Who this is for
This is for the person who has to sign the AI contract and will catch the blame if it goes wrong. Often that's a CTO, an IT lead, or a finance or operations manager who got handed a proposal and a deadline. The vendor's deck looks fine. Your job is to find the gaps before the ink dries.
The questions here are about the deal itself, not the model. They cover what happens to your data, how the price moves, who owns the output, and how you get out. Those are the things that cause trouble a year in, long after the demo.
You don't need to be an AI expert to ask them. You need to be the person who reads the contract slowly and asks the vendor to repeat themselves. If an answer sounds vague, that's information. Write it down.
Use this alongside your normal procurement and security review, not instead of it. The point is to surface the AI-specific risks that a standard software contract won't mention.
Before you begin
Get the vendor to name the exact model and hosting setup in writing. "We use a leading AI model" tells you nothing. You want the provider, the model name or version, and where it runs. Models get deprecated and swapped, so ask what notice you get when that happens and whether you can approve the change first.
Pull together your own facts before the call. Which data will the system touch? Does any of it include personal data, health details, financial records or trade secrets? Who inside your company will use it, and what decisions will they make from its output? A vendor answer only means something against your own context.
Decide your tolerance for error before you hear the sales pitch. If the tool drafts internal notes, a wrong answer is annoying. If it feeds a credit decision or a medical workflow, a wrong answer is serious. That difference changes every question you ask about accuracy and human review.
Bring the person who will actually run the thing day to day. They will spot operational gaps that a manager misses, like what happens when the vendor's service is down or an answer looks wrong.
The checklist
Ask these in a live call and ask for the answers in writing afterward. A vendor who is comfortable with scrutiny will send a clear summary. One who pushes back on written answers is telling you something.
Group the questions by risk. Data and privacy first, then accuracy and oversight, then money, then exit. Don't let a friendly account manager rush you past the data questions to get to the pricing slide.
Some answers will be "it depends." That's fine if the vendor then explains what it depends on. What you're listening for is whether they know their own system or are reciting a script.
- Which exact model and hosting provider will run our workload, and what notice do we get before you switch models or versions?
- Where is our data stored and processed, and does any of it leave our region or get used to train anyone's model?
- What is your data retention policy, and can we set our own retention period and delete data on demand?
- How do you measure accuracy on our actual data, and will you share the test set and results before we sign?
- What human review steps does the system support, and can we require sign-off before an output is used?
- How do you handle a wrong or harmful output, including who is responsible and what the response time is?
- What is the full cost structure: setup, per-seat, per-token or per-call fees, storage, support and overage charges?
- What usage limits apply, and what happens to price and service if our volume grows well past the estimate?
- What uptime do you commit to, and what service credits apply if you miss it?
- Who owns the outputs, the prompts, the fine-tuned models and any data we supply, and what happens to them when the contract ends?
- How do we export our data and configuration if we leave, in what format, and at what cost?
- What security testing do you do, and will you share the results and your incident notification process?
- Which subcontractors or third-party providers touch our data, and do we get notice before you add more?
- What does support actually cover, what are the response times, and is there a named person we can reach?
What good looks like
A good vendor answers the model and hosting question without flinching. They can tell you the provider and version, and they have a written process for model changes, including notice and your right to approve. If they dodge this, expect surprises later.
On data, good answers are specific. They name where processing happens, state clearly whether your data trains models, and can show you the retention and deletion controls in the product, not just in a policy document. They should be able to say what their subcontractors do and give you notice before adding new ones.
On accuracy, look for honesty about limits. A serious vendor will talk about where the system fails and how they test it, and will share results on data like yours. They'll support human review and describe how they handle bad outputs. Vague claims of high accuracy with no evidence are a warning.
On money and exit, good looks like a price you can model at low and high volume, clear overage terms, and a clean export path. They should be relaxed about you leaving. A vendor who makes exit hard is often hiding something about lock-in.
What to do next
Score the answers against your own risk tolerance, not the vendor's confidence. For each checklist item, mark it clear, vague or missing. Vague and missing items go into the contract as open points before you sign, or they become assumptions you can't enforce later.
Ask for a short paid pilot or a scoped trial on your own data before committing to a full term. Use it to test the accuracy claims, the support response and the export process. What you learn in two weeks beats what you learn from a longer sales cycle.
If you're still deciding what to build and want an independent view before you sign anything, the Cloudverg AI readiness assessment is $5,500/month, runs for two weeks, and ranks use cases, checks your data and produces a cost model and delivery plan. It's one option among others, including doing the scoping in house or hiring a specialist.
Whichever route you take, get the final answers in writing and attach them to the contract. That turns a friendly conversation into something you can point to when the model changes, the bill jumps or you decide to leave.
Frequently asked questions
What should I ask an AI vendor about my data?
Ask where data is stored and processed, whether it is used to train models, what the retention and deletion rules are, and which subcontractors touch it. Get the answers in writing and check that the product actually has the controls the vendor describes.
How do I check an AI vendor's accuracy claims?
Ask them to test on data like yours and share the test set and results before you sign. Look for honest talk about failure cases and human review steps. Treat unsupported accuracy claims as a warning sign.
What contract terms matter most with an AI vendor?
Model change notice, data ownership and export, the full cost structure including overage, uptime commitments and exit terms. These are the items that cause problems later, so pin them down before signing.