The governance wedge · Article 4 of 7
The AI clause you already signed
Long before your firm wrote an AI policy, it signed a stack of contracts with opinions about AI. The audit is reading your own paper, and it takes an afternoon.
1 · Your obligations did not wait for your policy
Firms taking AI governance seriously usually read the vendors’ terms first. That is reasonable, and the previous article walks through exactly that reading. But the sharper edges sit closer to home, in the contracts your own firm has already signed. Engagement letters. Non-disclosure agreements. Master service agreements, publishing and production agreements, some of them years old. Those documents were written before your AI question existed, and several of their clauses answer it anyway, often more strictly than anyone in the building remembers agreeing to.
This is the domain where “we didn’t think of it as AI” buys nothing. A confidentiality clause does not care that the disclosure happened through a chat window. An intellectual property assignment does not soften because a model helped with the first draft. The work here is easy to describe and rarely done: put your standard paper beside your actual AI usage and find the places they disagree.
Automation that would break a client contract isn’t efficiency; it’s liability. That sentence has been on our site since launch. This is the article where it stops being abstract.
2 · Four clause families with opinions about AI
Nearly everything relevant hides in four familiar places. The language below is typical rather than quoted from any one agreement, and yours will differ, which is rather the point.
| Clause family | What it typically says | What it means for AI use |
|---|---|---|
| Confidentiality and non-disclosure | Client information goes to no third party without consent, sometimes with narrow carve-outs for named service providers | Pasting client material into an outside AI service is a disclosure to a third party. Whether it is a permitted one turns on words written before anyone imagined the question |
| Subprocessors and third-party services | Vendors handling client data must be disclosed, approved, or bound to equivalent terms | An AI vendor processing client content may belong on that list. If your agreement promises a subprocessor register, the register has to be true |
| Work product and IP assignment | The firm assigns all right, title, and interest in deliverables to the client | You can only assign what exists. Purely machine-generated material may carry no copyright to assign, and a warranty of originality sits uneasily on unreviewed model output |
| Warranties and professional standards | Services performed with professional skill and care, by qualified personnel | The ABA’s 2024 ethics opinion on AI already answered this for lawyers. Accountability for the work stays with the person who signed it, however it was produced, and “the tool did it” is not a standard of care |
None of this says AI is forbidden. In most portfolios we have reviewed, nothing forbids it outright. What the paper does is set conditions, and conditions have to be known before they can be met.
The uncomfortable cases are always specific ones. The legacy NDA with an absolute no-third-parties clause and no service-provider carve-out. The government or healthcare engagement with data residency terms no consumer tool satisfies. The co-publishing agreement whose delivery warranty was drafted in 2019 by somebody who has since retired. Every firm has a few. The audit exists to find them before a dispute does.
3 · Who owns AI-assisted work
You can only assign what exists.
The ownership question deserves its own section, because both halves of it are moving.
Under current U.S. Copyright Office registration guidance, issued in March 2023 and still the operative framework, copyright protects “only material that is the product of human creativity.” Purely machine-generated output, produced from a prompt where the machine determined the expressive choices, is not registrable. AI-assisted work is a different animal. Where a person selects, arranges, or modifies AI material with enough creativity, the human contribution is protectable, and the Office expects applicants to disclose more-than-minimal AI content in registrations and to describe the human authorship.
Translate that into a working rule for a firm that assigns deliverables to clients. The more a deliverable is actually authored by your people, with AI as an instrument rather than the author, the more solid the thing you are handing over. The closer it drifts toward raw model output, the murkier the property becomes, and the less comfortable your originality and non-infringement warranties should feel. Which happens to align with what quality already demanded: qualified humans doing real work with good tools, the arrangement your engagement letter promised in the first place.
Publishers carry a second layer. If authors and freelancers deliver material to you, the same questions run in the other direction, and your contributor terms may need an AI representation you currently do not collect. What did they warrant, and would you know if it stopped being true?
4 · The audit is a reading exercise
Here is the version of this work that actually gets done, because it is scoped like an afternoon rather than a career.
- Pull the paper that matters. Your top clients by revenue and sensitivity, your standard engagement template, your NDA template, your contributor and vendor agreements. Ten to twenty documents covers most firms’ real exposure.
- Read for the four families. Note the exact language rather than a paraphrase. The difference between “no third parties” and “no third parties except service providers under equivalent confidentiality” is the whole game.
- Build the register. One row per contract: what it restricts, what that implies for AI use on that client’s work, and a flag wherever the answer is unclear. Unclear is a finding rather than a failure.
- Feed it back into the policy. This is where the tiers from earlier in the series stop being theory. Clients whose contracts restrict third-party processing move into your most restrictive tier by name, and the policy stops saying “sensitive work” in the abstract and starts saying which work.
The register produces something quietly valuable besides. When a client asks whether their contract even allows this, you answer from a document, on the day, which is rare enough that clients notice.
The last step is prospective. Your next engagement letter is a chance to replace archaeology with intention: a sentence or two in plain language, stating whether and how AI-assisted tooling may be used, under what safeguards, reviewed by qualified professionals, with confidentiality preserved through business-grade services under written data terms. Clients rarely balk at honesty offered in advance. What erodes trust is discovery after the fact. And when a client’s own paper arrives carrying a blanket no-AI clause, treat it as an opening position rather than a landmine to sign around.
Where this fits
Contracts are the fourth of eight domains in our AI Governance Readiness Checklist, and the first one where the answers are entirely your own. No vendor setting fixes a promise you made in writing.
☐ We’ve reviewed client contracts and NDAs for terms that restrict AI or subprocessor use.
☐ Our AI use doesn’t conflict with the confidentiality and IP clauses we’ve signed.
☐ We know who owns the intellectual property in the AI-assisted work we deliver.
☐ We’d be comfortable putting our AI use on an account in writing.
The next article turns to a newer stack of paper in the same neighborhood, the AI exclusions now appearing in commercial insurance policies, which is the other document set deciding your exposure while nobody reads it.
And the plain disclaimer, because we mean it. We are consultants rather than your lawyers. This article tells you where to look and what the patterns mean operationally. What your specific clauses require is a question for counsel, ideally asked with the register already in hand, which makes that conversation shorter and cheaper.