
Before connecting an AI to your business, ask the vendor twelve simple questions: what rights it will have, whether an action can run without you, what is logged, which data goes to the AI provider and for how long, what is excluded, what happens when an employee leaves, and how the AI says it does not know. A vague answer to just one of them is reason enough to wait. Print the list and keep the answers in writing.
Tuesday, 10 a.m., in the office behind the kitchen
Ibrahim owns two restaurants. A vendor offers to connect his AI assistant to his management software "in five minutes". The demo was convincing. Before saying yes, he prints a list of questions and sets it next to his laptop. He calls the vendor and ticks as he goes.
At the third question, "can I see what the AI did, and on whose behalf?", the answer is: "We don't keep that." Ibrahim puts his pen down. He is not turning the tool down; he is waiting for that answer to change.
(Scene and characters are fictional.)
Why a written list beats a demo
A demo shows what works
That is its job. It rarely shows what happens when the AI misunderstands, when an employee leaves, or when someone asks it to record something that should never be recorded.
The plug says nothing about the rules
These connections increasingly go through a shared standard called MCP. Like a plug standard, it says how to connect, not what is allowed through: the vendor's settings are what count, and they are what the twelve questions are about. We explain it in the pillar page MCP: connecting your AI to your business, explained for frontline trades.
A written answer is a commitment
"Yes, of course" on the phone is worth little. A written, dated answer that points to a public page or the contract can be checked and quoted back.
The annotated checklist
What the AI can see and do
- What rights will the AI have? The only good answer: exactly those of the person using it. If a line cook connects their assistant, the assistant sees what the line cook sees, nothing more. Be wary of "global access, it's simpler". See give your AI the same rights as you, no more.
- Can an action run without my confirmation? It should not: the right answer describes a draft, a summary of the effect (what, for whom, where) and your explicit go-ahead. Also ask whether access is checked again at the moment of confirming. The principle is set out in AI proposes, a human confirms: the two-step rule.
- Is there a log of actions? It should show who confirmed, when, and what changed. Without a log, you cannot understand a mistake or reassure a team.
Where your data goes
- Which data is sent to the AI provider, and where is it processed? When the AI reads information to answer you, that information goes through its provider. Ask what leaves, whether reads are limited (so nobody can pull the whole database at once), where the vendor hosts its own data, and check the terms of your own assistant subscription on your side.
- How long is the data kept? You need two answers: the software vendor's and the assistant provider's. "We keep everything, we're not sure for how long" is a warning sign.
- Can I get my data back and leave? Reversibility is not only about AI, but one more connection should not make leaving harder. Ask about format, timing and what the contract says. On this, read your restaurant data should remain yours: why reversibility matters.
What the AI must not do
- What is excluded from the AI's actions? A precise list is reassuring, for example: payroll, invoicing, permanent deletion and legal acts. "Nothing is excluded as long as you confirm" is not.
- What happens if someone asks it to record sensitive data? An employee's health, a pregnancy, a disability, their religion or opinions are special-category data under Article 9 of the GDPR, the EU data protection regulation. A good tool refuses, even when the request is reworded, and offers to log a dated, observed fact instead.
When people change
- Who can cut off access, and how fast? Removing a right should take effect immediately, not at the next renewal. Ask who can do it (the owner, an administrator) and whether the person can remove the connection from their own assistant.
- What happens when an employee leaves? Their access must be cut, including the assistant they had connected. Their work history, on the other hand, stays with the business.
The quality of the answers
- How does the AI say it does not know? It should say so plainly: "no data recorded", "two sources disagree". An AI that always finds an answer will end up inventing one.
- How do I test before rolling it out? A good vendor offers a trial on non-sensitive documents or in a separate environment. The test should include real questions, one forbidden request and one action to confirm.
