What to ask before you buy AI for your business.
Seven questions that separate a tool you can supervise from one you will regret. Written for owners of small local service businesses who do not have an IT department and do not want a surprise in front of a customer.
Ask seven things: what it can do while you are not watching, what waits for your approval, whose accounts it runs in, whether you can read what it did, what you keep if you leave, what it costs in your busiest month, and what happens after it gets something wrong. A tool worth buying answers all seven in plain language, without a demo call.
Most buying advice for software is about features. For AI it is the wrong question, because the thing you are actually buying is permission: you are letting a system act on your business without asking you every time. The features are easy to change later. The permissions are what create the bad afternoon.
Below is the list, in the order worth asking it. Each one has an answer you want and an answer that should stop the conversation.
What can it do without asking me?
The list of actions the tool can take with nobody watching, written as verbs. If the answer is a category like marketing or follow-up rather than a specific action like send a second email to a lead who has not replied in three days, the scope is not defined yet. Undefined scope is what produces surprises.
A sane default is that unattended work is small, reversible, and internal. Drafting, sorting, reminding, tidying a record, preparing something for you to look at. Anything a customer sees, anything that spends money, and anything you cannot take back belongs on the other list.
Write the list down during the sales conversation and keep it. It is the document you will check against on the day something unexpected happens.
What waits for my approval, and how do I approve it?
Which actions pause for a yes, and how long saying yes takes. An approval queue you have to hunt for gets ignored within a month, and an ignored queue is the same as no gate at all.
There are two failure modes and they sit at opposite ends. If everything waits for you, you have not delegated anything, you have added a second inbox. If nothing waits for you, you find out what the tool did when a customer tells you.
Ask to see the approval screen before you buy, not after. Then ask the question that actually matters: what happens if I do not answer for a week. Does the work sit and wait, or does it go out anyway? Only one of those answers is safe on a busy week.
Whose accounts does it run in?
Where the work happens, what permissions it actually needs, and how you get your data out and revoke access if you stop. A vendor who can answer plainly has thought this through.
Then ask which permissions it needs and why each one is needed. Full access to everything is rarely the real requirement, and a vendor who can explain the specific reason for each permission has thought about it. One who cannot has not.
Finish with the exit question: on the day I stop, what do I have to do to cut off access, and how long does it take? If the answer involves waiting on their support queue, that is your answer.
Can I read what it did, in plain language?
A real log line from a real account. You want a record that says what was done, when, and why, in the words you would use, not a technical trace. If you cannot audit the work in a few minutes a week, you cannot safely delegate the work.
Two tests are enough. First: can you tell from the log whether one specific customer was contacted, and what was said? Second: can you tell what still needs a decision from you without reading every line? A log that fails either test is a log you will stop opening by week three.
This is also the question that protects you later. When something goes wrong, the log is the difference between a five minute fix and an afternoon of guessing.
What do I keep if I leave?
What leaves with you: the record of what was done, the rules and knowledge you wrote down, and the accounts themselves. Ask for the export before you sign, not after. If the export exists the answer takes a minute, and if it does not the answer takes weeks.
Contract length matters less than exit friction. A month to month agreement is not worth much if everything you built over a year lives in a system you cannot get anything out of. The knowledge you feed a tool about your pricing, your customers, and how you like things said is the expensive part, and it should be yours.
What does this cost in a heavy month?
The price of your busiest month, not the starting price. Tools priced by usage, seats, credits, or messages can cost several times the headline number in exactly the month you need them most, which is the month you are least able to sit down and read a bill.
Ask it directly: is there a meter, what happens when I go over, and who tells me before it happens? Then write the answer down. A vendor who cannot state the worst case is telling you the worst case is unbounded.
Do the same for the setup. Ask whether there is an implementation fee, whether it is refundable if setup does not finish, and what you have paid for if you cancel in month two.
What happens after it gets something wrong?
What happens next when it sends something wrong, because it will eventually. Who notices, how fast you find out, what can be undone automatically, and what you would have to apologize for by hand.
The good answer is boring. The risky work was gated so the wrong thing never went out unattended, the reversible work is genuinely reversible, and the log shows exactly what happened and when. The bad answer is a promise that it will not happen.
While you are there, ask whether the tool promises outcomes. Anyone promising rankings, revenue, or a number of leads is promising a result they do not control. Work is something a vendor can commit to. Results are not.
