May 3, 2026
What to ask an AI vendor before you sign (the short list)
A practical list of questions about your data, accuracy, and exit—without needing a computer science degree.
If you have ever sat through an AI demo, you know the pattern: fast answers, slick slides, and a price sheet at the end. The hard part is knowing what to verify before anyone commits budget or customer trust.
This post is a companion to our vendor checklist guide. Use both before your next renewal or pilot.
Before the meeting
Send your messiest realistic example (redacted) and ask them to prepare a live run—not a canned video. If they push back, note it. The goal is to see how the product behaves when the world is not perfect.
Agree internally on non-negotiables: regions, SSO, logging, training opt-out, or whatever your industry demands. One aligned list keeps the meeting from becoming a feature beauty contest.
1. Training and your content
Ask in plain language: Will anything we type into this tool be used to train your models?
If the answer is long and fuzzy, slow down. You want a clear yes or no, plus anything that costs extra to turn off. This matters most when staff paste emails, tickets, or contracts into the product.
Follow up with: How do we delete or purge a mistaken paste? Incidents happen; recovery paths matter.
2. Where your data lives
You do not need to become a cloud expert. You do need to know which country or region the data is processed in, and whether that matches your contracts or industry rules.
If the vendor cannot point to documentation, assume you will be the one explaining gaps to your customers later.
3. What “good” looks like
Demos are curated. Bring a messy, real example from your world: a long support thread, a proposal with odd formatting, or a report that mixes tables and notes.
Watch whether the tool admits uncertainty, cites sources, or hands off to a human when it should. Confidence without evidence is a liability.
Ask what happens when the model is updated. Do you get notice? Can you pin a version for a regulated workflow?
4. Logs and access
Ask who can see activity in the system and whether you can export audit logs if something goes wrong. For many SMBs, “we will get back to you on that” is a sign to keep looking.
Clarify admin roles: who can change prompts, who can see customer content, and how access is removed when someone leaves.
5. Support and commercials
Ask about response times for production-impacting bugs. Beautiful roadmaps do not help if you cannot reach a human when invoices fail.
Understand limits: seats, tokens, API caps, overage pricing. Surprises here destroy ROI narratives.
6. Leaving the vendor
No one likes breakups, but they happen. Ask how you export workflows, prompts, and settings—and how long until data is deleted after the contract ends.
If export is “manual copy-paste,” price that labor before you sign.
Putting it together
You are looking for specific answers in writing, not vibes. If a vendor shines on product but wavers on data and logs, you are buying risk—not magic.
If you want an independent read before a big decision, our readiness audits are built for exactly that moment: contracts, data paths, and a sane pilot scope.