Trimio Field Notes

Your CLAUDE.md has a hole in it.

June 4, 2026 5 min read governancecompliancesecuritypolicy-enforcement

Organizations are writing AI governance policies at a record pace. The number of repos with a CLAUDE.md file grew from roughly 2,000 in early 2025 to over 90,000 by mid-2026, as teams realized that without documented guidance, AI coding assistants tend to operate as shadow IT. But here's what's rarely discussed: the policy is the easy part. Enforcing it is where things break down.

90K+
CLAUDE.md files
Repos with explicit AI usage guidelines — advisory, not enforced.
0
Enforcement points
Most policies sit in a repo file. No infrastructure control touches them.
100%
SOC 2 gap
"How do you enforce your AI policy?" → silence or hand-waving.

The enforcement gap is structural, not intentional

Essential
A CLAUDE.md file is guidance for a model, not a control for an infrastructure. The API call still originates from the developer's machine and bypasses any policy gate. Writing the policy and enforcing it are two completely different engineering problems.

A CLAUDE.md file tells a model how to behave. It sits in a repo root, gets read into context, and influences the model's output. That's its design — and it works reasonably well for that purpose.

What it doesn't do: intercept API calls, route based on content classification, block calls to non-approved providers, or generate audit logs. When an engineer runs Claude Code or any AI coding tool from their laptop, the model call goes directly to the API endpoint. The policy file is advisory context — not an infrastructure control.

The gap isn't a bug. It's a structural mismatch. The policy exists at the file level; enforcement needs to happen at the network level.

Stanford's framework became a SOC 2 control

Essential
Stanford HAI's AI governance frameworks started as research papers, got adopted as corporate policy, and landed in SOC 2 audit checklists. The question auditors now ask: "show us your enforcement evidence." The answer most companies can't provide — until now.

Beginning in late 2025, Stanford HAI's series of AI governance frameworks began appearing in enterprise compliance programs — initially as internal policy documents, then as de facto standards, and eventually as items in SOC 2 Type II control mappings. Security auditors started asking questions like:

For the first two questions, most companies have answers: a model allowlist, a data classification scheme. For the third — how do you actually enforce it — the answers get vague fast. "We trust our engineers" is not a SOC 2 control.

The auditors know it. The companies know it. The gap is real and getting wider as AI tool adoption scales.

The routing layer is the enforcement point

Essential
The LLM proxy sits exactly where enforcement needs to happen: between the call and the provider. Not inside the model (can't control that), not in a separate policy service (adds latency). The routing layer intercepts every call and applies policy as infrastructure — routing rules, not readmes.

The LLM proxy occupies the one infrastructure layer that sees every AI API call. It's the natural enforcement point — closer to the call than any code linter, and more reliable than a policy file someone might forget to read.

What that looks like concretely:

None of this requires code changes in the calling application. The enforcement is in the routing layer, invisible to the developer, auditable by the compliance team.

The compliance artifact that actually satisfies an auditor

Essential
A policy document is evidence of intent. An audit log from a routing layer is evidence of enforcement. Same effort to produce, completely different compliance value. The SOC 2 control isn't "we have a CLAUDE.md" — it's "our proxy blocked 847 non-compliant calls this quarter."

SOC 2 auditors have a specific evidentiary bar: not "what's your policy" but "show us how you enforce it." A CLAUDE.md file demonstrates intention. An audit log from a routing layer demonstrates control.

The difference in compliance value is categorical. Consider:

One satisfies a control mapping. The other doesn't. The routing layer generates the second type of artifact automatically — every call is a log entry, every block is a compliance event.

The enforcement hierarchy

Essential
Enforcement strength runs from weakest (advisory files) to strongest (proxy-layer blocks). Organizations typically progress through all four stages as AI usage scales. Stage 4 — infrastructure-level enforcement — is the only one that holds at 200+ engineers.
Enforcement levelMechanismReliability at scale
AdvisoryCLAUDE.md fileDepends on individual compliance
Code linterPre-commit hooks catch known patternsMisses anything not in the linter rules
Policy serviceSeparate microservice evaluates each callReliable but adds latency and operational overhead
Proxy-layer enforcementRouting rules block/redirect at the gatewayAutomatic, zero developer friction, full audit trail

Most organizations progress through this hierarchy as AI usage scales. At 10 engineers, the advisory file works fine. At 100, someone writes a linter. At 200+, the only control that holds is infrastructure-level — the enforcement needs to be invisible to the developer, automatic in the call path, and auditable by the compliance team.

What enforcement actually looks like in practice

Three scenarios where proxy-layer policy enforcement changes the compliance posture:

Compliance team needs GDPR-sensitive workloads routed to EU-deployed models only. Without a proxy: a document that engineers are expected to read before writing code that touches customer data. With a proxy: a routing rule — if the call contains a customer_id field or other PII signal, route to eu-west-1 deployed models only, block everything else, log the call. The policy is the infrastructure.

Security team needs "no AI calls in CI pipelines" enforced automatically. Without a proxy: a CLAUDE.md note in the repo's engineering guidelines. With a proxy: a routing rule that returns 429 for any call originating from a CI environment, with a log entry for every rejection. Engineers can't accidentally bypass it because the enforcement is in the routing layer, not the documentation.

CFO needs a hard cap on AI spend per team, enforced at the API level. Without a proxy: spend alerts that notify after the bill arrives. With a proxy: a hard cap that returns 429 when the team hits its budget ceiling, with per-call attribution logged so finance can see exactly what generated the spend.

In each case, the proxy makes the policy the infrastructure — not a document someone's supposed to remember, but a control that's built into the call path.

The control plane for enterprise AI

Essential
AI governance is moving from advisory (files, guidelines, intentions) to enforced (routing rules, blocks, audit logs). The enforcement mechanism isn't a policy document — it's the infrastructure layer closest to the API call. That's the routing layer, and that's where the SOC 2 control lives.

The underlying shift is simple: AI governance is becoming a first-class infrastructure concern, not a developer culture concern. The questions compliance and security teams are asking — "which models are we authorized to use, how do we enforce data residency, what's our hard cap" — are the same questions asked about every other critical infrastructure dependency. And the answer is the same: you enforce it in the control plane.

For AI, that control plane is the routing layer. Trimio's LLM proxy sits at that layer — intercepts every call, applies policy as infrastructure, generates audit logs that satisfy SOC 2 control mappings, and enforces CLAUDE.md-equivalent governance automatically, without developer friction.

The policy document is the starting point. The routing layer is the control.

Trimio handles AI policy enforcement at the routing layer — model allowlists, data residency routing, spend caps, and full audit logs. See how it works →

Finance & Engineering

See your AI spend at team granularity.

Trimio routes every LLM call through a single endpoint with live cost attribution by team, model, and workflow. No code changes required.

Field Notes

Data-led essays on AI cost governance.

One email per week. No filler. Written for finance and technical leaders running AI at scale.