AI Clauses for Agency Contracts: Disclosure, Ownership, Approvals

Add clauses on three things: disclosure (which AI tools you use and for what, and whether the client can opt out), ownership (what you assign, the limits on rights in AI-generated material, and your tools' terms), and approvals (client sign-off before AI-generated images, voices or claims go live). Then cover confidentiality, warranties, indemnity limits and records.

The ownership clause is where most agency contracts are now out of date. A standard "we assign all intellectual property in the deliverables" promise assumes there is copyright to assign. Material generated by AI with little human creative input may have none, and the AI vendor's terms pass on only whatever rights it holds, "if any". Promise what you can deliver.

Follow me on Instagram@sagnikteaches

A note before the wording: this is practical drafting to decide what you want and to brief a lawyer, not legal advice. Contract law and copyright differ between jurisdictions, and the clauses that allocate risk (ownership, warranties, indemnities) should be reviewed by a lawyer who knows the law your contracts are under. For the general principle of mentioning AI in proposals and contracts, start with how to mention AI use in client contracts.

Connect on LinkedInSagnik Bhattacharya

Disclosure: choose one of three levels, then say it plainly

Agencies argue about disclosure because they treat it as all or nothing. It's easier as a choice between three levels, matched to the client and the work:

Subscribe on YouTube@codingliquids

Sample wording for level 2, which suits most agencies:

AI tools. The Agency uses artificial intelligence tools in research, drafting,
design and production, under the direction and review of its staff. The Agency
will identify in the delivery note any Deliverable in which an AI tool generated
a substantial visible or audible element (for example an image, a synthetic voice
or a video sequence). The Client may, by written notice, require that AI tools
are not used for a specified Deliverable or category of work; the Agency will
confirm any effect on timing or fees before proceeding.

The last sentence matters. An opt-out with no effect on fees invites every client to tick it; an opt-out that triggers a conversation about time and cost is fair to both sides.

Level 3 earns its admin cost where an AI-made element could mislead in a way that matters. Take an illustrative dental practice client: an AI-retouched smile on a treatment page is, in effect, a claim about results, so every generated or AI-edited image goes to the practice for written approval before use. For a retained social account posting seasonal opening hours, the same gate would be pure friction. That's why the level is set per client, and sometimes per deliverable, rather than once for the whole agency.

Ownership: what you can actually hand over

Three facts shape this clause:

• AI vendors assign what they have, not more. OpenAI's terms assign to the user "all our right, title, and interest, if any" in outputs, and Anthropic's commercial terms similarly assign its rights (if any) to the customer. That removes the vendor as a claimant. It doesn't create copyright where none exists.
• Human authorship still decides protection in many places. One major national copyright office concluded in 2025 that prompts alone don't give enough control to make the prompter the author, and that wholly AI-generated material isn't protected, while human selection, arrangement and creative modification of AI output can be. Other jurisdictions have their own rules for computer-generated works. Either way, the more human creative work in a deliverable, the stronger the client's position.
• Tool terms travel with the asset. Some image tools set conditions on ownership (Midjourney, for example, requires companies above $1 million in annual revenue to be on its Pro or Mega plan to own what they create), and stock or template licences may limit use. Check the terms of every tool that touches a deliverable.

Here is the core clause, rewritten:

Before: "The Agency assigns to the Client all intellectual property rights in
         the Deliverables and warrants that the Deliverables are original."

After:  "On payment, the Agency assigns to the Client all intellectual property
         rights it owns in the Deliverables. Where a Deliverable includes
         material generated by an AI tool, the Agency assigns whatever rights it
         holds in that material and in its selection, arrangement and
         modification, and the Client acknowledges that such material may not
         attract copyright protection in every jurisdiction. The Agency will not
         knowingly supply material that infringes a third party's rights."

The "after" version still gives the client everything you have. It swaps an absolute promise of originality you can't fully check for a knowledge-based one you can stand behind. For brand identity work, logos especially, go further: agree that the final artwork is drawn by a person, with AI used only for exploration. The designer's side of that process is in AI for freelance designers without rights issues.

