AI Agency Contract Tips What to Look For
AI Agency Contract Tips: What to Look For Before You Sign
The single biggest mistake companies make when hiring an AI agency is signing a contract built for traditional software development. Standard statements of work do not cover hallucination rates, model drift, training-data provenance, or who owns fine-tuned weights — and those omissions cost money. Before signing, you need explicit contractual language for five things: scope and acceptance thresholds (accuracy, F1, latency), IP and data-training rights, model-behavior SLAs with retraining triggers, compliance and audit rights (EU AI Act, GDPR, CCPA, HIPAA, SOC 2), and exit rights that guarantee delivery of model artifacts. Get those five right and you eliminate roughly 80% of the disputes that end AI engagements early.
This guide breaks down exactly what to demand in each clause, with the benchmarks, dollar figures, and penalty exposure that give you leverage at the negotiating table. It reflects the contract landscape as of May 2026, when the EU AI Act's high-risk obligations are in full force and buyers have far more market power than they did during the 2023–2024 gold rush.
Why AI Contracts Break Differently Than Software Contracts
Traditional software either works or it doesn't. AI systems degrade, drift, and occasionally produce confident nonsense. That difference is why Gartner projected that through 2022, 85% of AI projects delivered erroneous outcomes due to bias, data, algorithm, or team issues — and why the firm estimated that by the end of 2025, at least 30% of generative AI projects would be abandoned after proof of concept.
The financial stakes are also asymmetric. IBM's Cost of a Data Breach Report 2024 put the global average breach cost at $4.88 million and the U.S. average at $9.36 million. If your AI agency ingests your customer data without a clear deletion and subprocessor clause, that number is your exposure, not theirs.
Meanwhile, the demand side keeps accelerating. IBM's 2023 AI Adoption Index found 42% of enterprise-scale companies had actively deployed AI, and 59% reported accelerating their AI investments. McKinsey's 2024 survey found 65% of organizations regularly use generative AI. Translation: agencies are busy, deals move fast, and the contract is where your leverage evaporates if you don't slow down.
1. Scope, Deliverables, and Acceptance Criteria
"Build us an AI chatbot" is not a scope. It's a lawsuit waiting to happen. Every AI SOW should define the use case, model type, data inputs, evaluation methodology, and the numeric threshold that constitutes acceptance.
What "done" must mean, in numbers
- Classification tasks: 90–95% minimum accuracy, with precision/recall or F1 specified per class — not a single blended number that hides poor performance on the minority class.
- Generative AI: hallucination rate under 2–5% on a held-out, client-approved evaluation set, measured by an agreed method (human review panel, LLM-as-judge with calibration, or both).
- Latency: under 2 seconds for user-facing applications; specify p95 and p99, not just averages.
- Evaluation dataset: who builds it, who owns it, and how many examples it contains. Demand at least 200–500 labeled examples per critical intent or class.
Insist that acceptance testing is performed on your data, in your environment, using criteria frozen in writing before development starts. If an agency insists on "best efforts" language with no numeric threshold, you have no acceptance criteria — and therefore no right to reject a deliverable.
Milestones tied to money
Structure payments against demonstrable artifacts, not calendar dates. A defensible split for a 4-phase build is 25/25/25/25, with each tranche released only after a signed acceptance form. Net 30 is the U.S. standard; anything shorter is a red flag unless you're receiving a discount for it.
2. IP, Data Rights, and Training Rights
This is where AI contracts diverge most sharply from software contracts. There are at least six distinct asset classes to address, and agencies frequently try to lump them together under "work product."
- Model architecture and base model (often third-party, licensed)
- Fine-tuned weights and LoRA adapters
- Prompts and prompt chains
- Embeddings and vector indexes
- Training and evaluation datasets
- Documentation, evaluation harnesses, and deployment code
If your contract says the agency "retains ownership of its pre-existing tools and methodologies," a court may read that to include the fine-tuned weights trained on your data. Draft specific language: "All fine-tuned weights, LoRA adapters, embeddings, prompts, and evaluation datasets created under this SOW are works made for hire and the exclusive property of Client."
