AI vendor contracts: vetting after the DeepSeek ban

AI vendor contracts were once signed on the strength of a demo and a data processing addendum. The DeepSeek federal ban changed that. Buyers now have to ask where the model runs, who can see the prompts, and what happens if a regulator moves against the vendor overnight. This guide sets out the checks to run before you sign an AI vendor contract, and the clauses to fix when you renew.

What the DeepSeek federal ban changed for buyers

The ban is the clearest signal yet that model origin is a procurement risk, not a technical detail. A tool that worked in a pilot can become unusable if the provider is restricted, and the contract decides who carries that cost. Buyers who treated AI like any other SaaS subscription are now reading the fine print.

The practical shift is simple. Every AI vendor contract needs a named model, a named hosting location, and a named exit route. If a supplier cannot state those three things in writing, the deal is not ready.

Start by mapping every AI tool already in use, including the ones bought on a team credit card. You cannot renegotiate what you cannot see.

The AI vendor contracts clauses that decide your risk

Most AI vendor contracts are standard SaaS paper with a short AI schedule bolted on. That schedule is where the risk sits. Read it line by line and check the following.

• Model and provider named. The contract should name the model and the underlying provider, not just the reseller. A clause that lets the vendor swap models at will needs a notice period and a right to object.
• Sub-processor list. Model providers are sub-processors. You need the current list, a duty to notify changes, and a right to terminate if a new sub-processor is unacceptable.
• Data use for training. Confirm in writing that your inputs and outputs are not used to train the model. Silence is not consent, and silence is not protection either.
• Restriction and suspension cover. Ask what happens if a government restricts the provider. Who notifies you, how fast, and who pays for migration?
• Audit and evidence. You need the right to see security certifications and to ask questions, not just a marketing page.

Take the model-swap clause first. A vendor that can change the engine behind your workflow without notice has effectively sold you an open-ended promise. Fix it with a change-control process: notice in writing, a defined objection window, and a termination right if you reject the change. Put that process in the AI vendor contract itself, not in a side email.

Data residency, training and the sub-processor chain

Data residency questions used to end at the region dropdown. With AI, they run through the whole chain: your application, the vendor's platform, the model host, and any inference provider behind it. Each hop is a place where data can leave your control.

Ask for a data flow diagram, not a policy page. Then check the AI vendor contract matches the diagram. If the diagram shows inference in one country and the contract is silent on location, you have a gap.

Training rights deserve the same treatment. Get a clause that says your data is not used to train or improve models, and that any aggregated telemetry is stripped of identifiers. Where the vendor wants usage data for product improvement, make it opt-in and documented.

Sub-processors are the weakest link. Model providers change terms often, and a restriction can arrive with little warning. A workable clause gives you the sub-processor list, notice of changes, and a termination right with a pro-rata refund if a change is material. That is the same discipline buyers apply to any IT services vendor directory shortlist: know who is actually doing the work.

Exit rights and portability when a model is pulled

The ban made one question urgent: if the model disappears tomorrow, what do you still own? The answer should be in the AI vendor contract before you sign.

Check four things:

1. 1Export rights. You must be able to export your data, prompts, configurations and outputs in a usable format, on request, without a fee.
2. 2Transition assistance. Ask for a defined period of help after termination, at a stated rate or at no cost.
3. 3Escrow or continuity. For business-critical tools, ask what happens if the vendor itself is restricted or wound down.
4. 4Deletion proof. Get written confirmation of deletion within a set period, including from sub-processors.

A termination right without an export right is a trap. You can leave, but you cannot take your work with you. Test the export before you sign by asking for a sample file from a reference customer or a sandbox.

Engagement model matters here too. A fixed-scope build with clear deliverables is easier to exit than an open-ended retainer with the vendor holding the keys. The trade-offs between fixed price, time and materials, and dedicated teams apply directly to how much lock-in you accept in the AI vendor contract.