Approvals: the sign-off gates before anything AI-made goes live

An approval clause turns "the client signed off the campaign" into specific checks on the parts where AI creates new risk. Four gates cover most agency work:

1. Real people. No AI-generated likeness or synthetic voice of a real, identifiable person without that person's written consent and the client's approval.
2. Products. Any AI-generated or AI-edited image of the client's product must match the real product in shape, colour, size and features, and the client confirms it does. A lamp rendered with a brass finish it doesn't come in is a misleading advert.
3. Claims and facts. Statistics, comparisons, testimonials and product claims in AI-drafted copy are verified against the client's evidence before approval.
4. Labelling. Where a law or platform requires AI-generated content to be labelled, the parties agree who applies the label. If your client sells to customers in the EU, the EU AI Act's Article 50 transparency duties, including labelling deepfakes, have applied since 2 August 2026. The practical side is in how to label AI-generated images and video.

Sample wording:

Approvals. Before publication, the Client will approve in writing (email is
sufficient) any Deliverable that (a) depicts or imitates the voice of a real
person, (b) shows the Client's products in images generated or substantially
altered by AI, or (c) contains factual or comparative claims. The Client is
responsible for the accuracy of product information and claims it approves.
The parties will agree in the brief who is responsible for any labelling of
AI-generated content required by law or by the publishing platform.

The supporting clauses agencies forget

Disclosure, ownership and approvals get the attention. These items are where disputes start. Treat the list as a checklist against your current terms:

• Client material in AI tools. Say that you'll only put client confidential material into tools on your approved list, on business plans that don't train on it by default. Check: your tool list exists and the plans are business tiers.
• Client-supplied AI content. If the client sends you AI-generated images or copy to use, their warranties apply to it, not yours. Check: the warranty clause covers "Client Materials" including AI outputs.
• Warranty scope. Replace "original" with "not knowingly infringing", as in the ownership example. Check: no absolute originality promise remains anywhere, including in proposals.
• Indemnity limits. Cap your indemnity and exclude claims arising from client-directed prompts, client materials or client changes. Some vendors offer their own protection on paid business products: Microsoft's Customer Copyright Commitment covers commercial Copilot customers who use the built-in guardrails, and Adobe sells an IP indemnity entitlement for Firefly outputs to enterprise customers. Those protect you as the vendor's customer; they don't extend to your client unless you say so. Check: the cap is a figure, not "unlimited".
• Records. Keep, for each deliverable with AI-generated elements, the tools and plans used, the dates and the source files, for an agreed period (two years is common). Check: your project folders have somewhere to put this.
• Prompts and working files. Say whether prompt libraries, brand-voice instructions and style references you build for a client are theirs on termination. Clients increasingly ask for them.
• Fees. If you price by output rather than by hour, say so, so a client can't argue that faster AI-assisted work should cost less after the fact.

Vendor protection has edges worth checking tool by tool. Adobe's own Firefly models are trained on licensed content, but the Firefly app also offers partner models that those assurances don't cover, and the IP indemnity is for enterprise customers only. An illustrative slip: a designer generates a hero background with a partner model inside the Firefly app, the account manager tells the client "it's Adobe, so it's indemnified", and neither half of that survives a look at the terms. Record the model, not just the app, for each asset, and mention vendor protection to a client only once you've confirmed it covers the plan and model you actually used.

The working-files question tends to arrive at the worst moment, when a client leaves. Picture a two-year retainer ending and the client asking for "the AI set-up you built for us": the brand-voice instructions, a library of 80 tested prompts and the style references that keep images on brand. If the contract is silent, both sides have a reasonable argument, and the handover stalls while the relationship sours. One way to settle it in advance:

Working materials. On termination, and once all fees are paid, the
Agency will supply the brand-voice instructions and style references
created specifically for the Client. Prompt libraries and methods the
Agency uses across its clients remain the Agency's property.

That split hands the client what is specific to their brand and keeps your general know-how yours. Some agencies go the other way and sell the prompt library as an add-on at handover. Either is defensible; what isn't is deciding it on the day the client gives notice, when every answer sounds like a grievance.

How the clauses play out on a homeware brand's spring campaign

In this illustration, a content and social agency of eight people is running a spring campaign for an e-commerce homeware brand. The deliverables: 40 lifestyle images, 60 social captions, six product pages and one 30-second video with a voiceover.