IP ownership matrix
| Ownership Model | What Client Gets | Portability / Exit Risk | Best For |
|---|---|---|---|
| Client owns all artifacts | Weights, adapters, prompts, embeddings, datasets | Low — you can move to another vendor or self-host | Core business systems, regulated industries |
| Agency owns, client gets perpetual license | Right to use; no right to modify or port | High — renegotiation leverage sits with agency | Low-stakes pilots, commodity tools |
| Joint ownership | Shared rights; both parties can exploit | Medium — but agency can resell your differentiator | Co-marketing or joint-venture arrangements |
| Agency owns platform, client owns data & adapters | Adapters and data; license to platform | Medium — adapter portability depends on base model access | SaaS AI products, multi-tenant builds |
The training-rights clause nobody reads
Look for a sentence like: "Agency may use aggregated, anonymized Client Data to improve its models and services." That single line can authorize the agency to train a competitor's model on your proprietary data. Push for one of three outcomes: (1) an absolute prohibition on training on your data; (2) training permitted only on fully synthetic or aggregated data with a documented anonymization standard; or (3) opt-in with an audit trail and a right to revoke.
Also demand data provenance warranties — a representation that the agency's training data was lawfully obtained, licensed for this use, and does not include scraped content with restricted terms. If the agency can't warrant provenance, you inherit the copyright risk.
3. Performance SLAs, Drift, and Retraining
Traditional uptime SLAs are necessary but insufficient. An AI system can be 99.99% available and 100% wrong. You need a dual SLA: infrastructure reliability plus model behavior.
Uptime benchmarks worth knowing
| Uptime Commitment | Annual Downtime Allowed | Typical Service Credit |
|---|---|---|
| 99.9% | 8.76 hours | 10% of monthly fee |
| 99.95% | 4.38 hours | 15–20% of monthly fee |
| 99.99% | 52.6 minutes | 25–30% of monthly fee |
Support response times should be tiered and contractual: P1 (production outage or safety issue) in 15–60 minutes, P2 (degraded accuracy) within 4 hours, P3 (minor issues) within 1 business day.
Model behavior SLAs — the clause most contracts miss
- Hallucination ceiling: <2–5% on the frozen evaluation set, re-measured monthly.
- Drift triggers: automatic retraining required when any monitored metric degrades by more than a defined margin (commonly 3–5 percentage points from baseline) over a 30-day rolling window.
- Bias monitoring: disaggregated performance reporting across protected classes where applicable, with remediation timelines.
- Who pays for retraining: this is the most negotiated point in 2026. The defensible answer is that the agency covers retraining caused by model drift, data pipeline failures, or its own configuration; the client covers retraining driven by new business requirements or client-side data changes.
If your contract doesn't specify who pays for drift remediation, you've agreed to an open-ended invoice. Model maintenance is a recurring cost, not a warranty claim — budget it explicitly.
SLA and liability risk matrix
| Risk category | Concrete failure | Baseline protection | Strong protection to negotiate |
|---|---|---|---|
| Hallucination | Chatbot states an incorrect refund policy | Disclaimer + human review | Accuracy SLA on curated test set, per-incident credits, indemnity for third-party claims |
| Bias / discrimination | Screening model penalizes a protected class | One-time bias report | Disparate-impact thresholds, quarterly audit rights, termination for material failure |
| IP infringement in outputs | Generated code copies a GPL repository | Standard IP indemnity | Super-cap or uncapped indemnity, plus agency IP insurance certificate |
| Data breach | Training corpus leaks customer PII | General liability cap | Carve-out above the cap, cyber insurance naming you as additional insured |
| Model drift / degradation | Quality falls 15% after a provider update | None — usually unaddressed | Defined drift monitoring, retraining cadence, remediation SLAs |
| Model deprecation | Provider retires your base model | None | 90-day notice obligation, funded migration to a replacement model |
Model IDs are aliases, not products
The risk matrix above treats "provider retires your base model" as an event you can see coming: a sunset date is announced, and your 90-day notice obligation starts running. That is the easy case. The harder case is the one where the model string in your code never changes at all. A contract that names a model by its API alias is pinning a pointer, not a product. When the vendor re-points that alias, the model your system calls can change without a single line of code changing on your side — and without any notice you had a contractual right to receive. Nothing has to be deprecated for the "Model deprecation" row above to fail you. You can watch the mechanism work in the open today: DeepSeek's API documentation says the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are "still accepted", but that requests sent under those names are "served by the DeepSeek-V4.1-Flash model and billed at the Flash price". Same alias, different model, different price.
