AI agency contracts in 2026 have to handle scenarios that did not exist three years ago. Who owns the fine-tuned model? Who is liable when an LLM hallucinates a fact in client-facing output? What happens to the API bill when a client ghosts mid-month? A boilerplate marketing agency template will leave you exposed on all three. This is the structure and the specific clauses that actually hold up when a deal goes sideways.
Short answer: Use a two-document structure. An MSA (Master Services Agreement) defines the long-term legal relationship, IP, liability, indemnity, and termination. SOWs (Statements of Work) plug into the MSA per engagement and define scope, deliverables, timeline, and price. Inside the MSA, you need four AI-specific clauses most templates miss: prompt and model IP, AI-output liability cap, third-party API pass-through, and a kill fee tied to data and access rather than hours worked.
This is operational guidance from running an agency and reviewing dozens of contracts with peers. It is not legal advice. Run anything you adapt past a lawyer in your jurisdiction before signing.
Why Most AI Agency Contracts Fail
The contracts most AI agencies use are recycled marketing-agency or web-dev templates from 2019. They were written for a world where the deliverable was a static asset (a logo, a website, a campaign) and the failure modes were predictable (missed deadline, scope creep, bad creative).
AI changes the failure modes. The new ones include:
- Output liability: the AI generated copy that defamed a competitor, recommended an unlicensed practice, or quoted a fabricated source. Who pays?
- Model ownership: you fine-tuned a model on the client's data over six months. The client churns. Do they own the weights? Can you reuse them on the next client?
- API cost blowups: a campaign overruns on tokens, an MCP loop misfires, the bill triples. Whose problem is that?
- Vendor death: the underlying model gets deprecated, prices double overnight, or a provider shuts down. Are you obligated to deliver the same service at the same price?
If your contract was written before late 2023, none of these are addressed. Patch the gaps before your next signature.
The MSA + SOW Structure
One-document contracts are why agencies get into trouble. You sign a comprehensive 14-page agreement for a $3K/month engagement and three months later the client wants to add a content workstream and the whole document has to be redone. Or worse: the original contract terms get assumed to apply to work that was never scoped.
MSA (Master Services Agreement): the umbrella contract that defines the legal relationship between agency and client. It covers IP, confidentiality, liability, indemnity, dispute resolution, payment defaults, and termination. It is signed once and stays in force for the life of the relationship.
SOW (Statement of Work): the per-engagement attachment that defines what is actually being delivered. Scope, deliverables, timeline, milestones, fees, and any deviations from the MSA. A new SOW for each new workstream or each renewal cycle. Short, specific, and easy to amend.
The MSA is the legal armor. The SOW is the operational contract. Keep them separate and your life gets significantly easier when a client wants to expand, pause, or pivot.
Practical rule: if you find yourself rewriting the same liability or IP clauses for every new engagement, those clauses belong in the MSA, not the SOW.
Scope Clauses That Hold Up
Scope creep is the number one margin killer for agencies. AI agencies have a specific flavor of it: clients see AI output and ask for "just one more tweak" five times because it costs the agency nothing visible. The contract has to make the cost visible.
Write your SOW scope around three layers:
- Deliverable units, not hours. "30 LinkedIn posts per month, max 2 revision rounds per post" beats "content marketing services." Cap the revisions explicitly. Anything beyond becomes a change order.
- Inputs the client must provide. List them. Brand voice document, ICP definition, login credentials, API key access, response approval SLA. If the client misses an input deadline, your delivery clock pauses and that should be in writing.
- Out-of-scope, named. Explicitly list what is NOT included. Brand strategy work, legal review of output, custom integrations, additional channels. Vague exclusions get argued. Named exclusions do not.
For sequence-based deliverables (outreach campaigns, content pipelines), define the unit as the campaign or the pipeline run, not the message. "One 5-step multi-channel sequence per ICP segment, up to 3 segments" is a scope. "Outreach" is not.