How the clauses apply:

• Images. The agency photographs the real products, then uses an AI tool to generate the room settings around them. Under the level 2 disclosure clause, the delivery note flags all 40 as containing AI-generated backgrounds. The product gate applies: the brand's product manager confirms each product is shown accurately. Two images fail because the AI "improved" a vase's glaze, and are regenerated.
• Captions and product pages. AI drafts, copywriters rewrite. No flag needed under level 2, since no substantial AI element survives unedited. The claims gate catches one caption saying "handmade" about a machine-finished range.
• Video. The client wants a synthetic voice to save a studio session. The real-people gate doesn't apply because the voice isn't modelled on anyone, but the video is flagged in the delivery note and the brief records that the client will add a label where the platform requires one.
• Ownership. The photographs, copy and edit are human work and assign in the normal way. The AI-generated backgrounds are assigned "to the extent of the agency's rights", which the client accepted when signing because the agency explained it at the pitch.

The level 2 flag itself is one line per group of deliverables. The campaign's delivery note, illustrative:

DELIVERY NOTE: Spring campaign, batch 2
IMG-01 to IMG-40: product photographed; room background AI-generated
  ([tool, plan, model]); product accuracy approved by [client product
  manager] by email, 12 Mar
CAP-01 to CAP-60: AI-assisted drafts, rewritten by copywriter; no flag
VID-01: synthetic voiceover ([tool, plan]); client to label on platforms
  that require it (agreed in brief v3)

Written like this, the note doubles as the record the records clause asks for, so nobody keeps the same information twice.

The extra admin for the campaign was about two hours: the delivery-note flags, the product approval round and the records folder. That's what the clauses cost to run.

Checking a client's own contract for AI conflicts

Larger clients send their own master services agreement, and many now contain AI restrictions written by a procurement team. An approved business AI tool can do a first pass, as long as your plan doesn't train on the content. Try a prompt along these lines:

Below is a client's services agreement. Our agency uses AI tools to draft copy,
generate image backgrounds and edit video. List every clause that restricts,
requires disclosure of, or allocates risk for AI use, or that could be breached
by our use (confidentiality, originality warranties, IP assignment, data
handling, subcontracting). Quote the clause number and wording, then explain
the conflict in one sentence. Do not give legal advice; flag for review.

An illustrative extract of what comes back:

Clause 9.2 - "Supplier shall not use generative artificial intelligence in the
provision of the Services without Client's prior written consent."
Conflict: all current AI use needs written consent first.

Clause 12.1 - "Supplier warrants all Deliverables are its original work."
Conflict: AI-generated backgrounds may not be the agency's original work.

Clause 14.4 - "Supplier shall not disclose Confidential Information to any
third party." Conflict: may cover entering client material into AI tools,
depending on how "third party" is defined in clause 1.

Where a person still has to step in: the first two findings are right and are exactly what your lawyer needs to see. The third is a reasonable flag, but check clause 1 yourself: many agreements already permit disclosure to service providers under confidentiality terms. The AI also missed that the subcontracting clause (it wasn't asked to read schedules) requires approval of subcontractors, which might capture your freelancers. Always read the definitions and schedules yourself; a first pass shortens your lawyer's review, it doesn't replace it.

That subcontracting point is where agencies get caught in practice. An illustrative case: the agency agrees clause 9.2, gets the client's written consent for its own approved tools, then briefs a freelance copywriter without mentioning any of it. The copywriter drafts the product pages in a free chatbot on a personal account, pasting in unreleased product specifications. It comes to light only when the freelancer's invoice mentions the tool. The agency now has a possible confidentiality breach and has gone beyond the consent it was given. The fix is a flow-down line in every freelance brief for that client:

AI tools on this account: only [approved tools, business plans], through
accounts we provide or approve. Ask us before using any other AI tool,
including free ones. List the tools you used in your delivery note.

When the client pushes back

• "No AI at all on our account." Accept it for the account and price accordingly. Clarify the scope in writing: most clients mean generative content, not spellcheck, transcription or search. A realistic mistake here is agreeing to "no AI" and then using an image editor's generative fill without thinking of it as AI.
• "You must indemnify us for anything AI-related, unlimited." Offer an indemnity for knowing infringement with a cap, and explain that vendor protections exist on some tools you use. An unlimited indemnity for a risk nobody can fully assess isn't a reasonable ask of an eight-person business.
• "Disclose every use of AI on every asset." Offer level 2 and the records clause. Per-asset disclosure of every AI touch, including research and grammar checks, produces noise that helps nobody.

