Credit decisions, autonomy by fair lending risk
Auto approve the clear cases; the thin file applicant stays with a human.
Open the live lab · preloaded to this scenario
Human Review and Autonomy Control Simulator
Context
An agent processes credit decisions. Autonomy is set by risk tier: clear approvals/denials can auto decide, but thin file and borderline applicants, where fair lending exposure lives, stay with a human.
The decision
Autonomy is gated by fair lending risk, not throughput appetite. The level that clears the thin file edge cases is the ceiling, regardless of how much faster full autonomy would be.
What most miss
Teams set one global autonomy level. On credit the tier matters: a thin file auto denial that a human would have caught is exactly the disparate impact pattern regulators look for.
Stakes
One auto decided thin file case that should have been reviewed is a fair lending exposure, not a throughput win.
Firsthand · Agent Architecture and Protocol Strategy Artifacts · verified 2026-07-03
Sources: Credit-decisioning autonomy / fair-lending (ECOA) controls, firsthand (cards & lending); Risk-tiered human in the loop policy design