The answers you want, and the ones that should stop you
The same seven questions, with the signal to listen for in each answer.
| Question | Good sign | Stop sign |
|---|---|---|
| What runs unattended? | A written list of specific, reversible actions | A category name, or "it learns what you want" |
| What waits for approval? | You can see the approval screen today | Approvals are on the roadmap |
| Whose accounts? | Your accounts, with per-permission reasons | Their platform, full access, no detail |
| Can I read the log? | Plain language, one line per action | A dashboard of charts and no record |
| What do I keep? | Export it yourself, anytime | Request an export through support |
| Heavy month cost? | One number, stated up front | "It depends on usage" |
| After a mistake? | Gated, reversible, and logged | A promise it will not happen |
Use it as a scorecard. Fewer than five good signs is not a negotiation, it is a no.
How do I try it without betting the business?
Run one task for four weeks and write down the number you expect it to move before you start. One reversible task, one number, one month. That is enough to tell whether a tool earns its keep, and small enough that being wrong costs you a month.
- Pick a task that repeats weekly and is reversible. Follow-up on quotes that went quiet is a good first one. Anything that touches money or a public channel is a bad first one.
- Write the current number down first. How many quotes went unanswered last month, or how long a new inquiry waited for a reply. Without the before number, the after number means nothing.
- Leave approvals on for the first two weeks. Read what it wanted to send. That is the cheapest way to learn whether it understands your business.
- Read the log once a week, on the same day. Fifteen minutes. If that feels like too much work, the log is the problem.
- At four weeks, compare to the number you wrote down. If it moved, widen the scope by one task. If nothing moved and nobody can explain why, stop and keep your month.
How Gubernel answers these questions
Gubernel works from a written rulebook you control. Small reversible tasks run on their own, and anything big, public, or hard to undo pauses for your approval. Every action is logged in plain language and labeled with one of six signal states. It runs in its own hosted workspace and connects to your accounts with permissions you grant.
On cost, the plans are flat and monthly, from $249/mo, billed monthly, with no setup fee and no usage meters, so a heavy month costs the same as a quiet one. On the Autopilot plan the AI runs on your own provider account at cost, with no markup, which means you hold the key. See pricing for the three plans and the plain answers.
On what you keep: you can export everything in your workspace, item for item, with a receipt showing nothing was left behind, and take it with you if you go. There is no long contract and you can cancel whenever you want. On results, there are no ranking or revenue promises here, only a record of the work and a weekly brief you can judge for yourself. The full routine, the approval gate, and the brief are laid out on how it works.
It is also worth knowing what it does not do. Gubernel does not place or answer phone calls, and it is not a chatbot you sit and talk to. Ask every vendor the same question and expect a straight list back.
Read next: How to automate the busywork of a local service business.
Short answers to the rest of it.
What should I ask an AI tool before I let it contact my customers?
Ask for the written list of messages it can send without you, ask which messages wait for your approval, and ask to see the approval screen before you buy. If the scope is described as a category like marketing rather than as specific actions, the scope is not defined yet.
Is it safe to give AI access to my business accounts?
It is as safe as the limits you set and can check. Ask which permissions it actually needs and why, whether the work happens inside your own accounts or on the vendor's platform, and how access is removed on the day you stop. Full access to everything is rarely the real requirement.
How long should I test an AI tool before deciding?
About four weeks on one task is enough. Pick something that repeats weekly and is reversible, write down the number you expect it to move before you start, keep approvals on, and read the log once a week. If nothing moved and you cannot say why, stop.
What is a red flag when buying AI for a small business?
Three of them: a vendor who cannot tell you the cost of your busiest month, a log you cannot read without their help, and a promise about rankings, revenue, or leads. The last one is a promise about a result the vendor does not control.
Do I need to be technical to keep AI in check?
No, but you do need a record you can read. If you can tell from the log what was done, who it went to, and what still needs a decision from you, you can supervise the work in a few minutes a week. If the record only makes sense to the vendor, you cannot.
What this looks like in your line of work.
See the office work Gubernel runs for plumbers, cleaning businesses, electricians, contractors, and gyms and studios.
Ask us the seven questions, we answer them straight.
Start the intake and you see your first brief before you pay anything. A real person is on the other end if you would rather ask in person.
Get started