Delivering AI work to clients: the 4-clause contract addendum every solo agency needs in 2026

This piece lays out a legal brief framework, not finished contract language. Nothing in this article constitutes legal advice. Have your lawyer mark up every clause skeleton before you put it in front of a client.

You have a client deliverable due next month. You used Claude to draft the copy, Midjourney to generate the hero image, and GitHub Copilot to scaffold the front-end. The client knows you use AI in a general sense; they have never signed anything that addresses it specifically. That gap is manageable today. From 2 Aug 2026, when EU AI Act Article 50 transparency provisions take effect, it starts carrying regulatory weight. And the contract-law exposure on copyright warranties has been building for two years without a clear fix.

The four clauses in this piece are a structured brief: one for each liability vector the gap creates. You take the brief to your lawyer; your lawyer marks it up for your jurisdiction and your client profile. The article tells you what each clause needs to accomplish and why. Your lawyer decides the exact wording, the placement, and whether your existing contract structure needs a standalone addendum or surgical edits.

The reader assumption throughout: you are a solo or 2-10-person agency delivering design, development, copy, or marketing work. Your clients are businesses, not consumers. Your rate band is roughly €50-€500/hr. You operate in the EU or UK, or you serve clients who do.

Why four clauses, not one omnibus paragraph

The temptation with AI contract compliance is to add a paragraph to the MSA that says something like “Agency may use AI tools to assist with deliverables; client acknowledges and accepts.” That paragraph is not useless. It is, however, answering only one of four distinct legal questions.

Question 1 (Disclosure): Has the client been informed that AI was used, as Article 50 requires?

Question 2 (IP warranty): Does the agency’s standard “work is original” warranty still hold when AI-generated content has uncertain copyright status?

Question 3 (Training data): Is there a contractual commitment to prevent client confidential materials from being ingested by AI model training?

Question 4 (Liability allocation): When an AI-deliverable dispute arises, which party bears what share of the risk, up to what ceiling?

An omnibus paragraph typically addresses Question 1 partially and leaves Questions 2, 3, and 4 to the existing contract boilerplate, which was not drafted with AI deliverables in mind. Four targeted clauses address each question cleanly. That specificity is also easier for a client’s legal team to review and sign off on: a single vague AI paragraph invites negotiation on everything; four precise clauses let the negotiation focus on the points that actually matter for each party.

Clause 1: Disclosure of AI use

What it must accomplish

The clause needs to tell the client, per project: which AI tools were used, which deliverables contain AI-generated content, and how they can request a human-authored alternative if they object. Article 50 of the EU AI Act (EUR-Lex official text, Regulation 2024/1689, Art. 50) applies to deployers of general-purpose AI systems used to produce content for natural persons. A solo agency using Claude or ChatGPT to write copy, Midjourney to generate images, or Copilot to generate code is a deployer within the definition. The transparency obligation requires disclosure in a “clear and distinguishable manner.”

The Article 50 transparency provisions became mandatory on 2 Aug 2025 for providers of general-purpose AI models, and the provisions covering deployers who produce AI-generated content for natural persons take effect on 2 Aug 2026. Your lawyer will confirm your specific position in the supply chain.

Why the SOW, not the MSA

Put the disclosure in each statement of work, not the master services agreement. The reason is specificity: the tools you use and the proportion of AI-generated content vary by project. A disclosure clause in the MSA says “agency may use AI” in the abstract; a disclosure in the SOW says “this engagement will use [Claude API, Midjourney v6] to produce [first-draft copy, hero imagery].” The latter satisfies Article 50’s requirement for clear and distinguishable disclosure; the former does not obviously do so.

Skeleton language — your lawyer’s starting point

The language below is a brief skeleton. Every marker in brackets is a placeholder; your lawyer will fill these in based on your jurisdiction and your typical client profile. Do not send this to a client without legal review.