IP Ownership: Prompts, Models, and Output
This is the clause most templates get wrong because they were written before fine-tuned models and shareable prompt libraries existed. There are four IP buckets in an AI engagement and your contract should address each one separately:
| Asset | Default ownership recommendation | Why |
|---|---|---|
| Final deliverables (posts, emails, sequences, reports) | Client | This is what they paid for. Transfer on full payment, not before. |
| Prompts and prompt libraries built by the agency | Agency, with limited client license | Your prompts are your IP. License them to the client for the engagement, do not assign them. |
| Fine-tuned models trained on client data | Joint or agency, with client data carve-out | The weights embed client data. Client gets the right to receive a copy or have the model deleted on termination. You retain the methodology. |
| Underlying frameworks (workflows, blueprints, SOPs) | Agency, period | This is your business. Never assign it. License it during the engagement only. |
The single most important sentence in this section: "Agency retains all right, title, and interest in pre-existing methodologies, prompt structures, system instructions, and reusable frameworks, regardless of customization performed during this engagement." Without this line, a clever client lawyer can argue that everything you built using their data belongs to them.
On the client side, give them what they actually need: ownership of the final output, the right to a copy of any fine-tuned model on request, and a perpetual license to use deliverables produced during the term. That covers their legitimate concerns without giving away your business.
AI-Output Liability: The Hallucination Clause
An LLM in your pipeline will eventually generate something wrong. A fake statistic. A misattributed quote. A claim about a competitor. A recommendation that crosses a regulatory line. When the client publishes it and gets sued, the contract decides who pays.
Reality check on liability exposure: standard agency contracts cap liability at fees paid in the prior 12 months. For a $3K/month engagement, that is $36K. A single defamation suit or false advertising complaint can exceed that ten times over. AI-generated output multiplies the chance of a triggering incident because volume increases. The cap matters more, not less.
The clauses you need:
- Liability cap tied to fees paid: standard is 12 months of fees, capped. Keep this.
- Client review and approval clause: client is responsible for reviewing and approving AI-generated content before publication. If they publish unreviewed, liability shifts to them. This is the most important sentence in the contract for AI agencies.
- Mutual indemnity, narrowed: you indemnify them for IP infringement caused by the agency's deliverables. They indemnify you for misuse of deliverables, publication without review, or use outside the agreed scope.
- Third-party AI provider carve-out: you are not liable for outages, price changes, deprecation, or output errors caused by underlying model providers (OpenAI, Anthropic, etc.). You pass through their terms.
- No warranty on AI accuracy: explicitly disclaim warranties that AI-generated content is accurate, fact-checked, or fit for regulated use cases. The client's review obligation is the safety net.
The client review clause is not adversarial. It is the only honest way to do AI work. You generate at scale; they approve before publication. If they want to skip approval, they accept the risk in writing.
Kill Fees and Termination Clauses
Clients leave. Sometimes early. The question is what you walk away with.
Most agency contracts have a 30-day termination clause with no kill fee. For monthly retainer work that is fine. For setup-heavy AI engagements where you have spent two weeks configuring infrastructure, training models, and integrating systems before any deliverables ship, it is a disaster. The client can churn at day 31 and you have built an asset they walk away with for nothing.
Use a flat 30-day notice when: the engagement is pure retainer delivery (content production, outreach execution) with low setup cost and the value is in ongoing output.
Use a kill fee structure when: the engagement has meaningful setup work (custom workflows, fine-tuned models, integrations, lead system builds) that the client retains value from after termination. Charge 50% to 100% of remaining contract term, or a flat reimbursement of setup costs, whichever is greater.
Use a minimum term when: infrastructure costs are paid upfront by you and amortized over the engagement. 6-month minimum is standard for build-heavy work. After the minimum, revert to 30-day notice.
The kill fee is not about punishing the client. It is about protecting you from doing the expensive work in month 1 and losing the easy revenue in months 4 through 12. Frame it that way in the conversation and it lands.
Two other termination details that matter:
- Data and credential return: on termination, you return client data and revoke access to client systems within a defined window (typically 14 days). Spell out what "return" means: export format, delivery method, deletion confirmation.
- Transition assistance: if termination is requested, you will provide reasonable transition support at your standard hourly rate. Reasonable means defined hours, defined scope, no obligation to train a replacement.
Payment Terms and API Cost Pass-Through
Two payment clauses save you the most grief.
First: payment in advance, not in arrears. Net-30 invoicing for retainer work is how agencies die. The first month is paid before kickoff. Every subsequent month is paid by the first business day of the month. Work stops if payment is not received within a defined grace period (5 business days is reasonable). This is non-negotiable in the MSA, not a per-deal conversation.
