The Fine Print Is the Product: Five AI Contract Traps to Fix Before Renewal
Photo: The Société Commerciale Française de l'Indochine., Public domain, via Wikimedia Commons
AI contracts are not ordinary software contracts. The product changes after signature. The model can be updated or retired. The vendor's data practices can shift. The price is often metered in units that are difficult to forecast. Renewal is the moment to fix the terms that create operational and financial risk.
The five red flags below are not theoretical. They appear in enterprise AI agreements in various forms, and they tend to surface after the pilot has become a production dependency. The goal is not to negotiate a perfect contract. The goal is to avoid the clauses that leave the customer with no leverage when the vendor changes course.
1. The model can change under you
A common clause gives the vendor the right to update, replace, or discontinue model versions at its discretion. The language often says the replacement will be 'materially similar' without defining what that means. For a customer that has tuned prompts, workflows, and downstream systems around a specific model, a change can break accuracy, latency, or cost assumptions.
Ask three questions. How much notice will we receive before a model is deprecated? Can we pin a model version for the term? What evidence will you provide that a replacement performs comparably on our use case?
Negotiate a notice period of at least 90 days for major changes and 180 days for deprecation. Require support for the prior model version for a defined period, such as 12 months. Add a right to terminate for material performance degradation, with a clear definition of degradation based on agreed benchmarks. Avoid 'materially similar' unless it is tied to measurable criteria.
2. Your data may train the model
Many consumer AI terms allow the vendor to use inputs and outputs to improve its services. Enterprise agreements often include a similar license unless the customer negotiates an opt-out. The clause may be buried in a definition of 'customer data' or in a subprocessor list.
Ask whether training on customer data is the default or the exception. What is the retention period? Can we get deletion certification? Which subprocessors touch the data? Will we receive notice before a new subprocessor is added?
Negotiate a prohibition on training on customer data without explicit written consent. Require zero or short retention, deletion within 30 days of termination, and a current subprocessor list with notice of changes. Add audit rights or third-party attestations. If the vendor insists on using data for model improvement, limit it to aggregated, de-identified data and require an indemnity for unauthorized use.
3. Indemnity stops at the model boundary
AI contracts often include intellectual property indemnity for the software itself but exclude outputs. The vendor may argue that it cannot control what the customer prompts. That is fair up to a point. But if the model generates infringing content in ordinary use, the customer should not bear the entire risk.
Ask who owns the output. What happens if the output is alleged to infringe a third-party right? Does the indemnity cover the model's training data? Are there exclusions for prompts or modifications?
Negotiate mutual indemnity for IP infringement. Require the vendor to defend claims arising from its model and training data, not just its software. Carve out customer misuse, but do not accept a 'sole control' exclusion that gives the vendor a broad escape hatch. Check the liability cap. An indemnity that is capped at a small fraction of fees is not meaningful for a high-value use case.
4. Usage caps and escalators hide the real cost
AI pricing is often metered by tokens, API calls, seats, or compute minutes. The contract may include a minimum commitment, an overage rate, and an annual escalator. The problem is that usage is difficult to forecast, especially when a pilot becomes a production system.
Ask how the unit is defined. What happens if we exceed the cap? Is overage billed retroactively? Can we burst without penalty? What is the cap on annual price increases? Does the price change if the vendor moves us to a different model tier?
Negotiate a burst allowance or a grace period before overage charges apply. Cap annual increases at a reasonable percentage, such as 5% or the consumer price index, whichever is lower. Require notice before a price change and a right to terminate if the increase exceeds an agreed threshold. Make sure true-ups are prospective, not retroactive. If the vendor will not budge on price, ask for credits or a longer commitment in exchange.
5. Exit is theoretical
Many AI contracts promise data export in a 'standard format' and transition assistance at the vendor's discretion. In practice, the customer may discover that the export excludes prompt libraries, fine-tuned weights, evaluation data, or logs. The vendor may charge professional services fees to help with migration. The clock may run out before the data is usable.
Ask what data we can export and in what format. Does it include prompts, outputs, fine-tuning artifacts, and audit logs? How long will you provide transition assistance? What will it cost? Will you certify deletion after termination?
Negotiate a machine-readable export within 30 days of request, at no additional cost. Require transition assistance for at least 90 days at capped fees. Include deletion certification. If the customer has invested in fine-tuning or custom model artifacts, negotiate ownership and escrow terms. Do not accept a transition clause that depends on the vendor's 'commercially reasonable efforts' without a defined scope.
A negotiation sequence
Start 120 days before renewal. Inventory every AI tool in use, including shadow deployments. Map data flows and identify the contracts that cover them. At 90 days, build a data map that shows which systems depend on which model versions. At 60 days, create leverage: competitive bids, internal alternatives, or a credible plan to reduce scope. At 30 days, push for signature.
Know when to walk. If the vendor will not offer a training opt-out, deprecation notice, a defined export process, or a meaningful indemnity, the risk may outweigh the convenience. The best contract is not the one with the lowest price. It is the one that still works when the model changes, the data practices shift, and the renewal notice arrives.