AI assistants have moved past the novelty stage: they read supplier emails, sort out invoices, prepare replies, and remind you about deadlines. But along with the benefits, you hand the service your most sensitive material — correspondence, prices, contract terms. Before connecting any tool, put seven questions to the vendor. The answers will quickly show you what you are dealing with: a mature product or a polished demo.
1. Where is our data stored, and who has access to it?
Ask for specifics: the country and jurisdiction of the servers, encryption in transit and at rest, and which of the vendor's employees can technically see your documents. For an importer in Kazakhstan or the EAEU, this is no formality: your correspondence with factories contains purchase prices, discounts, and terms that must never end up in a competitor's hands.
2. Does the model train on our data?
The defining question of the large language model era. If the vendor fine-tunes models on customer data, your commercial information can indirectly leak into answers shown to other users. The right answer sounds like this: "the model does not train on your data, and the contract says so." Insist on a written clause, not a verbal promise made during the demo.
3. What does the assistant do on its own, and what requires approval?
This is the dividing line between a suggestion tool and an operator. Reading an email and preparing a draft is safe. Sending a supplier an order worth tens of thousands of dollars is not. In a mature product, autonomy levels are configurable: what the AI does on its own, what it escalates for human approval, and what is off-limits entirely. If no such setting exists, control is nothing more than a promise.
4. Is there an action log, and are decisions explained?
When the assistant has sorted two hundred emails or checked an invoice, you need to see exactly what it did, on what grounds, and where it hesitated. Without a log, you can neither trace an error nor prove your case to a partner. The regulatory trend — including the European AI Act — is moving toward transparency and explainability, so an action log will eventually be the norm, not a bonus.
5. How does the system handle its own mistakes?
Language models make mistakes: they mix up numbers and invent line items that do not exist. The question is not whether the system errs, but what it does with uncertainty. Sound architecture means double-checking extracted amounts and part numbers, reconciling the invoice total against the sum of its lines, and honestly flagging "not confident — needs a human." Ask directly: "what happens if the AI misreads an amount on an invoice?" — and listen for whether the answer includes the word "verification."
6. How is access separated within our team?
A purchasing manager does not need payroll records, and an accountant does not need the sales team's correspondence. Find out whether roles and permissions exist, whether the assistant can be limited to specific mailboxes and folders, and how quickly a departing employee's access gets revoked. One shared login for the whole company is a warning sign.
7. What happens if we decide to leave?
Data export in an open format, clear timelines for deletion from the servers, written confirmation that the deletion happened. If the vendor starts hedging on this question, you are looking at future vendor lock-in that will be expensive to escape.
How to use this checklist
Send the seven questions by email before the demo — the quality of the answers tells you a great deal in advance. Good signs:
- they answer specifically, with references to the contract and documentation, rather than "everything is secure with us";
- they proactively offer a mode where critical actions require human approval;
- they discuss your exit scenario calmly.
To estimate how much time and money you free up by automating email and document routine, try the calculator, and short answers to common questions are collected in the FAQ.
At HORUVIA, we build an AI operator for importers around exactly these seven answers — from the action log to human approval of critical operations — and we take a closer look at each topic in other posts.