For AI in Financial Services

Runtime evidence for AI in regulated financial workflows.

When the OCC, CFPB, NYDFS, or a bank counterparty asks what your AI workflow actually does at runtime, an MRM binder and a SOC 2 letter aren’t the answer. Glacis runs inside your infrastructure, instruments one production AI workflow with runtime controls, and produces signed evidence receipts MRM, internal audit, and examiners can verify — with zero sensitive-data egress.

Or read the SR 11-7 & AI guide →

The gap

SR 11-7 was written before generative AI.

SR 11-7 expects effective challenge, independent validation, and ongoing monitoring. Your model risk program documents all of it — but the guidance was written before generative AI, before non-deterministic outputs, before agentic workflows that call tools and rewrite data.

Examiners and counterparties have shifted from “show me the policy” to “show me the runtime evidence.” OCC, CFPB, and NYDFS inquiries increasingly ask which controls fired on a specific decision, not which controls exist on paper.

The Sprint scopes that runtime layer on one production AI workflow in three weeks — runtime controls, signed receipts, and an evidence pack that maps to your existing MRM framework.

How the sprint runs

Inside your stack. Mapped to your MRM framework.

Glacis instruments runtime controls beside your existing model-risk pipeline, signs each control outcome, and packages an evidence pack MRM and examiners can verify without seeing customer data.

Runtime controls in your stack

Bias checks, output validation, human-review gates, content filters — the controls your MRM team already specified now execute and emit signed evidence at runtime.

Zero sensitive-data egress

Customer data, model inputs, and proprietary algorithms stay inside your infrastructure. Hashes and signed metadata are the only things that cross the boundary.

Pack for MRM & examiners

Timestamped, third-party notarized, cryptographically signed receipts assembled into an evidence pack mapped to SR 11-7, fair-lending, and counterparty review language.

Where evidence matters most

Four AI workflows under model-risk review.

01 / VALIDATION

Model validation

Prove your validation tests actually ran against production models. Not recreated for audit, not simulated — the real thing, timestamped and notarized.

EXAMINER · “Show me the validation runs on this quarter’s model.”

02 / MONITORING

Ongoing monitoring

Every control check, every threshold evaluation, every human review decision — captured as verifiable evidence. Continuous, not periodic.

EXAMINER · “What controls fired between examinations?”

03 / FAIR LENDING

Fair-lending compliance

Prove your bias controls executed on every decision. Cryptographic evidence that fairness checks ran — without exposing individual applications.

CFPB · “Did the fairness check run on this denial?”

04 / VENDOR AI

Third-party AI oversight

When you use vendor AI, prove your oversight controls executed. Evidence that you validated vendor outputs, not just that you have a policy to.

OCC · “Show me the oversight on the vendor model.”

What MRM gets

Receipts at runtime. Pack on examination.

SDK · TypeScriptWrap a model-risk control.

import { attest } from '@glacis/runtime';

const receipt = await attest({
  workflow: 'underwrite.decision',
  policy:   'mrm.fairlending.v3',
  decision: 'PASS',
  rules:    ['fairness.adverse_impact'],
});

// → signed OVERT receipt; no customer data egress

MRM packet · OVERT 1.0What the examiner reviews.

Evidence PackUnderwriting AI · Q2 2026
Verified
Issuer
did:web:notary.glacis.io
Workflow
underwrite.decision · v4.1
Decisions
87,412 control executions
Fair-lending
87,412 / 87,412 · check fired
Schema
overt://schema/v1.0/runtime-attestation
Maps to
SR 11-7 · OCC 2021-39 · NYDFS
ED25519 · ed25519-2026-q2 · chain depth 87,412

Direction of travel

From “documented” to “operational.”

The OCC, Fed, and FDIC are paying attention. The EU AI Act treats credit scoring as high-risk. State regulators are adding AI-specific requirements to existing frameworks.

The pattern is consistent: regulators want evidence that AI governance is operational, not just documented. They want to see that controls executed, not just that they were planned.

Institutions that can demonstrate continuous, verifiable AI governance will face less friction. Those that can’t will face more scrutiny, more MRAs, and more constraints on AI adoption.

Related guides

Framework-specific reading.

Scope a model-risk evidence sprint on one production AI workflow.

Three weeks. One production AI workflow. Runtime controls instrumented inside your stack, and an evidence pack mapped to SR 11-7 and counterparty review. No rip-and-replace.

Or learn how continuous attestation works →