A no training clause with indefinite retention is a promise with your data still in the room
An enterprise AI contract is a data governance document. It decides who may train on your inputs, where the data sits, how long it stays, and who owns what the model produces. Those five terms are usually negotiated as a checklist of independent items, and they are not independent: one of them determines whether the other four mean anything.
Prepared by Redress Compliance · August 10, 2026 · GenAI advisory. Based on 20 to 30 enterprise AI contract reviews, 2024 to 2025.
Executive summary
Default or standard tiers allowed input reuse for model training in 50 to 70 percent of first drafts.
Enterprise tiers usually disable it, but the setting is a selection rather than an automatic property of buying at enterprise scale, which means the common assumption that the major vendors are enterprise safe by default is wrong in the specific way that matters.
The commitment has to be a clause in the agreement rather than a statement on a product page, because only the agreement binds.
Residency, retention, and output ownership were silent until the buyer raised them.
Silence is not neutrality: unspecified residency means the vendor chooses the region, unspecified retention means the storage window is open ended, and unspecified ownership leaves copyright in machine generated work resting on unsettled law.
So the honest description of a standard order form is not an enterprise agreement with gaps, it is a set of consumer defaults printed on enterprise paper.
Retention is the clause that makes the other four enforceable, and it is the one most often skipped. A no training commitment with an open ended retention window leaves your inputs available to a future policy change, a change of ownership, or a subpoena.
A named residency region without a deletion window means the data sits in that region indefinitely rather than briefly. Output ownership matters less if the inputs that produced it are retained and reusable. Fix retention and the other four terms acquire a boundary in time.
Close the governance terms before the commercial conversation, not alongside it. Our AI procurement work finds that data terms settled after price are settled without leverage, because the discount has already been conceded and reopening the paper costs goodwill the buyer no longer has.
A discount agreed against unresolved governance terms is a discount you pay for twice: once in the concession you traded for it, and again in the term you accept to avoid reopening the deal.
The five terms, and what silence means on each
| Area | Default risk | Buyer side fix |
|---|---|---|
| Training use | Inputs reused to improve the model | An explicit no training clause bound in the agreement |
| Residency | The vendor chooses the region | Named regions written into the order |
| Retention | Open ended storage of prompts and outputs | A fixed deletion window |
| Output ownership | Ambiguous, resting on unsettled law | Customer ownership stated explicitly |
| Indemnity | Silent, with the customer carrying output risk | Vendor indemnity on third party IP claims, with the cap read |
Read the five as a dependency chain rather than a checklist and the priority order changes.
Retention sits underneath the others because it decides how long the question stays open: a no training commitment governs what may be done with data that is still held, and an open ended hold means the commitment has to survive every future change of policy, ownership.
And jurisdiction rather than merely the current one.
Residency has the same relationship to it, since naming a region without naming a deletion window fixes where the data lives without limiting how long it lives there.
Output ownership is the weakest of the five on its own, because ownership of a result matters less when the inputs that produced it remain available and reusable. Fix retention first and the other four acquire an end date. The wider category framework sits in the AI procurement framework.
Why governance closes before price
- Settle the five terms before discount talks open, because a term negotiated after the price is negotiated without leverage and reopening the paper costs goodwill you have already spent.
- Find the clause, not the page. A no training commitment on a help article or a blog post is not a contractual commitment, and the two are routinely conflated during evaluation.
- Confirm the clause applies to the tier you are actually buying, since the protection is a property of the tier and the selection rather than of being an enterprise customer.
- Bind it in the master agreement rather than in a referenced policy the vendor can revise unilaterally, which is the same amendment risk that governs every other unilateral vendor document.
- Read the indemnity cap and its conditions, because an indemnity with a low cap or a long list of exclusions is a term that reads as protection and functions as a disclosure.
The enterprise AI procurement strategy brief
The five categories, the governance clauses that decide whether a deal is safe to sign, and the commitment arithmetic behind the commercial terms.
Get the white paper →What the pattern across reviews actually shows
The common advice is that the major AI vendors are enterprise safe by default, so a standard order form is fine to sign. The review file contradicts that in a specific and useful way.
It is not that the vendors are careless or that enterprise protections do not exist: they exist, and they are usually available, and they are not applied unless the buyer selects the tier and binds the clause.
Input reuse for training was permitted in half to two thirds of first drafts, and residency, retention, and output ownership were simply absent until raised.
Absence is the pattern rather than adversarial drafting, and it has a straightforward commercial logic: a term nobody asks for is a term nobody has to give.
That reframes the buyer side task from detecting hostile language to noticing what is missing, which is a harder review discipline because a silent contract reads as clean. It also explains why sequencing matters more here than in most software categories.
In a seat based negotiation the commercial and legal tracks can run in parallel because the terms and the price are largely independent.
