AI triage with humans in the loop — why we don't auto-deny at MVP
AI triage with humans in the loop — why we don't auto-deny at MVP
AI is good at reading documents and bad at calibrating denial decisions on small datasets. Our pilot triage runs in human-review mode by design.
What the AI does
The triage pipeline reads uploaded documents, extracts evidence, classifies the refund request against the 12 covered reasons, and produces a recommendation with a confidence score. The recommendation is a starting point for the human reviewer, not a binding outcome.
What a human does
A human reviewer reads the AI's recommendation, scans the uploaded documents themselves, and either approves or denies. They can also send back for more information. The decision is logged with the reviewer's identity, the time, and a free-text rationale.
Why no auto-deny at MVP
Three reasons. First, calibration — we don't have enough labeled data yet to know the AI's denial precision in production conditions. Second, consumer protection — a wrongful auto-denial is harder to undo than a wrongful approval. Third, learning — every human decision is training data for the calibration phase that gates auto-approval.
What changes after calibration
Once the labeled-decision corpus reaches the threshold and the calibration analysis shows reliable performance, hard-cases auto-approve becomes a real option for clean-classified situations. Denial automation lives further down the roadmap, behind a confidence floor that we will publish before we use it.