Six questions to ask before you let AI touch your customer list

Your software vendor has an AI agent now. It will handle your follow-ups, write your estimates, and answer your customers. The demo is genuinely impressive.

Before you switch it on against a list of real people who can call you, ask these six questions. The answers separate the products that will save you time from the ones that will cost you a customer on a Saturday.

1. Can it contact my customers without me?

This is the whole question and everything else is detail.

There is a real difference between software that drafts and software that sends. Drafting is enormously useful and carries almost no risk: the worst case is you delete something. Sending is a different product with a different failure mode, and the failure mode is a text message to fourteen people that you did not write and cannot recall.

Ask specifically: is there any configuration, any setting, any automation, in which this thing emails or texts a customer without a human seeing the message first? If the answer is “you can turn that off”, note that the answer was yes.

The stronger answer is that the restriction lives in the software itself rather than in a settings screen, so changing it would take a new release rather than a checkbox. A limit that an administrator can raise on a busy afternoon is not really a limit.

2. Can I see what it did, afterwards?

Ask to see the log. Not the chat history, the log: every action taken, when, on whose instruction, and against which record.

This is the single most common complaint about the AI agent products that are growing fastest right now. Not that they are wrong. That nobody can see when or why they acted. Software you have to keep checking is not saving you time, it is moving the work from doing to supervising.

3. What will it cost me next month?

The second most common complaint is unpredictability. Usage-based AI pricing is easy to sell and hard to budget, and “it depends how much you use it” is not a number you can plan around.

Ask what a unit is, what is included, and what happens when you exceed it. Ask whether background automations consume the same budget, because a daily instruction quietly runs thirty times a month. Ask whether you get warned before the overage or invoiced after it.

4. What happens when it is wrong?

It will be wrong sometimes. Anyone telling you otherwise is selling.

The question is what that costs. If being wrong means you delete a draft, the worst case is ten seconds. If being wrong means a customer received a price that was never approved, the worst case is a refund and a review.

Here is the part nobody in this category will volunteer, so take it from someone building in it: undo does not exist. Not in ours, not in any of the agent platforms, not in the workflow tools that added agents last year. Once a message is delivered it is delivered. Any vendor implying otherwise has not thought about it. The only real answers are prevention and visibility, and a vendor who says that plainly is telling you more than one who promises a rollback.

5. Is my data training somebody’s model?

Ask directly and get it in writing. Your customer list, your pricing, and your job history are competitive information, and “we may use aggregated data to improve our services” is a sentence doing a lot of work.

The answer you want is that your data is sent to the model to answer your own question, in your own account, and is not retained for training.

6. Can it see other businesses’ data, or they mine?

In multi-tenant software this comes down to whether isolation is enforced at the database, or by application code remembering to filter every query. The second one works right up until one query forgets.

You will not be able to audit this yourself. What you can do is ask the question and listen to whether the answer is specific. Vague reassurance about enterprise grade security is not an answer. “Row level security in the database, so a query that forgot to filter returns nothing rather than somebody else’s customers” is an answer.

The pattern in all six

Every question above is really asking the same thing: can this software be left running?

An assistant you have to watch has not saved you anything. An assistant that can act unattended on things you cannot take back has cost you something you have not been billed for yet. The useful middle is software that does the work, proposes the change, shows you exactly what would happen, and waits.

That is a slower product to build and it is the only version worth putting in front of your customers.