AI Tool Disclosure (to be completed per engagement)

In connection with the deliverables described in this SOW, Agency will use the following AI tools: [LIST TOOLS BY NAME OR CATEGORY, e.g., large-language-model text generation, AI image generation]. The following deliverables contain content generated or materially assisted by AI: [LIST DELIVERABLES]. Agency has reviewed all AI-generated content editorially before delivery. If Client objects to the use of AI in any deliverable, Client may request a human-authored alternative within [X] days of delivery [LAWYER TO CONFIRM: remedy, timeframe, and any associated fee adjustment]. This disclosure is made in accordance with applicable transparency obligations under EU AI Act Article 50 [LAWYER TO CONFIRM: jurisdictional scope and any additional national implementing measures].

Clause 2: IP warranty carve-out for AI-generated portions

Why the standard warranty breaks

Standard agency MSAs contain a warranty that reads, in effect: all deliverables are original works; they do not infringe third-party intellectual property rights; and title passes to the client on payment. That warranty is drafted for human-authored work. It breaks in two specific ways when applied to AI-generated content.

First, the originality question. The US Copyright Office’s March 2023 guidance on AI-generated works established that copyright protection requires human authorship, and AI-generated content without sufficient human creative control does not qualify. The US Copyright Office has continued refining this position through subsequent registration reviews and public guidance updates into 2026. The UK Intellectual Property Office’s consultation on AI and copyright has been running since 2021, with active policy development ongoing; the UK currently extends copyright to computer-generated works under Section 9(3) of the Copyright, Designs and Patents Act 1988, but the application of that provision to modern generative AI is unsettled. The WIPO Standing Committee on Copyright and Related Rights has been tracking global divergence on this question since 2020. There is no international consensus.

Your standard warranty warrants more than you can deliver on AI-generated portions. The carve-out narrows the warranty to what you can actually warrant.

Second, the infringement question. AI models are trained on large corpora of existing works. The legal question of whether AI-generated output infringes the training-data copyright is actively litigated in multiple jurisdictions. You cannot warrant, with confidence, that an AI-generated image or a text passage does not reproduce protected expression from the training data. You can warrant that you have taken commercially reasonable steps to check for known third-party matches.

Skeleton language

IP Warranty — AI-Generated Content

The warranty in [MSA Clause X] that deliverables are original works is modified as follows with respect to AI-generated portions identified in the applicable SOW disclosure: (a) Agency does not warrant that AI-generated portions qualify as original works capable of copyright protection in any jurisdiction; (b) Agency warrants that it has reviewed AI-generated portions using commercially reasonable means to identify obvious reproduction of known third-party works; (c) Agency warrants that it has exercised editorial judgment over AI-generated portions before delivery; (d) copyright ownership and protectability of AI-generated portions is subject to applicable law in the relevant jurisdiction, which continues to develop. Client acknowledges that copyright status of AI-generated content varies by jurisdiction and that Agency has disclosed this uncertainty. [LAWYER TO CONFIRM: whether any additional indemnity or escrow mechanism is appropriate for the client’s industry or risk profile.]

Bird & Bird’s AI briefings (twobirds.com/en/trending-topics/generative-ai) and Pinsent Masons’ coverage at pinsentmasons.com cover the EU and UK IP law dimension in practical terms for commercial contracts and are worth sharing with your lawyer as background.

Clause 3: Training-data exclusion of client materials

What this protects

When you paste a client’s unpublished brand guidelines, a confidential brief, or a draft business plan into a chat interface, the question is: does that input end up in future model training? The answer depends on the configuration you are using.

Anthropic’s Claude business and API tiers exclude training on user inputs by default. OpenAI’s ChatGPT Business and API configurations do the same. This is a vendor-side contractual commitment, not a technical guarantee, and it is an auditable, named commitment that a client’s legal team can reference. The consumer-tier seats (Claude.ai free and Pro, ChatGPT Plus) have different default terms; check the current terms before relying on them for client-confidential inputs.

