doesn't work.

Each proposed agent is marked MISSING / PARTIAL / CLEAR (0/1/2) on six use-case-fit dimensions, weighted to 100: genuine autonomy need (28), bounded scope (20), integration readiness (16), measurable outcome defined (16), ROI survives mitigation (12), and failure fallback (8). The weighted marks produce the 0–100 fit score. Then you enter two dollar figures — annual mitigation cost and projected annual return — and the gate runs the money against the marks to land BUILD, PILOT FIRST, or DON'T BUILD.

DON'T BUILD fires when either gate trips — the mitigation erases the ROI, or there's no genuine autonomy need (automation rebranded as an agent). BUILD is earned: no gate fires, fit is 75 or higher, and no dimension is MISSING — strong fit, the return survives the controls, nothing missing. PILOT FIRST is the floor for any gate-clear project that isn't a clean BUILD yet — worth a bounded pilot to prove the value on a low-risk path before you commit the budget.

Score a whole slate of proposed agents and the gate rolls them into one program verdict — SLATE READY (every project is BUILD), SOME TO PILOT (no DON'T BUILD, at least one PILOT FIRST), or STOP-FIRST (at least one DON'T BUILD) — and names the single project to work first (worst verdict, then lowest fit). The mean fit is shown as context only; it never decides anything, because an average can quietly promote a slate that has a stop-first project hiding in it.

No. This is a pre-build decision aid — it scores the proposal you describe from your own marks and your own two dollar figures, entirely offline. It never builds, runs, deploys, or connects to an agent, and it uploads nothing. Once you decide to build, the operational control check is a separate step: the AI Agent Go-Live Readiness Gate grades the live controls, and the Agent Reliability Harness evaluates a built agent at the trajectory level.

No. It's a planning decision aid that scores the proposal you describe, not financial, investment, or legal advice, and it scores no person — only the proposed use case and the numbers you supply. The cost and return figures are your own estimates; confirm them before you act on a verdict. The deterministic Python engine, the workbook that reproduces it, and the on-page demo all run the same logic, so the verdict is auditable and repeatable rather than a black box.