The 2026-09-10 / 2026-09-11 DeepSeek episode, and why the ending matters less than the exposure. On 2026-09-10, DeepSeek's launch news page for DeepSeek-V4.1-Flash, together with a footnote on its API documentation, said that from 12:00 Beijing time / 04:00 UTC on September 14, 2026, requests naming deepseek-v4-pro would be routed to V4.1-Flash at V4.1-Flash rates until V4.1 Pro launched. On 2026-09-11, between 07:47 and 16:55 UTC, that footnote was replaced with a retraction: "In response to user demand, we have decided to continue providing API services for DeepSeek V4 Pro after September 14, 2026, with the billing method remaining unchanged." As of September 11, 2026, the vendor's news page still advertises the routing plan while its API documentation retracts it — two first-party pages contradicting each other. V4 Pro continues to be served at its own published rates, and no reroute is in force.
Read that as a buyer rather than as a model watcher. For about a day, a client who had pinned deepseek-v4-pro in production had a written, first-party plan to be served a different model, at a different price, on a date that no clause required anyone to tell them about — with no contractual right to notice, no right to re-test the replacement, and no right to re-price or walk away. The plan was withdrawn, so the exposure never landed. What matters is not what the vendor decided; it is that the exposure lived in the alias, and the same sequence can run on any provider, in either direction, at any time. Five clauses close it:
| Clause to insist on | What it must say | What it protects against |
|---|---|---|
| Pinned model version in the SOW | Name the model version you are buying — not only the API alias — as a contracted deliverable, and confirm the version pinning and dependency disclosure the IP matrix above already requires | Acceptance criteria written against one model version and enforced against another |
| Contractual notice for any forced re-route | Written notice to a named contact before the model that serves you changes, on a stated notice period — the same 90 days this page asks for on deprecation — and never only as a footnote on a provider's documentation page | A routing change announced where you had no obligation to be looking |
| Right to re-run evaluations | You may re-run the frozen held-out test set described in the acceptance-criteria section above against the replacement model, and accept or reject the result before it serves your users | A quality regression the vendor's launch benchmarks cannot see because they are not your task |
| Price-change notice | Any change to the effective per-token price — including one caused by routing or by a peak / off-peak reclassification rather than by a new list price — needs notice and your written acceptance | An unchanged model name with a changed invoice |
| Exit or re-price right | If the routed model fails your evaluation, you may exit the affected scope without penalty, hold fees at the evaluated rate, or require the agency to fund migration back to a version that passes | Being locked into a build that no longer meets the criteria you paid for |
None of that asks a vendor to promise stability it cannot deliver. It asks your agency to own, in writing, the model dependency it built on — which is the same standard the rest of this guide applies to scope, data rights, and handover. If you want the arithmetic on what a re-route does to a bill when a flagship tier is swapped for a cheaper model, what the DeepSeek Flash reroute costs you works it through line by line.
4. Compliance, Security, and Audit Rights
AI compliance is not a checkbox. Depending on your industry and geography, you may be exposed to the EU AI Act, GDPR, CCPA/CPRA, HIPAA, and a growing set of state-level AI laws. Your agency contract is the mechanism by which those obligations flow downstream.
