1. Read

   ~/.claude/plugins/config/claude-for-legal/ai-governance-legal/CLAUDE.md

   . Confirm vendor governance positions are populated — if not, stop and direct to setup.
2. Use the framework below.
3. Confirm document type (AI addendum / main agreement AI provisions / ToS). If only an AUP was provided, ask for the full terms.
4. Term-by-term review: training on data, confidentiality of inputs, model changes, output IP, liability, incident notification, human review rights, use restrictions, audit rights.
5. AI addendum gap check if DPA exists but no AI addendum.
6. AI policy consistency diff vs.

   ~/.claude/plugins/config/claude-for-legal/ai-governance-legal/CLAUDE.md

   .
7. Output: bottom line, term-by-term, recommended redlines, if-they-won't-move routing.

/ai-governance-legal:vendor-ai-review openai-enterprise-agreement.pdf

• A standalone AI agreement or AI addendum (most structured)
• A vendor's universal terms of service with AI provisions embedded (often buried)
• An acceptable use policy (tells you what you can't do; says nothing about what the vendor can do with your data or outputs)
• A combination — master agreement + DPA + AI addendum (common for serious enterprise AI vendors)

~/.claude/plugins/config/claude-for-legal/ai-governance-legal/CLAUDE.md

## Vendor AI governance

## AI policy commitments

~/.claude/plugins/config/claude-for-legal/ai-governance-legal/CLAUDE.md

[PLACEHOLDER]

I notice you haven't configured your practice profile yet — that's how I tailor vendor governance positions to your practice.

Two choices:

• Run

  /ai-governance-legal:cold-start-interview

  (2 minutes) to configure your profile, then I'll review tailored to YOUR positions.
• Say "provisional" and I'll review against generic defaults — US jurisdiction, middle risk appetite, lawyer role, no playbook — and tag every output

  [PROVISIONAL — configure your profile for tailored output]

  so you can see what I do before committing.

[PROVISIONAL]

"That was a generic run against default assumptions. Run

/ai-governance-legal:cold-start-interview

to get output calibrated to YOUR practice — your vendor governance positions, your jurisdiction, your risk appetite. 2 minutes."

"Can you share the vendor's AI terms? The most useful thing is the actual contract language — the AI addendum if there is one, or the main agreement with AI provisions highlighted. An acceptable use policy alone won't tell us what the vendor can do with our inputs; it only tells us what we're allowed to do."

"This is the acceptable use policy — it tells us what we can't do with the vendor's AI. That's useful context, but it doesn't address the commercial terms: whether the vendor can train on our data, what their liability is for AI errors, whether they notify us when the model changes. Do you have the service agreement or AI addendum?"

~/.claude/plugins/config/claude-for-legal/ai-governance-legal/CLAUDE.md

[Term name] 🟢 / 🟡 / 🟠 / 🔴 Vendor says: [summary of what the contract actually says] Our position: [from

~/.claude/plugins/config/claude-for-legal/ai-governance-legal/CLAUDE.md

] Gap: [specific delta — or "Aligned"] Proposed fix: [specific redline language, or "escalate — outside fallback"]

~/.claude/plugins/config/claude-for-legal/ai-governance-legal/CLAUDE.md

• 🟢 Aligned — at or better than the standard position in the playbook.
• 🟡 Note — within fallback but worse than standard; flag for awareness, not a blocker.
• 🟠 Significant — outside standard position but within fallback; needs redline before signing.
• 🔴 Critical — outside fallback; deployment should not proceed without resolution. Escalate per

  ~/.claude/plugins/config/claude-for-legal/ai-governance-legal/CLAUDE.md

  .

"There's a DPA in place but no AI-specific addendum. The DPA covers data protection obligations but doesn't address: training on our data, model change notification, liability for AI outputs, or incident notification for AI system failures.

For a [Standard / Elevated / High] tier use case, this gap is [acceptable at Standard tier / a blocker at Elevated or High tier]. Recommend requesting an AI addendum or at minimum negotiating AI-specific terms into the next renewal."

"There are no AI-specific terms in this agreement. The vendor is providing an AI-powered service under general service terms — which means we have no contractual protection on the highest-risk AI governance items (training, liability, model changes). This is a 🔴 for any Elevated or High tier use case."

