GenAI Practice
OpenAI agreement. Seven clauses to push back on.
OpenAI enterprise agreements move fast and favor the vendor on data, capacity, and price. Seven clauses decide whether the deal protects you. Push back on these before signing.
OpenAI enterprise agreements move fast and favor the vendor on data, capacity, and price. Seven clauses decide whether the deal protects you. Push back on these before signing.
OpenAI enterprise agreements are drafted to favor the vendor on data use, capacity, and price escalation. Seven clauses carry most of the buyer risk. This guide gives the push back language and the reasoning behind each one.
Start here. The data clause decides whether your prompts and outputs can be used to improve the model. Require explicit contractual confirmation that they cannot.
OpenAI publishes an enterprise privacy position, but the executed contract is what binds. Hold the signed terms to the same standard as the public page.
Require language that your business data is excluded from model training by default. Make it a representation, not a setting that can change.
Set a maximum retention window and a deletion right on termination, in line with the data processing addendum. Confirm logs and fine tuning artifacts are covered, not only prompts.
Capacity is where production workloads break. Best effort throughput is not a commitment. Push for committed capacity tied to your use.
Negotiate guaranteed tokens or requests per minute, with a remedy if the vendor misses it. Reference the published pricing and rate structure as the floor, then commit it.
A deprecated model can break a live pipeline. Require advance notice and a migration window before any model is retired.
Seven OpenAI clauses and the buyer side position
| Clause | Vendor default | Buyer side position |
|---|---|---|
| Data and training | Ambiguous | Contractual exclusion from training |
| Retention | Vendor set | Capped window plus deletion right |
| Capacity | Best effort | Committed throughput with remedy |
| Model deprecation | Vendor discretion | Notice plus migration window |
| Price protection | Absent | Renewal increase cap in writing |
| Output indemnity | Customer bears | Vendor indemnifies for model output |
| Exit | Limited | Export plus certified deletion |
Headline price is not the cost. Escalation, minimums, and overage decide what you actually pay across the term.
Cap the renewal increase in writing. Without a cap, a successful pilot becomes leverage the vendor uses against you at renewal.
Negotiate the overage rate before you need it. Align any committed minimum to a conservative usage forecast, not the optimistic one.
Regulated buyers cannot sign without an exit. The contract must let you leave with your data and without disruption.
Require full export in a usable format and certified deletion on termination. Tie deletion to a defined timeline.
Confirm output indemnity sits with the vendor, in line with the published business terms. Add a transition period so a switch does not strand a live workload.
The common advice is that the standard enterprise terms are fine because the vendor already promises not to train on business data. We disagree. In the reviews we ran, the marketing promise and the signed contract language did not always match, and capacity and price protection were almost never in the buyer's favor by default. The buyer side move is to treat every published assurance as a starting point and require it in the executed agreement, with committed throughput and a renewal cap written in. A promise on a web page is not a contractual right.
Source: Redress Compliance advisory engagement file, 2024 to 2025.
A model vendor will give you the capacity and the price you negotiate, not the capacity and the price you assume. Write both into the contract.
OpenAI states that it does not train on business data submitted through its enterprise and API products by default. Require that assurance in the executed contract as a representation, because the binding protection is the signed language, not the published page.
The data and training clause is first, because it governs whether your prompts and outputs can improve the model. Close behind are capacity commitments and price protection, which decide reliability and cost across the term.
Yes, through committed throughput rather than best effort limits. Negotiate guaranteed tokens or requests per minute with a remedy if the vendor misses the commitment, so production workloads are protected during demand spikes.
Add a written cap on the renewal increase before you sign. Without a cap, a successful pilot becomes leverage the vendor uses at renewal, and the headline price you negotiated does not hold.
Push for the vendor to indemnify for model output rather than the customer bearing that risk. Check the business terms, since indemnity scope varies and often excludes the cases buyers care about most.
A clean exit clause requires full data export in a usable format and certified deletion on termination within a defined window. Regulated buyers should treat this as non negotiable.
Rarely. The first draft typically favors the vendor on data, capacity, and price. Reviewing the seven clauses and requiring published assurances in writing closes the gaps the standard paper leaves open.
At the first draft, not at signature. Early redlining on data, capacity, and price gives the most leverage and avoids accepting vendor defaults under deadline pressure.
Enterprise AI contract review, data use posture, capacity protection, and the buyer side moves across the major model vendors.
Used across more than five hundred enterprise clients. Independent. Buyer side. Built for procurement leaders running the next renewal cycle.
Enterprise AI deals reward the buyer who reads past the data sheet. Fix the seven clauses, require every assurance in writing, and the contract finally protects the workload it runs.
