Human-in-the-Loop Design Regulators Actually Accept
Full autonomy is rarely the right answer in enterprise, especially in banks. The design skill is deciding exactly where a human must approve, and making that approval fast enough that ops teams do not bypass it.
The problem in prod
- Agents that draft compliance reports, move money logic, or touch customer records cannot fail silently.
- Review queues designed by engineers get skipped by operators under time pressure.
- Regulators ask who approved what, and clickwrap logs do not count as evidence.
Loop engineering
Classify every agent action: auto-approved (read-only summaries), one-click approve (drafts, flags), committee review (external effects). BBVA runs thousands of GPTs inside exactly this kind of tiered trust.
Every override or edit by a reviewer becomes a labeled training and eval case. Override rate per action type is the metric that tells you when to widen or narrow autonomy.
Each approval records who, what version, what evidence was shown, and the policy cited. Exportable per case for BNM or internal audit.
Production hardening checklist
- Default deny on external actions. Agent proposes, human disposes.
- Show the reviewer the exact retrieved sources, never just the draft.
- Time-box approvals with escalation, so queues cannot stall operations.
- Simulate reviewer absence: deputy routing and break-glass with full logging.
Deploying this in MY
Malaysian compliance teams run Shariah review cycles on top of standard control checks. Design the approval chain with both gates visible from day one, plus explainability notes for credit and underwriting-adjacent drafts.
Malaysia FDE guide →