Retrofitting an old MSA for AI production
Most agreements in force predate generative production and are quietly hostile to it. The six sections to amend, and how to raise it without alarm.
The master services agreement you signed in 2022 does not forbid AI production. That is the problem. It says nothing about it, which means every question it raises gets answered by clauses that were drafted for a different way of making things, and some of those clauses now say something you did not intend to promise.
Three sections do most of the damage, and none of them mention AI.
The clauses that turned hostile without being edited
The originality warranty. Almost every production MSA contains a line where you warrant that deliverables are original works, created by you, not infringing, and assignable in full. Read literally against generative output, that is a promise you may not be able to keep. The US Copyright Office position is that protection attaches to human contribution: substantial modification of generated material, human-authored elements, and the creative selection and arrangement of AI output. Raw output with no meaningful human contribution may carry no copyright at all, and you cannot assign a right that does not exist. Practitioner writing on this is blunt about it, including Neolemon's treatment of client contracts for AI-generated illustration, which frames the assignment as a promise you may already have broken.
Model providers assign contractual ownership of outputs to the user in their terms, and that is real and useful. It is not the same thing. A contract between you and a vendor cannot manufacture copyright that statute declines to grant, and an assignment clause that assumes it can is the single most common defect in a retrofitted agreement.
Confidentiality. Standard confidentiality language restricts disclosure of client materials to third parties. Uploading a client's product photography or unreleased brief into a hosted model is, on a plain reading, disclosure to a third party. Nobody drafting that clause in 2022 was thinking about inference endpoints, but the words do not care what anyone was thinking about.
Subcontractor consent. Many MSAs require written approval before engaging subcontractors. Whether a hosted model is a subcontractor is a genuinely open question, and you do not want it answered for the first time during a dispute.
The six sections to amend
Practitioner guidance converges on a small, consistent clause set. Agent Mode AI's four-clause minimum covers disclosure, an IP warranty carve-out, a training-data exclusion, and a liability cap tied to fees paid. Numonic's fuller version adds metadata preservation and audit rights. Mapped onto an existing MSA, that is six sections.
| Section | What it says now | What it needs |
|---|---|---|
| IP and assignment | Deliverables are original works, assigned in full | A carve-out for AI-generated portions and an honest statement of what is being assigned |
| Confidentiality | No disclosure to third parties | An express permission for processing through named tool categories, with a training exclusion |
| Subcontractors | Written consent before engaging third parties | A definition that separates tools from subcontracted humans |
| Disclosure | Silent | Who labels, on which surfaces, and who owns the consequence of not labelling |
| Indemnity and liability | Broad indemnity, uncapped or oddly capped | Two-sided allocation and a cap tied to fees paid under the relevant statement of work |
| Records and audit | Silent, or a general books-and-records right | A scoped production record and a bounded inspection right |
Six amendments is a lot to hand a client's legal team at once. It is also the honest number, and bundling them into one amendment is easier to get through review than six separate requests over six months.
What each amendment says
IP and assignment. Replace the flat originality warranty with a two-part statement: you assign all rights you hold in the deliverables, and you do not warrant that AI-generated portions are protectable by copyright, since that status varies by jurisdiction and turns on the degree of human authorship. Pair it with a positive commitment that the human contribution is real, because that is what actually determines the answer. This is the amendment clients push back on hardest and the one that protects them most, which is worth saying out loud in the covering note.
Confidentiality. Add an express permission to process client materials through the tool categories you use, with a hard training exclusion: no client material may be used to train, fine-tune, or improve any third-party model. Name it as an exclusion rather than a preference, because it is the specific fear behind most client hesitation. Whether a vendor honours that is a diligence question, and licensed training data as a buying criterion is the version of that question aimed at the models themselves.
Subcontractors. Define a subcontractor as a person or firm performing services, and state that software tools, including hosted models, are not subcontractors. One sentence, and it closes an argument nobody wants to have.
Disclosure. Assign it explicitly. In practice the labelling obligation almost always attaches to the publishing account, which is the client's, and clients frequently do not realise this until a platform surfaces a label they did not apply. What to tell clients about using AI separates the commercial disclosure from the compliance one, and the amendment needs to cover the second.
Indemnity and liability. Split it in both directions and cap it at fees paid under the relevant statement of work. An uncapped indemnity on a five-figure engagement is not a negotiating position, it is an unpriced liability.
Records and audit. Commit to keeping a production record, define what is in it, and bound the inspection right. Left undefined, an audit right expands to whatever the auditor asks for.
Raising it without setting off a fire alarm
The framing mistake is leading with "we've started using AI." That sentence invites a procurement review you did not need to trigger. It also invites a price conversation, and the data suggests that conversation is less common than the fear of it: reporting aggregated by TechBullion puts 73% of agencies as never having been asked to cut prices despite adopting AI, and among the 27% who did face discount requests, only 13% actually lowered their rates. The fear of that conversation is doing more work than the conversation itself. It still comes up eventually, and you want to be the one who raised it, in a document, on your timing.
So raise it as housekeeping, at renewal, bundled with other administrative updates.
"Ahead of the renewal I've had our agreement reviewed and there are a handful of sections that predate how production actually works now. Nothing about scope or rates changes. The updates cover how we handle your material, who owns what, who applies platform labels, and what records we keep. I've marked up the existing document rather than replacing it, so it should be a short read. Happy to walk your team through it, or to send the plain-English summary first if that's easier."
Three things that note is doing. It states that commercial terms are unchanged in the second sentence, which is the only sentence a busy client reads carefully. It presents the changes as protective of them, which four of the six genuinely are. And it offers the summary-first path, because the reason these amendments stall is not disagreement, it is that nobody has time to read a redline.
If your rates are also moving, move them separately. Bundling a rate change with a legal amendment makes both harder, and raising prices on existing content clients is a different conversation with a different rhythm.
If they say no
Some clients will decline part of it, usually the IP carve-out, because their legal team reads any warranty softening as risk transfer. Two workable fallbacks.
Narrow the carve-out to a defined list of deliverable types rather than the whole agreement. A client will often accept "the following asset classes are produced with generative tools and carry this warranty treatment" when they will not accept a blanket amendment.
Or offer a human-authorship commitment instead of a warranty softening: you warrant meaningful human authorship in every deliverable, which is the thing that actually determines protectability, and you accept the originality warranty on that basis. It is a genuine commitment and it costs you nothing if your process is real.
If a client refuses everything and their existing contract forbids the tools outright, that is a separate decision with its own path. Governance-side clients often just need to see the controls written down, which is what AI content governance for brands covers from their side of the table.
FAQ
Do I need an amendment, or is an addendum enough?
An addendum that expressly supersedes the conflicting sections is usually cleaner than a redline, because it avoids reopening the whole agreement. Say explicitly which numbered clauses it modifies. An addendum that is silent about precedence just creates two documents that disagree.
What about work already delivered under the old terms?
Leave it. Retroactively amending a completed statement of work draws attention to the exact question you are trying to close and offers no benefit. Start the new terms at the next statement of work or the renewal date, whichever comes first.
Does this apply to a one-page freelance contract too?
Yes, and it is easier there because there is no legal team. The six sections compress to about a paragraph each. If you work through agencies rather than direct, check the pass-through terms as well, since white-label arrangements often bind you to an end-client agreement you have never seen.
Should I get a lawyer to draft this?
For anything above a modest annual value, yes. The sections above are what to ask for, not a substitute for advice in your jurisdiction, and the IP treatment in particular differs enough between markets that a template written for one is actively misleading in another.