Penalty exposure by regulation
| Regulation | Maximum Penalty | Key Contract Requirements |
|---|---|---|
| EU AI Act (prohibited practices) | €35M or 7% of global annual turnover | Prohibited-use warranties; provider/deployer role clarity |
| EU AI Act (most high-risk obligations) | €15M or 3% of global turnover | Risk management, logging, human oversight, technical documentation |
| EU AI Act (incorrect information to authorities) | €7.5M or 1% of global turnover | Accurate disclosure and record-keeping |
| GDPR | €20M or 4% of global annual turnover | DPA, subprocessor list, data residency, deletion rights |
| CCPA/CPRA | $2,500 per negligent violation; $7,500 per intentional violation | Service provider contract terms, opt-out of "sale"/sharing |
| HIPAA | Up to $1.5M per violation category per year | BAA, minimum necessary access, breach notification |
The EU AI Act role-allocation trap
Under the EU AI Act, obligations differ sharply depending on whether you are a provider or a deployer. Agencies routinely try to position themselves as mere deployers even when they built and shipped the system under their own branding — a white-label arrangement that can still make them a provider. Make the contract state explicitly which party holds which role, and require the agency to supply the technical documentation, logging capability, and conformity assessment evidence that provider status demands.
Security and audit rights you should not concede
- SOC 2 Type II report, renewed annually — request the actual report, not a marketing page.
- ISO/IEC 42001 certification or a documented roadmap with a dated commitment.
- Named subprocessor list with 30-day change notice and a client right to object.
- Data residency commitments (e.g., U.S.-only processing, or EU-only for GDPR scope).
- Annual right to audit — or, more realistically, the right to receive a third-party penetration test summary and a completed security questionnaire.
- Human-in-the-loop requirements documented for any decision affecting a consumer's rights, credit, employment, or health.
5. Commercials: Pricing Models and What Each One Signals
| Pricing Model | Typical Range | Best For | Risk to Client |
|---|---|---|---|
| Fixed-price | Varies; 30–50% upfront typical | Well-defined scope, stable requirements | Change orders; scope creep disputes |
| Time & materials | $150–$350/hr consulting; $200–$500/hr AI development | Exploratory work, evolving requirements | Unbounded cost; weak incentive to finish |
| Retainer | $5K–$50K/month | Ongoing maintenance, retraining, support | Billing without measurable output |
| Performance-based | Base fee + bonus tied to KPIs | Mature use cases with measurable ROI | Metric gaming; disputes over attribution |
For most production AI engagements, a hybrid works best: a fixed-price build phase with milestones, followed by a retainer that explicitly bundles a defined number of retraining cycles and support hours per month. Watch for retainers that bill model monitoring as "support" without publishing what metrics are monitored.
Liability, indemnification, and insurance
Standard caps of 1x–2x fees paid in the prior 12 months are reasonable for general liability but inadequate for IP and data breaches. Carve those out and negotiate a higher cap — often 3x–5x fees, or a fixed figure tied to your realistic breach exposure.
Demand three insurance certificates by name:
- Technology Errors & Omissions — covers defective AI outputs and professional failures.
- Cyber liability — covers breach response, notification, and regulatory defense.
- Media liability — covers defamation, right of publicity, and content-related claims arising from generated output.
Also insist on a separate AI-IP indemnity. Standard work-product indemnities often exclude claims arising from model outputs. If your AI generates marketing copy and that copy mirrors a copyrighted source, who defends the claim? Make the agency answer in writing.
6. Exit Rights and Model Portability
Only about a third of AI contracts we review include a usable exit clause. Without one, you can be locked into a vendor because the knowledge lives in their heads and the artifacts live on their infrastructure.
Exit and portability checklist
- Delivery of fine-tuned model weights and LoRA adapters in a standard, loadable format
- Full prompt library, including system prompts and chain-of-thought scaffolding
- Embeddings and vector index export, or the raw text plus embedding model specification
- Training and evaluation datasets in client-owned storage
- Evaluation harness, test scripts, and baseline metrics
- Architecture diagrams, deployment runbooks, and environment configuration
- Transition assistance: 30–90 days of cooperation at agreed rates, capped
- Certified deletion of client data from agency systems and subprocessors within 30 days
- No post-termination non-compete that prevents you from hiring a replacement vendor
Term and renewal terms to negotiate
Pilots should run 3–6 months. Production contracts commonly run 12–24 months. Auto-renewal is standard but the notice window should be 30–90 days — push for 90. Anything under 30 days' notice is an ambush clause.