The training-data exclusion clause binds you, the agency, to use only configurations that carry those contractual protections. It cannot bind Anthropic or OpenAI directly in your client contract. What it does is give the client an enforceable undertaking from you: if training-data protection fails because you were running a consumer-tier seat or a model that trains on inputs by default, the contractual liability is yours.

The Society for Computers and Law has published practical guidance on drafting data-handling obligations for AI tools in commercial contracts. Stibbe’s AI practice (stibbe.com/tags/artificial-intelligence-ai) covers the GDPR intersection with AI training data in an EU commercial context and is a useful reference for your lawyer when the client’s materials include personal data.

Skeleton language

Training-Data Exclusion

Agency warrants that it will not submit Client Confidential Materials (as defined in [MSA Clause X]) to any AI system that trains on user inputs as a default or that lacks a contractual exclusion of training on user inputs. Agency will use only the following configurations, which as of the date of this agreement contractually exclude training on user inputs: [LIST SPECIFIC VENDOR + TIER, e.g., Anthropic Claude API with zero-retention; OpenAI ChatGPT Business; LAWYER TO REVIEW AND CONFIRM CURRENT TERMS]. If Agency changes its AI vendor or configuration in a way that materially affects training-data handling, Agency will notify Client within [X] days and obtain Client’s written consent before submitting Client Confidential Materials to the new configuration. [LAWYER TO CONFIRM: definition of ‘Client Confidential Materials’ cross-reference, notification period, and consent mechanism.]

Clause 4: Liability cap tied to fee paid

The practitioner default and why it exists

For a solo or small agency, the liability cap on AI-deliverable disputes follows the same logic as the general MSA liability cap: the ceiling should be proportionate to the value of the engagement, not to the theoretical maximum harm that could flow from a copyright dispute or a training-data incident.

The practitioner default in the EU and UK commercial agency market is fees paid in the prior 6 months. The 6-month look-back period is widely used because it is proportionate (it covers the value at risk in a typical project), it is not so low that courts set it aside as unconscionable in a commercial context, and it is not so high that it functions as unlimited liability for a solo agency.

EU AI Act Article 50 creates a transparency obligation on you as a deployer. It does not create a strict-liability private right of action for the client. The regulatory enforcement mechanism is through national competent authorities, not through your client contract. This matters when setting the cap: the client’s legal team may try to argue that Article 50 liability should sit outside the contractual cap because it arises from regulation, not from the contract. Your lawyer’s position is that Article 50’s transparency obligation is satisfied by your Clause 1 disclosure, and that any residual dispute between you and the client about AI use is contractual, not regulatory, and therefore capped.

Skeleton language

Liability Cap — AI-Deliverable Disputes

Notwithstanding any other provision of this Agreement, Agency’s total liability arising from or relating to AI-generated content in any deliverable, including but not limited to liability arising under the IP warranty carve-out in [Clause X], the training-data exclusion in [Clause X], or any failure of disclosure under [Clause X], shall not exceed the total fees paid by Client to Agency in the six (6) months immediately preceding the event giving rise to the claim. This cap applies regardless of the form of the claim (contract, tort, statute, or otherwise) and regardless of whether Agency has been advised of the possibility of such losses. [LAWYER TO CONFIRM: interaction with existing general liability cap; whether any carve-outs (gross negligence, wilful misconduct) apply; jurisdictional enforceability in client’s jurisdiction.]

Where these four clauses live in your contract documents

Getting the placement right matters as much as the content. The four clauses are not all the same type of provision.

SOW placement: Clause 1 (Disclosure of AI use). Every SOW that involves AI-generated content needs a completed disclosure. Blank disclosure clauses in SOW templates are fine; your lawyer can build a standard completion checklist.

MSA or standing addendum placement: Clauses 2, 3, and 4 (IP warranty carve-out, training-data exclusion, liability cap confirmation). These govern how you work generally. If you already have a well-drafted MSA, the cleanest approach is a standalone AI Addendum that modifies the relevant MSA clauses by reference, rather than re-drafting the MSA wholesale. A standing addendum lets you update AI-specific terms as the law develops without renegotiating the entire MSA with every client.

