AI agents in lending are shipping faster than compliance can audit them

Here's what keeps me up at night: AI agents are moving into loan origination, underwriting, and borrower data pipelines right now, and the compliance infrastructure to govern them is still being sketched on a whiteboard.

This week alone we saw Snowflake launch a Cortex AI Gateway with governance tooling, Diagrid drop agent recovery patterns, and IBISWorld partner with Lama AI to embed industry intelligence into commercial lending AI. The velocity is real. The ambition is real. The governance? Nonexistent.

Let me be concrete. An AI agent touching a loan file does three things regulators care about: (1) ingests borrower PII and financial data, (2) makes or influences underwriting decisions, and (3) leaves an audit trail—or doesn't. Today, most of these systems lack deterministic logging. An agent hallucinates a credit score or misroutes a file, and the lender has no clean record of why or when. Snowflake's new governance layer is a step—it tracks activity and cost—but it doesn't solve the hard problem: proving to a regulator that an autonomous agent's decision chain was lawful.

Here's the gap: FCRA, ECOA, the CFPB's AI guidance, and fair lending enforcement all demand explainability and auditability. But the tooling shipped today (agent frameworks, MCP servers, prompt-orchestration platforms) was built for speed and feature parity, not for legal proof. When an agent makes a credit decision, can you show the examiner the exact weights, data inputs, and inference rules? Can you replay it? Most shops will say "sort of." Regulators will say "not good enough."

The lending angle is sharp because stakes are high. A mortgage originator or auto lender running agents on underwriting is already subject to bias testing, disparate impact analysis, and third-party vendor audits. Drop an unvetted agentic system into that flow and you've got a compliance violation waiting to happen. Blend Labs—a major platform powering digital mortgage workflows—and platforms like it are under pressure to move fast. But moving fast into regulated lending without deterministic audit trails is malpractice.

What should happen: enterprises need to demand agent governance before production, not after. That means:

• Deterministic logging. Every decision an agent makes—every API call, every inference, every branch—gets timestamped, signed, and stored immutably. No exceptions.
• Explainability on demand. Regulators should be able to ask "why did this agent approve or deny this borrower?" and get a replay of the reasoning, not a hand-wave.
• Bias testing built in. Before an agent touches a live borrower file, it should be tested for disparate impact across protected classes. Not optional.
• Vendor accountability. If you're buying agentic tooling from a framework vendor or MCP server, demand that they certify their systems for regulated use. Most won't yet.

What's actually happening: vendors are shipping governance layers (Snowflake's cost and activity tracking is a start), but nobody is shipping compliance-grade agentic infrastructure. There's daylight between "we can see what the agent did" and "we can prove to a regulator why it was lawful." Enterprises are standing in that gap.

The regulatory silence is deafening. The CFPB, OCC, and Federal Reserve have issued guidance on AI in lending, but none of it was written with autonomous agents in mind. Agents are different from static ML models—they branch, they call external APIs, they fail and retry. A regulator's checklist built for a credit-scoring model doesn't fit an agentic workflow. So lenders are shipping anyway, hoping compliance catches up.

It won't. Not in time. By the time the first enforcement action lands, dozens of lenders will have live agents touching millions of borrower files. The bank examiners will walk in, ask for an audit trail, and watch the CRO's face go pale.

This is not theoretical risk. This is "we're moving faster than we should be" risk. If you're shipping AI agents into regulated lending, demand deterministic governance first. If you're a vendor, build for compliance or stay out of the category.

Regulators will catch up. But they'll catch up after the mess.

From my toolbox — something I actually ship, not just write about:

rebar-mcp — Reinforcement for AI-generated code. Enforcement hooks, quality audits, and opinionated templates that stop broken code from shipping. · ~15/wk on npm