AI AI at work: policies, tools and what to watch
AI vendor contracts: the clauses to check before you sign
What to look for in an AI supplier agreement: training rights over your data, retention, subprocessors, uptime, indemnities for output and exit arrangements.
The short answer
- The single most important clause says whether the vendor may train its models on what you put in, and a business plan usually says no while the free tier usually says yes.
- Most of the terms that matter are not in the document you sign but on policy pages it links to, and those pages can change without telling you.
- Ask for a retention figure in days, a list of subprocessors, and the countries your data is processed in, in writing, before money changes hands.
- A model your workflow depends on can be retired, so look for a notice period on model changes and deprecations, usually 6 to 12 months for business plans.
- If you are in or selling into the EU, duties from the AI Act flow down to you as the deployer, so the contract should say the vendor will give you what you need to meet them.
Before you sign an AI contract, find the answers to seven questions: can the vendor train on your data, how long do they keep it, who else touches it and where, who owns the output, will they defend you if the output infringes someone's copyright, what happens when they change or retire the model, and how do you get your data out. Everything else in the agreement is ordinary software contracting. Those seven are where AI deals differ from the rest, and they are where small buyers get hurt. You do not need a lawyer to check them. You need about two hours and a habit of searching the document for the right words.
What you are actually signing
An AI purchase is rarely one document. There is usually an order form setting price and plan, terms of service, a data processing agreement (a DPA, the contract that governs personal data), an acceptable use policy, and a trust or security page. The order form incorporates the rest by reference, which means the linked pages are part of your deal even though you never opened them.
That structure matters, because pages the vendor hosts can be edited. A term you relied on in March may read differently in November. So when a promise matters to you, save a dated copy of the page as a PDF on the day you sign, and look for a clause saying material changes come with advance notice. Thirty days is common, and anything that lets the vendor change terms effective immediately is worth questioning.
Your role has a name too. If you decide what personal data goes in and why, you are the controller and the vendor is your processor. That is not vanity vocabulary: it decides who answers a customer's deletion request, who reports a breach, and who carries the duty if a regulator asks. Your workplace AI policy should match what the contract says here, because a policy promising staff more protection than the contract delivers is worse than none.
The clauses to find first
| Clause | What good looks like | What should slow you down |
|---|---|---|
| Training on your inputs | Explicit: no training on customer content, on by default | "We may use content to improve our services" |
| Retention | A number in days, with zero retention available | "As long as necessary for business purposes" |
| Subprocessors | Published list, notice before additions, right to object | No list, or "affiliates and partners" |
| Data location | Named regions, with an option to pin processing | Silence, or "worldwide" |
| Output ownership | You own outputs, vendor claims no rights | Vendor takes a broad license to your outputs |
| IP indemnity | Vendor defends and pays copyright claims on outputs | No indemnity, or one capped at one month of fees |
| Model changes | Notice period before a model is changed or retired | Vendor may modify the service at any time |
| Exit | Export in a usable format, deletion certified on request | Data deleted immediately on termination, no export |
| Liability cap | Twelve months of fees, with data breach carved out | One month of fees, covering everything |
Search the terms for the words "train", "improve", "retain", "subprocessor", "indemnif", "terminate" and "export". Those seven searches find most of the table in a few minutes.
Training, retention and deletion
Training rights are the clause people get wrong most often, usually by reading the vendor's general position rather than the one attached to the plan they bought. The pattern across the industry is consistent: consumer and free tiers often reserve the right to use your conversations to improve models, while business, team and enterprise tiers typically do not. Moving your staff from personal free accounts onto a paid business tenant is therefore a data control measure, not just a billing tidy up. What happens to text after you send it is covered in more depth in where your prompts are processed.
Retention is separate from training and easy to confuse with it. A vendor that never trains on your content may still store it for 30 days for abuse monitoring, which means it exists on their systems, is reachable by their staff under some process, and is discoverable in a dispute. Ask for the figure in days, whether zero retention is available on your plan, what you lose by turning it on (usually chat history), and whether human reviewers can read flagged content.
Outputs, indemnities and who pays when it goes wrong
Two questions live here. Who owns what the model produces, and who pays if it turns out to resemble someone else's work.
Ownership is usually generous: most vendors assign outputs to you and claim nothing. Read it anyway, and watch for a license back that lets the vendor reuse your outputs. The harder question is infringement, because generated text, images and code can reproduce protected material and the person who published it is the obvious defendant. Larger vendors now commonly offer an IP indemnity: if a third party sues you over an output, they defend the claim and pay the damages. These indemnities always carry conditions, typically a paid tier, safety filters left on, no prompting the model toward a specific copyrighted work, and prompt notice of a claim. An indemnity you void by pasting in a competitor's brochure and asking for something similar is not protection.
Check the liability cap too, because an indemnity sitting above a cap of one month of fees is close to symbolic. Twelve months of fees is the normal ceiling for small contracts, with data protection breaches often carved out to a higher figure.
Model changes, service levels and exit
The clause that surprises buyers most is deprecation. The model behind your workflow is a product, and products are retired. When it goes, prompts tuned over months behave differently, output formats shift, and anything built on top needs retesting. Business agreements increasingly promise a notice window before a model version is withdrawn, often 6 to 12 months. If nothing in your contract addresses this, ask, and write the answer down.
Service levels matter less than people expect, because the credit for an outage is usually a small percentage of a monthly fee. The practical question is what your staff do for the day. Assume a few hours of downtime a year and keep a manual fallback for anything customer facing.
Exit is the last check. You want an export in a format something else can read, a window in which to take it (30 to 90 days is typical), and deletion after that, certified in writing if you ask. The realistic lock in is rarely the raw data anyway. It is the prompt library, the integrations and the habits, so keep your prompts in a document you control rather than only inside the vendor's interface.
A review you can run before you sign
- Save the terms, DPA and trust page as dated PDFs.
- Run the seven word searches above and fill in the table for your plan.
- Email the vendor the gaps as numbered questions and keep the reply.
- Decide what may never be entered into the tool, given what you found, and write that into your internal rules.
- If you operate in the EU or sell there, ask which AI Act duties the vendor expects you to carry as the deployer and what documentation they provide. Systems used for hiring or staff evaluation are treated as high risk under the AI Act, and the obligations reach the organization using the tool, not only the company that built it.
- Diary a date six months out to recheck the pages you saved.
What to check first
If you only do one thing today, open the plan you are actually on and find the sentence about training. Then find the retention period. Those two lines decide whether the tool is safe for client work or only for public information, and they are the two that most often differ from what everyone in the office assumes.
Common questions
We are a five person company. Do we really need to read all this?
Yes, but it is a short read, not a legal project. The clauses below are findable with a text search of the terms, and the questions you cannot answer from the page are the ones to email the vendor. Keep their reply, because a written answer from sales is evidence even when it is not a contract term.
Does the paid plan really stop them training on our data?
On most major business and enterprise plans, yes, and it is normally stated plainly in the terms or the trust page. Free and consumer plans frequently reserve the right to use your content to improve the service, sometimes with an opt out buried in settings. Check the exact plan you are buying, not the vendor's general marketing.
What is a subprocessor and why should I care?
A subprocessor is another company the vendor uses to deliver the service, such as a cloud host, a moderation provider or a support desk. Your data reaches them too, so their locations and security posture become part of your risk. Good vendors publish a current list and promise notice before adding one.
The vendor will not negotiate. Is that the end of it?
Standard terms are common below a certain contract size and refusing to change them is normal, not a red flag on its own. What you can still do is choose the plan whose published terms suit you, restrict what data you put in, and record the gap so nobody later assumes protection you never bought.