Second: API and third-party cost handling. Decide your model and put it in writing:
- BYOK (Bring Your Own Key): client connects their own API keys for OpenAI, Anthropic, etc. They pay providers directly. You charge a platform fee on top. Cleanest, lowest risk, easiest to scale.
- Pass-through with markup: you pay API providers and bill the client at cost plus a markup (20% to 30% covers reconciliation overhead). Requires monthly reconciliation and a cap clause to prevent runaway bills.
- All-in flat fee: API costs included in retainer. Only works if you control consumption tightly and have margin to absorb spikes. Risky for high-volume content or outreach work.
Whichever model you pick, the contract must address what happens when usage exceeds projected volume. A consumption cap or overage clause prevents the conversation no one wants to have at month-end.
Getting a Working Template
Templates speed up the work but they do not replace local legal review. Different jurisdictions, different client sizes, different deliverables all change the right structure. What is reasonable indemnity language in the US may be unenforceable in the EU. What works for a $3K/month retainer is overkill for a $500/month engagement and dangerously thin for a $25K/month build.
The pattern that consistently works for AI agencies:
- Start with an MSA template adapted from a tech-services or SaaS agreement, not a marketing-agency template.
- Add the four AI-specific clauses: prompt/model IP, AI-output liability, third-party pass-through, AI accuracy disclaimer.
- Build SOW templates per service line (outreach, content, full-service) with named deliverable units and revision caps.
- Have a lawyer in your jurisdiction review once. Reuse that base for every client.
Inside the ACA community we share template language, redlines, and the actual MSAs and SOWs members are using right now, including the AI-output liability and kill-fee clauses described above. If you want the working documents to adapt rather than starting from a blank page, that is the place to get them.
Frequently Asked Questions
Do I need a lawyer to write my AI agency contract?
You need one to review it before you use it, yes. You do not need one to draft from scratch. Use a template adapted for AI services, customize it for your specific deliverables and jurisdiction, then pay a lawyer for a one-time review. Budget $500 to $2,000 depending on jurisdiction. After that you can reuse the base contract for every client without further legal cost until something material changes.
Who owns AI output: the agency or the client?
The default and most defensible position: the client owns the final delivered output (the posts, emails, reports, sequences they paid for) on full payment. The agency retains ownership of the methodology, prompts, frameworks, and any pre-existing IP used to produce the output. Fine-tuned models trained on client data sit in the middle - typically the client receives a copy on request and the agency retains the right to apply the same training approach to other engagements, with client data fully removed.
What happens if AI-generated content causes a lawsuit against the client?
This is exactly why the client-review clause matters. If your contract requires the client to review and approve AI-generated content before publication, and they publish without review, liability shifts to them. If they reviewed and approved, the agency's liability is capped at the contract's stated limit (typically 12 months of fees paid). Mutual indemnity language further narrows exposure on both sides. Without these clauses, the agency is exposed to the full value of any judgment.
Should I charge setup fees or work them into the retainer?
Charge them separately. A setup fee paid upfront covers the build-heavy work in month 1 (configuration, integration, model training, sequence creation) that is not repeated in subsequent months. Working setup into the retainer makes early-churn engagements unprofitable. A typical structure: setup fee equal to 1 to 2 months of retainer, paid before kickoff, plus the retainer starting at go-live. Pair this with a kill-fee or minimum-term clause to protect the retainer side.
How long should an AI agency MSA be?
Between 8 and 15 pages for most engagements. Shorter than that and you are missing protections. Longer than that and you are scaring small clients and creating clauses no one will read. The SOWs that attach to the MSA should be 1 to 3 pages each, deliverable-focused, easy to redline and re-sign per engagement. The MSA is the legal document. The SOW is the working agreement.
What is the minimum contract term I should require?
For pure retainer work with low setup cost, no minimum is fine; 30-day notice is enough. For build-heavy engagements (custom workflows, fine-tuned models, integrations, lead infrastructure), a 3 to 6 month minimum is standard and defensible. The justification is straightforward: you front-load investment in month 1 that the client benefits from for the rest of the engagement. Minimum terms align incentives with the actual cost structure of the work.