Most AI agency contracts are written by the agency's lawyers to protect the agency. The client frequently signs them without negotiation and discovers the problems later — when the project goes sideways, when they want to switch agencies, or when they discover the model they paid to build belongs to the agency.
This is a guide to the contract provisions that matter, what each one should say, and what to reject when you see it. It's not a substitute for legal counsel — for a significant engagement, you should have an attorney review the final agreement. But you can and should understand these terms before you get to that review.
IP Ownership: The Most Important Clause
Intellectual property ownership determines who owns what you paid to build. In a poorly written AI agency contract, you might pay $200,000 to develop a machine learning model and end up owning nothing more than a license to use it.
What you need: A clear statement that all custom code, trained model weights, fine-tuned parameters, training pipelines, evaluation frameworks, and associated documentation created during the engagement are assigned to you upon full payment. Not licensed — assigned. The difference matters.
What acceptable language looks like:
*"All work product, including but not limited to custom software, trained model weights, model architectures, training data (if created or transformed by the Agency), documentation, and any other deliverables created specifically for Client under this Agreement shall be considered works made for hire and shall be owned exclusively by Client upon delivery and final payment. Agency hereby assigns to Client all intellectual property rights in such work product."*
What to watch for and reject:
*"Agency retains ownership of all models and grants Client a perpetual license to use the deliverables."* — This is the clause that leaves you with a license rather than ownership. You cannot modify a licensed model without agency permission, cannot move it to a different vendor, and cannot have another firm work on it. Reject this.
*"Client owns deliverables excluding Agency's proprietary tools, frameworks, and pre-existing IP."* — Reasonable in concept (you shouldn't own their internal tools), but "pre-existing IP" can be defined to swallow the project entirely if you're not careful. Insist on a specific schedule or exhibit that lists the agency's pre-existing IP before you sign.
*"Agency retains the right to use anonymized work product for model training, benchmarking, and publication."* — This means your business data and the patterns in your trained model can be used to benefit other clients. If your model was trained on proprietary business data, this clause is unacceptable. Strike it.
Data Handling and Privacy
AI projects involve data — often sensitive data. The contract must specify how data is handled during the engagement, who has access, and what happens to it when the engagement ends.
Required provisions:
Data use limitation: The agency may only use your data for the specific purpose of performing the services described in this agreement. They cannot use it for agency research, benchmarking against other clients, or model training for other purposes.
Data security: Specify minimum security requirements — encryption in transit and at rest, access controls limited to project team members, audit logging of data access. If you're in a regulated industry, reference the specific compliance framework (HIPAA, SOC 2, GDPR).
Data return and deletion: Upon project completion or termination, the agency must return all your data and certify in writing that all copies have been deleted from their systems within a specified timeframe (typically 30 days). "Return and delete" sounds obvious, but without this clause, your data may sit on agency servers indefinitely.
Subcontractor limitations: Agencies often use subcontractors for portions of AI projects. Your contract should require that any subcontractor be bound by the same data handling obligations as the agency, and that you be notified before any subcontractor receives access to your data.
Breach notification: The agency must notify you within a specific timeframe (72 hours is standard) if they discover or suspect a data breach affecting your data.
Payment Milestones
Never agree to a payment structure that pays the majority of fees upfront. The standard for a well-structured AI project engagement is milestone-based payment, where each milestone payment is tied to a specific, verifiable deliverable.
Recommended payment structure:
- 20–25% upfront (covers project kickoff and initial infrastructure costs)
- 25–30% at first major milestone (typically proof of concept or working prototype)
- 25–30% at second major milestone (typically testing complete, production-ready)
- 15–20% at final delivery and client acceptance
The acceptance process matters: each milestone payment should be conditioned on written client acceptance of that milestone, with a defined process for what happens if acceptance is withheld (typically a cure period, then escalation).
What to reject:
*"50% upon signing, 50% upon project completion."* — This gives you no leverage if the project derails at week 6 of a 16-week engagement. You've already paid half and can't stop payment without breaching the contract.
*"Payment due net-15 from invoice, regardless of project status."* — This is standard in service agreements and not inherently problematic, but it should be tied to milestone achievement, not arbitrary invoice dates.
*"Client agrees to pay for all time and materials at rates specified in Schedule A."* — Open-ended time-and-materials contracts without caps are appropriate only for ongoing support engagements. For project delivery, insist on fixed-price or capped time-and-materials with a project change order process.
Service Level Agreements (SLAs)
For AI systems that go into production, you need SLAs that specify how the system must perform and what happens when it doesn't.