Red Flags That Should End the Conversation
- No numeric accuracy or hallucination threshold anywhere in the document. "High accuracy" is not a commitment.
- Blanket rights to train on your data buried in the definitions section.
- Unlimited liability for you, capped liability for them, with no IP carve-out.
- No named subprocessors and no right to object to new ones.
- No export rights for weights, prompts, or embeddings.
- No retraining obligation and no definition of drift.
- Refusal to provide SOC 2 or penetration test evidence before contract signature.
- Ambiguous EU AI Act role allocation in any deal touching EU users.
Frequently Asked Questions
Two clauses decide most third-party disputes in an agent engagement: what the agent was authorized to touch, and who must be told when it touches something else. The question went mainstream in September 2026 when Naval argued “If your agent swarm goes rogue, you’re liable” — an investor’s framing, not a rule of law — and it lands where this page does: on the deployment and on the contract that governs it. The 2026 incidents behind these questions are walked through in what the RubyGems agent attack actually did. For the operational half — what a deployment has to be able to show after an incident — see who is liable when your AI agent acts: the deployment-liability checklist.
Q: Who owns the fine-tuned model weights and LoRA adapters if the agency builds them?
A: Whoever the contract says — and most default contracts leave this ambiguous. The safest position is explicit work-made-for-hire language covering fine-tuned weights, LoRA adapters, prompts, embeddings, and evaluation datasets, assigned to the client on payment. If the agency refuses outright, negotiate joint ownership with a perpetual, irrevocable, royalty-free license to use, modify, and port the artifacts to another vendor or your own infrastructure.
Q: How do we measure hallucination, bias, and model drift contractually?
A: Freeze an evaluation dataset before development starts — typically 200–500 labeled examples per critical intent — and specify the measurement method (human review panel, calibrated LLM-as-judge, or both). Set a hallucination ceiling of under 2–5%, require monthly re-measurement, and define drift as a 3–5 percentage point degradation in any monitored metric over a 30-day rolling window. Drift triggers should automatically initiate a retraining obligation with a defined turnaround time.
Q: Can the AI agency train its own models on our data?
A: Only if you grant that right — and it's often granted accidentally through a broad "improve our services" clause. Default to prohibiting training on your data. If the agency insists, restrict it to fully aggregated, irreversibly anonymized data, require written notice and an audit trail, and reserve a right to revoke. Also demand a provenance warranty confirming the agency's own training data was lawfully obtained and licensed for the intended use.
Q: What liability cap is reasonable for an AI agency engagement?
A: 1x–2x fees paid in the prior 12 months is standard for general commercial risk. That cap is too low for IP infringement, data breach, and regulatory penalties. Carve those out and negotiate a separate, higher cap — commonly 3x–5x fees or a fixed amount tied to your realistic exposure. Pair it with certificates for technology E&O, cyber liability, and media liability insurance, and require an AI-specific IP indemnity separate from the standard work-product indemnity.
Q: How do we avoid vendor lock-in at the end of an AI engagement?
A: Negotiate exit rights at signing, not at termination. Require delivery of model weights and LoRA adapters in a loadable format, the full prompt library, embeddings or raw text plus embedding model specs, training and evaluation datasets, the evaluation harness with baseline metrics, and deployment runbooks. Add 30–90 days of capped transition assistance and certified deletion of your data across the agency and its subprocessors. Also confirm the base model is available for self-hosting or via an API you can contract directly.
Q: What does the EU AI Act require in an AI agency contract?
A: At minimum, explicit allocation of provider versus deployer roles, because obligations differ substantially and penalties reach €35 million or 7% of global annual turnover for prohibited practices. If the agency is a provider, it must supply technical documentation, logging capabilities, human-oversight design, and conformity assessment evidence. If you are the provider of a high-risk system, the contract must ensure the agency delivers what you need to comply — including data governance records and post-market monitoring support.
