Contracts you’ve signed before will hurt you here
The lawyers reviewing your AI vendor contracts have not been reviewing AI vendor contracts very long. Most of the boilerplate they reach for was written for SaaS — predictable workloads, mature exit paths, well-understood data flows. AI vendor contracts in 2026 contain a dozen specific clauses that will quietly transfer cost and risk to you, and that an experienced SaaS reviewer will let through because they look like everything they’ve seen before.
Below are the twelve we negotiate out, modify, or refuse on every AI procurement we touch. None of them are exotic. All of them are routine.
1. The “model improvement” clause
The vendor reserves the right to “improve their models using customer data.” This sounds benign and is not. It typically means your prompts, your outputs, and sometimes your underlying data are being added to a training corpus. Even when “anonymized,” that data is now embedded in a model that will be used by your competitors.
Fix: Strike the clause entirely. Replace with explicit no-training language and require positive opt-in for any data use beyond serving your requests.
2. The “third-party model provider” carve-out
The vendor lists their model providers (OpenAI, Anthropic, etc.) and disclaims responsibility for the providers’ terms. This means the vendor’s commitments to you flow through provider commitments they don’t control.
Fix: Require the vendor to assume responsibility for their providers’ compliance with your data terms. If they won’t, treat it as a flag and either price it in or walk.
3. Unbounded “fair use” rate limits
The contract specifies a price per seat or per API call, with “reasonable use” caveats. The vendor reserves the right to define reasonable in their sole discretion. In practice this means your costs are unbounded.
Fix: Replace “reasonable use” with explicit numeric caps and overage rates. If the vendor refuses to specify, treat their pricing as a fiction.
4. The “model substitution” clause
The vendor can substitute the underlying model at their discretion, with notice. This is often presented as a feature (“we’ll automatically move you to the best model”). It is also a one-way door for cost increases. The next model can be priced 3x higher.
Fix: Require notice + opt-out for any substitution that changes the per-token cost or the model family. Lock in the current model as the minimum SLA.
5. Indemnification gaps for “hallucinated output”
Vendors almost universally disclaim liability for outputs the model generates. This is reasonable in narrow form and outrageous in broad form. The version we won’t sign is “vendor is not liable for any harm arising from generated content, including but not limited to defamation, IP infringement, regulatory violation.”
Fix: Carve back. Vendor remains liable for IP infringement on training data the vendor selected (large vendors carry insurance for this; small vendors should disclose). Vendor is liable for breaches of their specified data-handling commitments even if those breaches manifest as bad outputs.
6. The data deletion deadline you can’t hit
“Customer data will be deleted within X days of contract termination.” X is sometimes 90 days, sometimes 180, sometimes silent. The number is meaningless if the data is embedded in a fine-tuned model or an embedding index. Deletion of the source row does not delete the model artifact.
Fix: Require deletion of all derived artifacts (fine-tuned weights, embeddings, indexes) on the same timeline, with attestation. If the vendor cannot delete derived artifacts, you must understand what that means before signing.
7. The “audit rights” clause you’ll never use
The contract grants you audit rights “upon 30 days written notice, no more than annually, at the customer’s expense, during business hours.” Translation: you cannot meaningfully audit anything. Real audit happens after incidents, not on schedules.
Fix: Add a “for cause” audit right that triggers on any data incident, with vendor cost-share. Require SOC 2 Type II reports annually as a substitute for routine audits.
8. The DPA that isn’t really a DPA
The vendor offers a “Data Processing Agreement” that’s actually their service terms with a cover page. It doesn’t name sub-processors, doesn’t commit to data-residency, doesn’t specify breach-notice timelines.
Fix: Require a real DPA aligned to GDPR Art. 28 (even if you’re not in EU — it’s the cleanest reference frame). Sub-processor list, breach notice in 72 hours, data residency by region.
9. Auto-renewal at undisclosed rates
Annual contract auto-renews. Price for the renewal year is “at vendor’s then-current rates.” You discover the new rate when you get the invoice.
Fix: Cap auto-renewal at CPI + 5%, with the right to renegotiate or exit if the increase exceeds the cap. No vendor that wants the business will refuse this.
10. The “innovation roadmap” promise that’s not enforceable
The vendor’s sales deck promised features X, Y, Z by Q3. The contract specifies a different set of features, none of which include X, Y, Z. By Q3, X is now an upgrade, Y is “deprioritized,” and Z requires their enterprise tier.
Fix: Either get the roadmap commitments into the contract or accept that they are sales theater. Don’t price the deal as if they’re real.
11. The portability clause that’s data export only
“Customer can export their data in standard formats.” Standard for the vendor, often not standard for the next vendor. Workflow definitions, prompt templates, fine-tunes, custom integrations — none of these is “data” in the export sense.
Fix: Require export of workflow artifacts, prompt templates, and any custom configuration in human-readable form. If the vendor can’t define what’s exportable, that’s the answer.
12. The exclusivity ratchet
Buried in the discount terms: “to maintain volume pricing, customer agrees not to use comparable services from named competitors for the duration of the agreement.” This is rare in SaaS and increasingly common in AI procurement.
Fix: Refuse. Discount or no discount, exclusivity in a fast-moving market is too expensive. If you can’t refuse it, you’re paying for the discount in optionality you no longer have.
How to actually negotiate these
A few tactical notes from our practice:
- Ask for the redlined version first. Don’t ask the vendor whether they “support” your terms. Ask them to redline the actual paper. Their willingness to redline is the signal.
- Negotiate the data clauses before the price. Vendors will sometimes flex on data terms when they realize you won’t sign without them. They almost never flex on data terms once you’ve signed off on price.
- Use the smallest vendor that has the capability. Large vendors have less flexibility on contract terms in the mid-market. Smaller vendors will redline more aggressively, which is sometimes better for you.
- Get an MSA review from a lawyer who has done AI deals before. General commercial counsel will miss most of these. Specialized counsel will catch them. The fee is small relative to the cost of getting one of them wrong.
What this saves you
In our practice, applying these twelve fixes to a typical mid-market AI procurement saves between 8% and 30% on the deal’s expected total cost over 24 months — almost all of it through avoided lock-in, avoided price escalation, and avoided cleanup work when something goes wrong. The legal cost of doing this well is in the low five figures. The cost of not doing it is in the six figures, just not immediately.
The framing
AI vendor contracts in 2026 are written by vendors who know more about the market than your procurement team does. The negotiation is asymmetric and the contracts reflect that. The twelve clauses above are the most common ways the asymmetry shows up. Closing them isn’t paranoia — it’s the same baseline diligence you’d run on any contract that touches your data and your operations. The fact that nobody is running it on AI deals yet is exactly why the deals end up looking like this.
Fractional CAIO engagements include vendor-contract review as part of the role. If you have an AI deal in flight and want a second pair of eyes, schedule a call.
