Agent Governance for Financial Services Teams

Financial services teams ask three questions about agent automation before they expand it: who approved the change, which model produced it, and what record survives the release. Modulo answers all three as the work happens.

Segregation of duties

A pipeline step can require human approval before it continues, and a human-only gate cannot be claimed or decided by an agent principal, so a workflow cannot approve its own output. Decisions run through a claim and decision flow with an exclusive claim window, and every decision writes an audit event whether it approves or rejects. Organization roles (admin, operator, runner, viewer) and team roles scope who can create, approve, and view.

The model record

Every agent binds to a named model backend, and each run freezes the whole pipeline definition into a snapshot before execution: the agent versions, the schema pins, the connector bindings, and the model backend. A run paused for review and resumed later behaves the same, and the record shows which model and which prompt version produced each output. When a model risk review asks which model touched a system, the run holds the answer.

The evidence layer

Every state-changing action appends an audit event, and each event carries the SHA-256 hash of the previous one. Events are append-only at the application layer and at the database layer, the chain is verifiable on demand, and it exports to CSV for an internal audit or a supervisory request. The chain lives in your database, where you control retention.

Spend as a control

Daily spend limits apply at the organization and team level, and a run that would exceed a limit is rejected before it starts. Cost reports export to CSV and can be scheduled, and a day whose spend exceeds twice the rolling seven-day average is flagged for review. Cost control reads as a risk control when an agent can loop.

Fit to your framework

An internal audit, a supervisory request, or a model risk review asks for the same thing each time: a record of who approved what, and on what basis. The frameworks differ (SR 11-7, DORA, PCI DSS, your own change policy) and each one lands on your institution. Modulo supplies the mechanism those frameworks expect to find: approvals a named human made, a frozen version of what ran, and an audit trail that cannot be edited after the fact.

Placement is self-hosted, model calls run through your own provider contracts, and the audit chain sits in your database. How the split works.