The fee conversation after a "no AI" request is easier with numbers. On the homeware campaign, say AI drafting brings each caption to about 10 minutes of copywriter time, against roughly 20 minutes written from scratch. For 60 captions that's 10 hours against 20. At an illustrative internal rate of $60 an hour, the no-AI captions cost $600 more to produce, and the images would cost far more, because the room settings would need a styled photo shoot. Put the difference in a change note so the client decides with the cost in front of them. That's what the opt-out sentence in the disclosure clause is for.

Whatever you agree, make your internal process match it. The terms only protect you if the delivery notes, approvals and records they describe actually exist, which is why the workflow in an AI content workflow from brief to approved draft is the other half of this job.

Questions agencies ask before updating their terms

Do existing clients need a contract variation?

If your current terms promise things AI use now makes awkward, such as assigning all copyright or warranting that every asset is wholly original, update them at renewal or by a short written variation both sides sign. Sending a notice that your terms have changed may not bind a client under a signed contract. Where you haven't used AI on a client's work yet, add the clauses before you start.

We subcontract to freelancers. Do these clauses need to flow down?

Yes. If your client contract restricts AI tools, requires disclosure or promises an assignment of rights, your freelancer agreements need matching terms, or you'll be promising what your suppliers haven't agreed to. Ask freelancers which AI tools they use, require them to follow your approved list for that client, and get the same records you promise the client.

Can a small agency use this wording without a lawyer?

Use it to decide what you want to say and to brief a lawyer efficiently, which keeps the bill down. Have a lawyer review the ownership, warranty and indemnity clauses at least, because those allocate risk and depend on the law that governs your contracts. The disclosure and approval wording is more operational and easier to adapt yourself.

Further reads

• How to Write Terms of Business With AI and What Needs a Lawyer — Which parts of your terms AI can draft, and which need a lawyer.
• How Agencies Keep Brand Voice Consistent in AI-Written Content — Keeping client voice consistent once AI drafts the copy.
• AI Rollout Plan for a Ten-Person Marketing Agency — Where these clauses fit in a wider agency AI rollout.
• How to Write an AI Content Policy for Your Marketing — The internal policy that sits behind the client-facing clauses.
• Should You Tell Customers When You Use AI in Your Marketing? — Your clients' own disclosure duty to their customers.
• How Small Agencies Use AI to Write New-Business Pitches — Using AI in pitches without promising what the contract won't.
• How Marketing Agencies Use AI to Automate Client Reporting — A four-layer setup for automated client reports: data pipes, the right reporting tool, a commentary prompt that can't invent wins, and a five-minute check.
• Which Agency Tasks AI Handles Well, and Which It Doesn't — Sort agency work by evidence, review effort and consequences, then test a reporting workflow with a clear human sign-off.
• Should Freelancers Tell Clients They Use AI? — Decide when to disclose AI use, ask for specific approval and explain your process without making promises you cannot support.
• How to Track Contract Renewals and Deadlines With AI — Use AI to read every agreement, a spreadsheet to calculate notice deadlines, and two reminders per contract so auto-renewals stop catching you out.
• Automate Client Onboarding With AI: Forms, Contracts, and Emails — One intake form, a template contract filled from it, and emails triggered by the signature, with AI writing the kick-off brief and chasing what's missing.
• Can AI Review a Contract? What Small Business Owners Should Know — What an AI first read of a contract catches and misses, when to pay a lawyer, a clause-by-clause prompt and a tour operator's hotel contract read end to end.
• AI Content Approval Workflow: Draft, Check, Sign Off — A three-stage approval workflow for AI-drafted content, with risk lanes, a filled-in check sheet, sign-off records and the numbers for a small agency.
• How to Check AI Content for Plagiarism Before You Publish — A pre-publish plagiarism routine for AI-drafted copy: checker prices, a match-triage table, cross-client duplicates and why AI detector scores prove little.
• AI Consulting Contracts: 9 Clauses to Check Before You Sign — Nine clauses for an AI consulting contract, from acceptance tests to exit terms, with illustrative wording and a farm shop's contract worked through.
• How to Choose a Marketing Agency That Uses AI Responsibly — Seven questions that separate careful agencies from careless ones, a test brief for the pitch, and how a wedding planner compared three agencies.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI and Anthropic terms (assignment of rights in outputs); Copyright and Artificial Intelligence Part 2 report on copyrightability; Microsoft Customer Copyright Commitment; Adobe Firefly enterprise legal FAQs; EU AI Act Article 50.