How to run the diligence before signature

Diligence on AI vendors should look like diligence on any critical supplier, with extra weight on the model layer. Run it in a fixed order so nothing gets skipped under deadline pressure.

1. 1Inventory the use case. Write down what data goes in, what comes out, and who inside the business depends on it.
2. 2Request the AI schedule. Ask for the model name, hosting region, sub-processor list and training stance in one document.
3. 3Check the security evidence. Certifications, penetration test summaries and incident history. Ask when the last test was and who ran it.
4. 4Model the restriction scenario. Write a one-page plan: who calls the vendor, who informs the business, what the fallback is, and how long it takes.
5. 5Negotiate the gaps. Put the model-swap, training and exit clauses on the table early, while you still have leverage.
6. 6Record the decision. Keep the evidence pack with the signed AI vendor contract so the next review starts from facts.

Give the vendor a deadline for the evidence pack. A supplier that cannot produce a sub-processor list in a week will not produce one in an incident.

Red flags in AI supplier paperwork

Some clauses should stop a deal until they are fixed. Watch for these in any AI vendor contract.

• The vendor will not name the model provider, citing "proprietary technology".
• Training rights are described as "service improvement" with no detail.
• The sub-processor list is a link that can change without notice.
• There is no restriction or suspension clause at all.
• Export is offered only as a PDF or a screenshot.
• Liability for data breaches is capped at a token amount.
• The vendor's security page is the only evidence offered.

None of these is automatically fatal. Each one is a negotiating point, and each one tells you how the vendor behaves when things go wrong. Vendors who expect to win serious work prepare for these questions. The ones who win RFPs on a marketplace without racing to the bottom usually arrive with the answers ready.

What to put in the renewal checklist

AI vendor contracts signed before the ban need a refresh. Build a short checklist and run it at every renewal, and whenever a provider's status changes.

• Confirm the model and provider named in the schedule still match reality.
• Re-check the sub-processor list against the data flow diagram.
• Re-test the export and confirm the format still works.
• Review incident logs and any security certification renewals.
• Update the restriction plan with current contacts and fallback options.
• Check that training and telemetry clauses still match what the product does.

Put an owner and a date against each line. A checklist without an owner is a wish. If the review turns up gaps you cannot fix with the current supplier, start looking at alternatives in the vendor directory and on the open RFPs list.

Frequently asked questions

What should an AI vendor contract say about the model provider?

It should name the model and the underlying provider, not just the reseller. It should also require notice before any model change and give you a right to object or terminate. If the provider is restricted, the contract should say who notifies you and how fast.

Can I stop an AI vendor from training on my data?

Yes, if the contract says so in plain terms. Ask for a clause that bars training and model improvement on your inputs and outputs, and that strips identifiers from any telemetry. Silence in the contract is not protection.

What exit rights do I need if a model is pulled?

You need export of data, prompts, configurations and outputs in a usable format, at no fee. Add transition assistance for a defined period and written deletion confirmation from the vendor and its sub-processors. A termination right without export rights leaves you stuck.

How often should I review AI supplier contracts?

At every renewal, and immediately when a provider's regulatory status changes. Run the same checklist each time: model named, sub-processor list current, export tested, restriction plan updated. Keep the evidence pack with the signed contract.

What this means for buyers

The DeepSeek ban is a warning about concentration. Any single model provider can become unavailable for reasons outside your control, and the AI vendor contract is the only lever you hold when that happens.

Buyers who name the model, control the sub-processor chain, protect training rights and test the exit will keep operating through the next restriction. Those who signed a generic SaaS agreement will be negotiating from a standing start, with live systems and no fallback.

Start with the tools already in production. Pull the contracts, run the checklist, and put the gaps on the table at the next renewal. If you are sourcing a new supplier, describe the project and the contract terms you need on Scopelyst before you compare proposals, so every vendor answers the same questions about the AI vendor contract.