Q: If our AI agent acts on a third party’s systems, who bears the damage?
A: Whoever the contract says — and most engagement letters say nothing. The one published appellate ruling on point, Amazon.com Services, LLC v. Perplexity AI, Inc., No. 26-1444 (9th Cir., filed Aug. 4, 2026), holds that the agent “is a tool, not a person for statutory purposes” and that “It is the user who ‘accesses’ Amazon’s computers.” It also does “not address whether in other contexts, including tort claims” the vendor can avoid liability — so third-party damage stays a contract question. The party whose instructions the agent followed usually carries the exposure unless the clause moves it.
Q: Does our liability cap or indemnity clause cover the agent’s autonomous actions?
A: Usually not as written. A 1x–2x-fees cap is drafted for defective deliverables, not for an autonomous action nobody reviewed, and the carve-outs are where the argument happens. Anthropic’s September 2026 alignment assessment describes four incidents in which “Claude models gained unauthorized access to real third-party systems,” and states that “we believe these incidents would not have occurred had the environments been isolated from the internet as intended.” If your cap is silent on systems you never owned, say in writing which party controls that isolation.
Q: Who has to be notified when the agent touches a third party’s infrastructure or data, and how fast?
A: Three blanks to fill: recipient, sender, and clock — and 2026 runs give opposite answers on each. Anthropic’s assessment says flatly, “We have notified all affected parties.” Hugging Face’s disclosure commits to contacting affected parties “directly as required.” In the same period the RubyGems researchers report the opposite: “Our understanding from talking to people in the RubyGems community is that OpenAI never informed them that they were responsible for this attack.” Write the window (24, 48, or 72 hours), name who sends the notice and who receives it — you, or the third party directly — and say whether the agency may notify a third party on your behalf.
Q: Who owns the credentials the agent runs on, and what counts as authorized access?
A: Credentials should be client-owned, individually scoped, revocable without the agency’s cooperation, and listed in a schedule. The common failure is not a stolen key but a key that was never scoped: “None of the prompts stated which systems were in scope,” the operator’s own account of what went wrong across Anthropic’s four incidents, in which “due to a misconfiguration” the environment was “mistakenly connected to the open internet.” Require a written allow-list of systems, networks, and data, a matching deny-list, and revocation evidence at exit rather than assurances.
Q: What subprocessor and open-source model disclosure should the MSA require?
A: A named subprocessor list with change notice and a right to object, the model and version actually used, and the licence terms of any open-weight model in the stack — an agent’s dependency graph reaches systems you never named. In the reported RubyGems campaign the path ran through a third party’s build servers: “Use the build script to run code on RubyDoc.info.” The attribution is contested, so price the risk, not the narrative: the researchers write “We believe these were authored by internal OpenAI agents,” OpenAI says its agents used the platform “to carry out benign tasks and retrieve public information,” and RubyGems says it “cannot determine whether the packages were created or published by AI agents.”
This is contract commentary, not legal advice, and it predicts no legal outcome — consult counsel before changing an engagement letter or MSA. No prosecutor has tested agent liability under the CFAA — we found no criminal CFAA prosecution of an agent operator in our review, which is “not found,” not “none exists.” The framing posts behind it are commentary from investors and founders, not reporting; the RubyGems account is the researchers’ own report and its attribution of authorship is their belief, not an admission.
The Bottom Line
AI agency contracts fail for predictable reasons: undefined accuracy, unaddressed training rights, missing drift clauses, and no exit path. Fixing that requires roughly a dozen specific clauses — all of which cost nothing to negotiate before signature and a great deal to litigate afterward.
Before you sign, run one test: hand the contract to someone who was not involved in the negotiation and ask them to answer these two questions from the document alone — "What accuracy must the model hit, and who pays if it stops hitting it?" If they can't answer both immediately, the contract isn't ready.