In an AI agreement the governance terms are the product characteristics, so conceding price first removes the leverage needed to change what you are actually buying.
Close training use, residency, retention, output ownership, and indemnity, in that order of dependency, and only then discuss the rate. Vendor assurances outside the contract are marketing, and what is not written is not enforceable.
The Microsoft platform terms sit in the Azure OpenAI guide, and the AWS equivalent in the Bedrock pricing guide.
- Percentile standing for your exact deal size and industry, from real closed transactions
- Scenario simulation before the call: test alternative terms and see the financial impact of each
- A negotiation playbook, talking points, and a two page executive brief on day one
What we saw across enterprise AI contract reviews, 2024 to 2025
Across roughly 20 to 30 enterprise AI contract reviews supported between 2024 and 2025, the default terms favoured the vendor on training and intellectual property in most first drafts:
First drafts where default or standard tiers allowed prompts and outputs to be reused for model training.
Residency, retention, and output ownership were silent rather than adverse, which is a harder review problem because a silent contract reads as clean.
Three patterns recurred: default or standard tiers allowing input reuse for training in 50 to 70 percent of cases, data residency unspecified or vendor chosen unless the buyer raised it, and output ownership and indemnity silent until the buyer side negotiated them in.
The buyer side move is to treat the order form as an opening position, review for what is missing rather than only for what is adverse, fix retention first because it bounds the other four in time, and close all five before the commercial conversation opens.
An assurance outside the contract is marketing. Only the contract binds.
Your first five moves
- List every AI tool in use and the tier each one is bought on, because the protections are properties of the tier and the selection rather than of being an enterprise customer.
- Fix retention first with a defined deletion window, since an open ended hold leaves a no training commitment dependent on every future policy, ownership, and jurisdiction change.
- Locate the training use clause in each contract and confirm it applies to your tier and is bound in the master agreement rather than in a policy the vendor can revise alone.
- Name the required data regions and state customer ownership of outputs explicitly, because silence on either point resolves in the vendor's favour by default rather than neutrally.
- Require an IP indemnity and read its cap and conditions, then block signature until all five terms are agreed and before any discount is conceded. The GenAI practice runs the review with you.
Frequently asked questions
Can AI vendors train on enterprise data?
Only if the contract allows it, and in our reviews default or standard tiers permitted input reuse in 50 to 70 percent of first drafts. Enterprise tiers usually disable it, but that is a selection you make and bind rather than an automatic property of buying at enterprise scale.
Verify the clause in the agreement, not the marketing page.
Why is retention the most important clause?
Because it bounds the other four in time. A no training commitment governs what may be done with data still held, so an open ended retention window means that commitment must survive every future policy change, ownership change, and legal process.
Naming a residency region without a deletion window fixes where data lives without limiting how long.
What does silence mean in an AI contract?
It resolves in the vendor's favour rather than neutrally. Unspecified residency means the vendor chooses the region, unspecified retention means storage is open ended, and unspecified ownership leaves copyright in machine generated work resting on unsettled law.
A silent contract reads as clean, which makes reviewing for absence harder than reviewing for adverse terms.
Is a vendor blog or help page commitment binding?
No. Only the contract binds. A no training statement on a help article can be revised unilaterally and without notice, which is the same amendment risk that governs any policy document a vendor controls.
Find the clause, confirm it covers your tier, and reference it in the master agreement rather than in a linked policy.
Who owns the output an AI model produces?
Most enterprise terms assign output to the customer, but copyright in machine generated work is unsettled, which is why an explicit ownership clause matters.
It is also the weakest of the five terms on its own, because owning a result matters less when the inputs that produced it remain retained and reusable under an open ended window.
Should we require an IP indemnity?
Yes, and then read the cap and the conditions. Several vendors now offer indemnity against third party IP claims on outputs for enterprise tiers, but an indemnity with a low cap or an extensive exclusion list reads as protection and functions as a disclosure.
Make it explicit and check what it actually covers before relying on it.
When should governance terms be negotiated?
Before the commercial conversation. In an AI agreement the governance terms are the product characteristics rather than legal boilerplate, so conceding price first removes the leverage needed to change what you are actually buying.
A discount agreed against unresolved governance terms is paid for twice: once in the concession, once in the term you then accept.
How should an AI vendor review be structured?
Around a recognised control set so it is repeatable rather than personality dependent, paired with the applicable regulatory risk tiers for regulated use cases.
The practical addition is a review discipline that checks for missing terms as deliberately as it checks for adverse ones, since absence was the dominant pattern in our file.
Signing the Enterprise Agreement
Part 3 of the Negotiating Anthropic series. What the agreement actually has to cover: the commitment and its shape, the rate card, data and training terms, model deprecation, capacity, and what happens if you under consume.