Model performance SLAs: Define minimum acceptable performance metrics (accuracy, precision, recall, latency, throughput) at delivery and for the duration of any support period. Specify what happens when the model's performance degrades below these thresholds — the typical remedy is the agency's obligation to retrain or remediate at no additional cost.
System availability SLAs: For production-deployed AI systems, specify minimum uptime (99.5% is reasonable for most applications; 99.9% for critical systems). Define what constitutes downtime and how it's measured.
Response time SLAs: Specify how quickly the agency must respond to production incidents (typically 4 hours for critical issues, 24 hours for non-critical) and how quickly they must resolve them (typically 24–48 hours for critical, 5–10 business days for non-critical).
Model drift provisions: AI models degrade over time as the real world diverges from the training data. Your SLA should include provisions for model monitoring and a retraining obligation if performance drops below specified thresholds. Without this, you're responsible for catching drift and paying full project rates to fix it.
Exit Provisions
Exit provisions govern what happens when the engagement ends — by completion, by mutual agreement, or by one party terminating for cause.
Termination for cause: Either party should be able to terminate the agreement if the other materially breaches their obligations and fails to cure the breach within a specified period (typically 30 days). This requires clearly defined obligations on both sides — which is another reason to be specific about deliverables and milestones.
Termination for convenience: You should be able to exit the engagement for any reason with reasonable notice (typically 30 days), paying for work completed to date plus a modest cancellation fee. Agency contracts often try to make termination for convenience expensive enough to be a trap. A cancellation fee of more than 10–15% of the remaining contract value is excessive.
Transition assistance: Upon termination, the agency must assist with an orderly transition for a specified period (typically 30–60 days). This includes providing full documentation, granting access to all systems and repositories, and cooperating with any successor vendor. Without this clause, exiting becomes a hostage situation.
Continuity of personnel: If key personnel leave the project during the engagement, specify the agency's obligation to provide replacements with equivalent qualifications within a defined timeframe (typically 2 weeks), and your right to approve replacements.
Liability and Indemnification
Cap on liability: Agencies typically cap their liability at the fees paid in the preceding 3–6 months. For a short project engagement, this can mean your remedy for a catastrophic failure is a refund of the last month's invoice. Push for a liability cap equal to the total fees paid under the contract.
Carve-outs from the liability cap: Certain categories should be excluded from liability caps: data breaches caused by the agency's negligence, willful misconduct, and indemnification obligations. These are non-negotiable.
Indemnification: The agency should indemnify you against claims that their deliverables infringe third-party intellectual property rights. This matters because AI systems trained on copyrighted data or built using unlicensed models can expose you to IP liability. If they built it, they should indemnify you for IP claims against it.
Warranties
Fitness for purpose: The agency warrants that the deliverables will perform as specified for a minimum period after delivery (typically 90 days). During this period, defects should be corrected at no additional charge.
No third-party infringement: The deliverables do not infringe any third-party patents, copyrights, or trade secrets to the agency's knowledge.
Authority to perform: The agency has the legal right to enter into this agreement and to grant the IP rights described.
Red Flags: Walk Away or Renegotiate
These provisions should trigger either significant negotiation or walking away entirely:
"Agency retains all rights to deliverables including model weights." — You'd own nothing. Walk away.
"Client waives any claims related to model performance or business outcomes." — This eliminates your ability to hold the agency responsible for anything. Reject.
"Agency may modify the scope at its discretion to deliver equivalent value." — Gives the agency unilateral authority to decide what you receive. Not acceptable.
"Confidentiality obligations survive for 1 year after the agreement." — Standard confidentiality should survive indefinitely for trade secrets and for as long as the data remains sensitive. One year is too short.
"All disputes are subject to mandatory arbitration in [distant city] under [favorable agency terms]." — Mandatory arbitration isn't inherently bad, but it should be in a neutral venue with balanced arbitration rules. Dispute resolution designed to be expensive and inconvenient for you is not neutral.
"Agency's liability is limited to $5,000 for any claim." — If you're signing a $150,000 contract, a $5,000 liability cap is effectively zero liability. This is not a negotiating position; it's a warning sign about how the agency views client relationships.
Review any contract with these clauses carefully. The aiagencymap.com directory lists agencies broadly, but your due diligence on contract terms is the safety net that protects you after selection. Get your legal team involved before signature on any engagement above $25,000.
Ready to Find the Right AI Agency?
Browse 700+ verified AI agencies. Filter by tech stack, industry, location, and client ratings.
Browse AI Agencies