Most AI agency contracts fail in one of two ways. Either they're generic freelance contracts that say nothing about AI — leaving both parties without recourse when AI-generated output creates a dispute. Or they're overcomplicated, treating every clause as a liability shield, which signals distrust and makes clients reluctant to sign.
The contract that works for an AI agency is neither defensive nor vague. It's specific: it names what AI is used for, what the human review process is, who owns the output, and what happens when the output doesn't meet the agreed standard. This specificity is not just legal protection — it's a sales asset. Clients who see a thoughtful AI governance clause in a contract feel more comfortable, not less. It demonstrates that you've thought about what could go wrong.
The EU AI Act Article 29 places explicit obligations on deployers of AI systems — the businesses (and agencies) using AI to produce work that others rely on. These obligations include appropriate use of the AI system, monitoring for risks, and not modifying the system in ways that increase risk. Your contract is where these obligations become concrete commitments to your clients.
An AI content agency delivers twelve blog posts to a SaaS client over three months. In month four, the client's new marketing director reviews the archive and flags three posts as containing factual errors — product claims that were accurate when written but have since been updated, and two posts that mischaracterize a competitor in a way the client considers professionally problematic.
The client requests that all twelve posts be revised at no charge, on the grounds that the agency's "quality guarantee" — stated verbally during the sales call — covers accuracy. The agency argues that the verbal quality guarantee referred to writing quality, not ongoing factual accuracy as the client's product evolved. The contract says nothing about AI, nothing about factual accuracy, nothing about what "quality" means, and nothing about what the revision policy covers.
Both parties believe they're in the right. The client believes they paid for accurate content. The agency believes they delivered what was scoped. Neither is obviously wrong. But the dispute costs the agency forty hours of unpaid revision work and the client relationship.
A well-written contract would have prevented this. Not by being adversarial — by being specific about what the agency is responsible for, what the client is responsible for (approving content before publication), and what "quality" means in terms of what the agency checks versus what the client needs to verify. The AI disclosure clause would have made the human review process explicit, making clear that human review covers quality but not ongoing factual currency as products evolve.
The contract is not a weapon. It's a shared understanding, written down before the relationship starts.
Generic service agreements were designed for human-produced work. When AI is in the delivery chain, five additional clauses move from optional to essential.
Describe the deliverable in client terms — what they receive, in what format, by when. Specify what is not included. "Ten blog posts per month" scoped without further detail invites disputes about length, format, images, and links. The scope clause is where you operationalize the offer document from Module 2.
Name which parts of the work involve AI and what human review is applied. "Agency uses AI tools to assist in content production. All outputs are reviewed by a human editor for quality and accuracy before delivery. Client acknowledges this process and accepts AI-assisted delivery as compliant with this agreement." This clause satisfies EU AI Act Article 29 and creates a record that the client understood AI was involved.
Who owns the final output? In most service agreements: the client, upon payment. But AI-generated work raises a specific question — the agency's prompt templates and workflows are intellectual property. The output transfers to the client; the process does not. Make this explicit: "Upon receipt of full payment, client receives full ownership of delivered content. Agency retains ownership of prompt frameworks, templates, and workflows used in production."
Cap your liability for AI-generated output failures. "Agency's liability for errors in AI-generated content is limited to the cost of revision or replacement of the affected deliverable. Agency is not liable for business decisions made based on AI-generated content, or for factual changes that occur after content approval." This protects you from claims like the one in the Scenario — while preserving the client's right to a revision.
Define what counts as a revision versus a scope change, and how many revisions are included. "Two revision rounds per deliverable are included. A revision is a change to existing content. A scope change — new topics, additional content, or changes to the approved brief — is billed separately at the day rate." This language prevents scope creep while giving clients reasonable recourse for quality issues.
Article 29 places obligations on deployers of AI systems — including ensuring appropriate use, not using the AI in ways that increase risk beyond what was assessed, monitoring for risks during deployment, and implementing human oversight where required. For AI agencies, Article 29 means the contract should document: what AI is being used for, that human oversight is in place, and that the client has been informed. A contract that is silent on AI may leave the agency in violation of Art.29 user obligations if the output causes harm.
Every service agreement involves three tensions. Getting these wrong produces either a contract that doesn't protect you or one that prevents clients from signing.
The AI disclosure clause protects you legally and builds client trust. But some agencies worry that naming specific AI tools or prompt approaches gives competitors a roadmap. Resolve this by disclosing the category and process (AI-assisted drafting with human review) without disclosing the specific tools or prompt structure. The client needs to know AI is involved and that oversight is applied. They don't need the recipe.
A liability cap that eliminates all client recourse signals that you don't stand behind your work. A liability clause with no cap exposes you to outsized claims for AI errors. The right balance: cap liability at the value of the disputed deliverable (not the whole contract), preserve revision rights, and exclude liability for business decisions made based on the output. This gives clients real recourse while keeping your exposure proportionate to the work.
A revision policy so tight that every change triggers an invoice will make clients feel trapped and damage the relationship. A revision policy so loose that the client can request unlimited changes will destroy your margins. The key distinction is between a revision (changing what was delivered) and a scope change (changing what was agreed). Revisions are included; scope changes are not. Put this distinction in the contract in plain language.
You'll apply all three tensions in the lab — drafting each of the five contract clauses with specific language for your agency's service type.