AI Governance & Regulation
Essays on AI governance in regulated finance: NIST AI RMF and ISO 42001 mapping, fair lending, shadow AI, and controls that survive an audit.
AI governance from the practitioner's side: mapping to NIST AI RMF and ISO 42001, fair lending as the real governance problem in consumer credit, shadow AI, and why a policy PDF your agents can't read isn't a control. Written by Michael York, who stood up AI governance for a regulated fintech platform. These essays treat governance as an engineering discipline, not a document.
27 posts, newest first
Your AI Contract Covers a Third of the Traffic
Every enterprise AI control binds to an account. Two thirds of AI users on corporate devices use personal logins, so your contract governs a minority of use.
The Model Was Bad Is Not a Root Cause
A model defect, a provider stall, a bad route, and a broken check are four separate findings. If your AI failure log has one bucket, none can be acted on.
Four Verbs, and None of Them Is Delete
A memory store's tool manifest is its real retention control. When the verbs read capture, search, list, count, you built a system of record with no eraser.
Open Weights Are Not a Residency Plan
I caveated self-hosted open weights on capability, not infrastructure. Near-frontier weights need a rack, and downloading them makes you the model provider.
Your Agent Decides Which Record Wins
Reconstructing a customer across six systems means deciding which record wins. Hand it to an agent and unwritten precedence becomes code nobody approved.
Sanitized Is Not Authorized
"PII redaction before egress" is filed as the data-handling control. It answers whether identifiers remain, not whether this content may go to this destination.
Country of Origin Is Not a Control
Risk committees ask whether to allow Chinese models as if nationality were a control. Your regulator defines foreign by where the work happens, not the flag.
Your Tool Catalog Is Standing Access
A resident tool definition is a standing grant with no request, no approver, and no expiry. We scoped the MCP servers and never scoped the catalog.
Fair Lending: The Real AI Governance Problem
Mapping models to AI frameworks isn't governance for credit. The binding constraint is fair-lending law: ECOA/Reg B, FCRA adverse action, SR 11-7 model risk.
'We Don't Train on Your Data' Is Not Enough
An agent told to open no files obeyed, while the product uploaded the whole repo, canary included. "We don't train on your data" answers the wrong question.
Your Prompt Is the Approval. That's the Gap.
An MCP connector executes writes with no approval screen: your prompt becomes the one boundary nobody governed. That missing gate is a control-plane gap.
Your AI Policy Is a PDF. Agents Can't Read It
A model given thousands of extra words wrote better prose, and failed the delivery contract two runs in three. Rules agents can ignore fail audits.
Shadow AI: Your "Personal Tool" Is Production
A coding agent stands up a data-touching tool in an afternoon. The moment it needs a login or gets shared, it's a production system nobody reviewed.
Stop Gating the $40 Question
Two hours and $40 did what a top engineer says he couldn't, and no routing table would have assigned it. Gating frontier access defunds your own sensing.
The Second Agent Cheap, the Fiftieth Boring
Build cost was never the constraint in a regulated shop: agent #50 hits the same security review as agent #1. Make identity and action gates reusable.
Context Custody Is a Concentration Risk
Intelligence went cheap, yet enterprise buyers expect to pay more for Claude. You're not paying for the brain. You're paying for where your context lives.
Why We Built AgentOS
One model scored 78% in one agent harness and 42% in another. In regulated fintech the harness is where governance lives, so we built our own: AgentOS.
Bake Audit Evidence Into Your AI Pipeline
Audit-defensibility isn't a document you write after the fact. It's a property you engineer into the AI pipeline so its operation emits evidence as exhaust.
The 2026 AI Regulatory Map on One Page
Everyone read 'EU AI Act deferred to 2027' and exhaled, but the part fining 3% of global revenue turns on in August. The four 2026 rules with teeth.
Your Agent Dashboard Is Green and Lying
Uptime tiles tell you the service answered, nothing about whether the answer was right. That gap is where a model-risk review will eat you alive.
The Boundary Layer Is the Actual AI Control
Every AI governance framework describes the same controls. The one that matters is a design decision: does this output get acted on, or interpreted first?
Three Token Counts, Zero You Can Attest To
Codex says one number, Claude another, your gateway a third. That isn't a metering problem. It's an attestation problem regulated industries can't afford.
Concentration Risk in the Three-Lab AI Stack
Most of the AI on your roadmap traces to three labs on the same chips, supply chain, and balance sheets. That's a concentration risk your board hasn't priced.
Dark Code Is a Control Failure, Not Tech Debt
AI is filling repos with code nobody can explain. We call it tech debt; it's a control failure, and it should fail CI like a missing approver does.
Shadow-Agent Discovery for Regulated FIs
Unsanctioned AI agents already run in your environment with your credentials. Find, classify, and gate them before they touch member data or an exam does.
An AI Agent Dropped Prod: The Change Playbook
Coding agents are committing real change to real systems. The question isn't whether to let them. It's how to give them speed without a SOC 2-fatal mistake.
Agent Memory Is a Data-Residency Problem
Give every agent a durable, MCP-connected brain and you've stood up a new data lake of PII and PCI scope nobody classified, encrypted, or can purge.