Before approving an AI vendor, your IT team should ask eight questions, and the speed of the answers matters as much as their content. A vendor should be able to answer all eight in writing within a week. One who cannot has told you where their documentation stands, which is itself useful. The questions below apply to any AI tool that touches your data; our examples come from engineering reporting, where the data is a client's site record, because that is the review we sit on the receiving end of.
These are not questions we invented. They are the ones firms' IT reviewers actually put to vendors, sometimes at the first conversation, more often at the last one, when the deal is otherwise agreed and someone finally asks.
Where does the data physically live?
Ask for jurisdiction rather than "the cloud". For firms doing work for public sector or defence-related clients, servers located in North America is frequently a contractual requirement rather than a preference, and it is the wrong thing to discover after signature. A good answer names the region and the infrastructure provider. A weak one says the data is stored securely, which is not a location.
Is it dedicated hardware or shared infrastructure?
Ask which, then ask what the difference means in practice for your data. Reviewers ask this early because it shapes every other answer. The honest answer for most vendors is shared cloud infrastructure with logical separation, and that is a perfectly reasonable architecture. What matters is that the vendor says so plainly rather than implying something more isolated than what exists.
What happens to the data before it reaches a model?
Ask whether anything is removed or masked before content is sent for processing, and what exactly. This question separates vendors who have thought about their pipeline from those who have not. A useful answer describes what is scrubbed and what is not. "Nothing is removed" is a legitimate answer if it is true and the terms covering the processing leg are strong; the reviewer needs to know which it is.
Does the vendor's agreement with the model provider differ from consumer terms?
Ask for the substance. Few buyers realise the terms differ at all between a consumer product and a commercial agreement with the same model provider, which makes this question rarely asked and disproportionately informative. Our own answer: Tenera uses commercial API agreements with our AI model providers, under which your data is not used to train their models and retention is contractually limited. Ask us for it in writing anyway; that is the standard this article sets.
Is our data used for training?
Ask about the vendor's own models and about any third party's, separately: these are two questions, and a single denial often covers only the first. "We take privacy seriously" answers neither. Our own answer is no on both counts: we do not use your content to train AI models, and our agreements with model providers exclude it on their side too. Hold us to the article's standard and get that in writing rather than from this page.
What are the retention and deletion terms?
Ask what is kept, for how long, what deletion actually removes, and whether it reaches backups. Then check the answer against your own obligations, which usually run longer than a vendor's default: professional firms carry record-retention requirements measured in years, varying by jurisdiction and the nature of the work, so confirm your own rather than relying on a general figure. A tool whose retention is shorter than your obligation is a problem to solve at the firm level, and the mirror question matters too: the right to have data deleted on request, and what that guarantees.
What happens if the vendor disappears?
Ask about redundancy, continuity, and export. Fire, outage, acquisition, insolvency: reviewers ask about all four, and they are asking a business-continuity question rather than a security one. The practical form: can you get every document out in a usable format, on demand, without the vendor's cooperation? If the answer depends on goodwill, record it as a risk.
Who at the vendor can see our content?
Ask for the internal access model, and whether support staff can view customer content. Most vendors need some access to support the product; the reasonable position is that it is limited, logged, and disclosed. A vendor claiming nobody internally can ever see anything is either running an architecture they should be explaining in detail, or is not being precise.
Run the review early, and read the answers as evidence
This review usually runs last, after the champion has sold the tool internally, and by then the questions arrive as an obstacle: vendors rush, reviewers dig in, and reasonable products fail on process. Run first, it costs an hour and changes the conversation. Send all eight questions in one email before the demonstration, and note how long the reply takes and whether it arrives as a document or as prose in an email body. Both signals matter.
Two calibration points while you read the answers. A security page is not the review; the review is whether the vendor answers your eight questions specifically, in writing, about your situation. And a missing certification is not automatically disqualifying: a small vendor without a badge may still describe your data's actual path more clearly than a large one with a badge in the footer. Whatever you approve, the confidentiality duty to your client stays with the firm; no supplier agreement moves it.
Send the eight questions to us first
Tenera Reports is an AI tool that writes engineering reports from site photos, video, and voice notes, so this review applies to us in full. Send us the eight questions before any demo: book a call or write to info@tenerapro.com and ask for the answers in writing. The checklist at the top of this page is the download version of the eight questions, ready to send.