The summary table
| Question | Good answer | Warning sign |
|---|---|---|
| 1. AI rights | Same as the person, no more, no less | "Global access, it's simpler" |
| 2. Confirmation | Prepared, summed up, confirmed; access rechecked | "It acts alone to save time" |
| 3. Log | Who, when, what, available to review | "We don't keep track" |
| 4. Data sent out | What leaves is known, reads limited, hosting stated | "Everything is sent, it's fine" |
| 5. Retention | Written periods, vendor side and assistant side | "We don't know" |
| 6. Reversibility | Format, timing and terms in the contract | "You won't want to leave" |
| 7. Exclusions | Precise list (payroll, invoicing, permanent deletion and legal acts) | "Anything goes if you confirm" |
| 8. Sensitive data | Refused, with a dated fact offered | "It's up to you to be careful" |
| 9. Cutting access | Immediate, by a named person | "At the next renewal" |
| 10. Employee leaving | Access cut, history kept | "Just email us" |
| 11. "I don't know" | Said plainly, with the reason | "The AI always has an answer" |
| 12. Testing | Trial on non-sensitive data, with trick cases | "Plug it in and see" |
Common mistakes
- Asking out loud and keeping nothing. Get the answers in writing, or links to the public pages that give them.
- Stopping at technical questions. Questions 8 and 10 are about people, not technology, and they are often the ones that stall.
- Testing with the owner's account. It sees everything, so you are not checking question 1. Test with a site manager's account as well.
- Forgetting the assistant side. The software vendor cannot answer for your AI provider: you read their terms yourself.
- Confusing "I confirm" with "I review". A confirmation is only worth something if the summary is readable: the task, the person, the date.
What to measure
- How many questions on the list have a written, checkable answer. Twelve out of twelve before connecting more than one person.
- During the trial, the share of trick requests (non-existent data, sensitive data, out-of-scope action) correctly refused.
- The actual delay between removing a right and the assistant no longer seeing anything, measured once with a test account.
Tsuno's answers to this list
Here is what Tsuno publishes on its Integrations and Trust and data pages, question by question. ChatGPT, Claude, Mistral or another compatible assistant can query Tsuno.
Questions 1 to 3: the assistant stays within the person's scope in Tsuno. It drafts, the person reads the summary and confirms in a new message, with access checked again at that moment. The log keeps who, when and what, with the state before and after the action.
Questions 4 to 6: reads are capped to prevent bulk extraction, and the app and its data are hosted in Germany, in the EU. What your assistant reads to answer you is processed by its provider, under your terms with them. When your subscription ends, your data stays available to you for thirty days, in a standard format; the details are in the contract.
Questions 7 to 11: payroll, invoicing, permanent deletion and legal acts are out of the assistant's reach, even when confirmed. Tsuno refuses to record an employee's health, pregnancy, disability, religion or opinions, and offers to log dated, observed facts instead. Removed access takes effect right away, and an employee leaving ends their open sessions. When data is missing, Tsuno says so instead of filling the gap.
Question 12: the three-document trial runs in an isolated environment, with no personal data, deleted on the agreed date. On the voice side, Tsuno San follows the same rules.
Further reading
To understand the plug itself, read the pillar page MCP: connecting your AI to your business, explained for frontline trades. Questions 1 and 2 are covered in depth in give your AI the same rights as you, no more and the two-step rule. And for the written limits of the voice assistant, see what San never does. If a word on this list stops you, the AI at work glossary for restaurant owners translates it into kitchen language.
Key takeaways
A connected AI is judged on its rules, not its demo. Ask the twelve questions, keep the answers in writing, and only connect the whole team once you have tested the awkward cases.
Frequently asked questions
What should I ask before connecting ChatGPT or Claude to my business software?
The essentials cover the AI's rights, confirmation of actions, the log, the data sent to the AI provider and how long it is kept, exclusions, sensitive data, cutting off access and what happens when an employee leaves.
Should an AI connected to my business have access to everything?
No. It should have exactly the rights of the person using it, no more and no less. An assistant connected through a manager's account should only see what that manager already sees.
How can I tell if an AI is making up answers about my data?
Ask it something you know does not exist in the tool. A good tool says it does not have the information. If it invents an answer, the connection is not ready for daily use.
What happens when an employee who connected their AI leaves the company?
Their access should be cut right away, and the assistant they connected should no longer see anything. Their work history stays with the business for traceability.
How can I test a connected AI safely?
Start with non-sensitive documents or a trial environment, with real questions, one forbidden request and one action to confirm. Only roll it out after that test.