Schedule placement (alternative): For clients who will not accept MSA changes, a schedule signed at the same time as the SOW can contain all four clauses. This is a weaker structural position (it is easier for a client to argue a schedule clause does not override an MSA clause), but it is often the pragmatic solution for established client relationships. Your lawyer will advise on the precedence language needed to make a schedule override work.

What we are not recommending

Three configurations look like solutions but leave you exposed.

One omnibus ‘AI clause’ buried in fine print. A single paragraph at the back of a schedule does not satisfy Article 50’s “clear and distinguishable” disclosure requirement. It may not address the IP warranty or the training-data exclusion at all. Courts and regulators look at whether disclosure was prominent and specific, not whether it existed somewhere in the document stack.

Relying on the existing IP warranty. Your current warranty warrants more than you can deliver on AI-generated work. The gap is not theoretical: it is actively litigated in the UK, EU, and US. Leaving the standard warranty in place and hoping copyright law settles in your favour is an unallocated risk.

Using the AI vendor’s terms as your client-facing disclosure. Anthropic’s Terms of Service and OpenAI’s Terms of Use govern the vendor-to-agency relationship. They do not govern the agency-to-client relationship. Your client cannot rely on Anthropic’s terms to understand what you disclosed, what you warranted, or what cap applies to a dispute between you and them.

What changes this verdict

Cadence on this claim is 30-day through Aug 2026, then 45-day. The Article 50 secondary acts from the European Commission are expected to publish practical implementation guidance through 2026; any shift in what “clear and distinguishable” disclosure requires could revise Clause 1 above. After Aug 2026 the immediate deadline pressure is off, and cadence extends.

Four specific triggers move this verdict to Partial:

• European Commission publishes Article 50 implementation guidance that is more specific about what deployer disclosure must contain. This would update the Clause 1 skeleton.
• UK IPO consultation concludes with a position on AI and copyright. A definitive UK position, whether extending or clarifying Section 9(3) CDPA for generative AI, would update the Clause 2 skeleton for UK-serving agencies.
• US Copyright Office issues updated guidance that either grants or narrows copyright protection for AI-generated outputs with material human editorial input. The US guidance matters for agencies with US clients or US-incorporated entities.
• A named EU or UK court ruling on AI-deliverable copyright establishes case law on whether AI-generated work can infringe training-data copyright. The first such ruling would sharpen the Clause 2 carve-out materially.

Status: Holding as of 12 May 2026. Next review: 11 Jun 2026.

Closing: get this marked up before August

The four clause skeletons in this piece are a brief, not a final document. The right action is to share this article and the clause skeletons with a lawyer who handles commercial agency contracts and who is current on EU AI Act implementation. The brief should take a competent lawyer two to four hours to mark up for your existing contract structure.

Named law firms publishing practical AI contract guidance for commercial agencies and SMBs in 2026: Bird & Bird, Pinsent Masons, and Stibbe. These mentions are included because their public briefings are substantive starting points, not as endorsements of their services.

The 2 Aug 2026 Article 50 deadline is the external pressure. The internal pressure is the IP warranty: every AI-assisted deliverable you send without a carve-out is a deliverable where your standard warranty overpromises. That gap exists now, not in August.

Related operator reading

The GDPR and data-handling intersection when client materials contain personal data is at /operators/gdpr-ai-tools-small-business/. The one-page AI policy that covers the internal side of the same controls (how you govern AI use inside the agency before it reaches the client) is at /operators/1-page-ai-policy-for-small-business/. The AI invoicing and VAT compliance question for solo agencies selling AI-augmented services is at /operators/ai-invoicing-vat-compliance-small-business/.

Related enterprise reading

The enterprise procurement side of AI-deliverable IP and warranty questions, including the Article 50 scope for large-scale deployers, is at /eu-ai-act-article-50-transparency-disclosure/. The vendor-comparison framework for enterprise AI procurement with contract terms included is at /enterprise-ai-agent-vendor-comparison/.