~/.claude/plugins/config/claude-for-legal/ai-governance-legal/CLAUDE.md

• Our policy prohibits vendor training on our data — the vendor's terms permit it by default. (Contract needs explicit prohibition or opt-out confirmation.)
• Our policy requires human review for certain use cases — vendor's terms say AI outputs are final. (Workflow needs to impose the human step, not the vendor terms.)
• Our approved vendor list doesn't include this vendor — or blocklist does.
• Our policy requires disclosure to affected parties — vendor's terms impose a confidentiality obligation on AI system capabilities that would prevent disclosure.

• Replace a word before a phrase. ("twelve (12)" → "twenty-four (24)")
• Replace a phrase before a sentence. ("paid by the Buyer" → "paid and payable by the Buyer")
• Restructure a subclause before replacing the sentence. (Add "(a)" and "(b)" to split a compound condition.)
• Replace a sentence before replacing the clause.
• Only replace a whole clause when the counterparty's version is so far from your position that surgical edits would be harder to read than a fresh draft — and when you do, say so in the transmittal: "We've replaced §8.2 rather than marking it up because the changes were extensive. Happy to walk you through the delta."

## Who's using this

~/.claude/plugins/config/claude-for-legal/ai-governance-legal/CLAUDE.md

Signing this vendor AI agreement has legal consequences. Have you reviewed this with an attorney? If yes, proceed. If no, here's a brief to bring to them:

[Generate a 1-page summary: the vendor and the use case, the key terms reviewed (data use, liability, auditability, model change, human review), where vendor positions diverge from policy, what's being accepted, what could go wrong, what to ask the attorney.]

If you need to find an attorney, solicitor, barrister, or other authorised legal professional: your professional regulator's referral service is the fastest starting point (state bar in the US, SRA/Bar Standards Board in England & Wales, Law Society in Scotland/NI/Ireland/Canada/Australia, or your jurisdiction's equivalent).

[WORK-PRODUCT HEADER — per plugin config ## Outputs — differs by role; see `## Who's using this`]

*This review is derived from vendor contract terms that are typically confidential under NDA, and it may itself be privileged. It inherits the source's confidentiality and privilege status. Distributing it beyond the privilege circle (e.g., forwarding to the vendor, sharing in an open channel) can waive privilege and breach the NDA. Mark, store, and route accordingly.*

1. End-user SaaS application (e.g., a legal tech tool, a CRM with AI scoring, a document assistant) — the tool your org signs up for
2. API gateway / orchestration layer (e.g., Azure OpenAI Service, AWS Bedrock, Google Vertex, LangChain-hosted) — often invisible, always has its own terms
3. Model provider (e.g., Anthropic, OpenAI, Google, Meta) — the LLM
4. Hosted knowledge base / RAG source (e.g., a vector database, a third-party data corpus, a retrieval service) — the data Claude reads from
5. Additional subprocessors — analytics, logging, fine-tuning partners

1. Search the contract for flow-down language. Look for: "subprocessor obligations no less protective than," "flow-down of data commitments," "back-to-back terms," "Provider shall ensure that its subprocessors are bound by," "equivalent obligations."
2. If present: Quote it, verify it covers the specific flagged term, and flag whether it's enforceable (who can enforce it — you, or only the intermediate vendor?).
3. If absent: Produce a specific redline requiring it:

   "Add to §[X]: Provider shall ensure that any third-party model providers, infrastructure providers, or subprocessors used in delivering the Services are bound by obligations with respect to [Customer Data / AI training / data retention / confidentiality] no less protective than those set forth in this Agreement, and shall be responsible for any breach of this Agreement caused by such third parties."

4. Flag the gap with a severity: 🔴 if the term is training-on-data or liability and there's no flow-down; 🟡 if the term is less sensitive or there's partial flow-down.

• It doesn't review the DPA provisions of the same agreement — run

  /privacy-legal:dpa-review

  , if the plugin is installed, for that.
• It doesn't decide whether to accept terms outside the fallbacks. It routes those per the escalation table in

  ~/.claude/plugins/config/claude-for-legal/ai-governance-legal/CLAUDE.md

  .
• It doesn't evaluate vendor security posture beyond what's in the agreement — that's a security team function.