Governance is now a plugin

In May, Anthropic launched Claude for Legal: twelve practice-area plugins and more than twenty connectors into the systems legal teams already run on, available to all paying Claude customers. One of the twelve is ai-governance-legal, a plugin for the legal and compliance side of AI adoption. It drafts a firm’s AI usage policy from published templates. It builds an EU AI Act inventory that assesses each system’s role and risk tier individually. It triages proposed use cases, generates impact assessments, reviews vendor AI terms, diffs new regulation against your current posture, and sweeps for drift between what the policy says and what practice does.

Every one of those outputs is a governance artefact. The policy, the register, the assessment, the gap analysis. The paperwork of being in control. Producing that paperwork used to take a quarter. It now takes an afternoon.

The detection is real

Credit where it is due, because the detection machinery here is genuinely good. Citations that come from model memory rather than a live research source are flagged for verification, and if no research tool is connected at all, a reviewer note records that sources were not checked. Deviations from your playbook get logged rather than smoothed over. A scheduled agent sweeps saved assessments, triage results and vendor reviews for drift.

That is detection at authoring time, and it is exactly where detection belongs: at the moment the work is made, not reconstructed afterwards. The EU AI Act inventory deserves a separate mention, because forcing role and tier to be assessed per system rather than per company is the analysis regulators actually expect and most registers quietly avoid.

Nothing blocks

Now notice what happens to the flag. Nothing blocks on it. The verification flag, the drift report and the deviation log all land on a human desk. Anthropic is explicit about this by design: every output is a draft for attorney review, the tooling makes that review faster, and the reviewer carries responsibility for anything that leaves the building. The framing is honest and it is the correct legal posture. It is also a boundary worth naming precisely.

A control that is written down is documentation. A control that demonstrably changes behaviour at the point of action is governance. This tooling industrialises the first at a rate nothing before it could touch. It cannot, by design, do the second. And the review capacity it relies on does not scale with the production rate it creates. When governance artefacts become cheap and plentiful, a model of assurance built on a person reading carefully does not survive contact with the volume.

The tell in the repository

There is one place in the same suite where the control is not a person reading carefully. Before a community-built skill can enter your environment, it passes a mandatory trust check, a hidden-content scan and a restrictive-by-default allowlist. A non-lawyer does not even get an install button; they are routed to their attorney.

Read that asymmetry closely. Where code enters the environment, the builders installed a gate that refuses. Where governance documents leave it, they installed a reviewer who is already busy. That placement tells you where the builders believe risk really lives, and it quietly concedes the point: when it matters, a gate that enforces beats a reader who checks.

From Complete to Substantive

On the Cognitive Delivery Index, this tooling accelerates the climb to Level 4, Complete: the artefact exists, is current, and is maintained. Level 5, Substantive, asks the harder question. Does the control demonstrably change behaviour at the point of action, enforced outside the process that produced the work?

That last step is a wiring job, and it is yours. The pattern is the same for every output the plugin produces: take the artefact’s existence and its verdict, and make something outside the authoring process refuse to proceed without them. A system absent from the inventory does not deploy. A conditional triage verdict becomes tracked checks that hold promotion until the conditions close. A missing impact assessment fails the gate rather than generating a reminder. An open drift finding above threshold suspends the affected approval.

None of that requires the plugin to change. It requires the organisation installing it to decide that the artefacts it now produces so cheaply are preconditions, not records.

The plugin industrialises the paperwork of being in control. Whether you are in control is decided at the gate, and the gate is still yours to build.

A companion field guide accompanies this note: how to run the plugin well, an honest map of where it performs and where it stops, what to pair it with, and the Level 4 to Level 5 wiring for each of its outputs. Product details are drawn from the public anthropics/claude-for-legal repository as read on 17 August 2026; this is an independent practitioner’s analysis, not affiliated with or endorsed by Anthropic.

Jason Cruz Falzon is the creator of the Cognitive Delivery Framework and the Cognitive Delivery Index. cognitivedelivery.co.uk · jason@cognitivedelivery.co.uk