A customer calls, upset. Last night your automated email reply told them a job would cost one price, and today your office quoted another. “Your system promised me,” they say.

What did the automation actually read? What did it send? Which rule told it to say that? If you cannot answer those in a few minutes, you are guessing in front of an angry customer.

That question is about to get more common, and not only from customers.

Why this is suddenly a real question

On September 30, the Federal Trade Commission confirmed it is investigating Anthropic, OpenAI and other AI companies, CBS News reported. A spokesperson said the agency is looking at whether the companies broke the FTC Act, the federal consumer protection law, and that the probe opened this summer.

That investigation is about the companies that make AI models, not about small businesses that use them. But the head of the FTC has been clear about how he sees responsibility. Speaking at a Reuters event in late September, FTC Chair Andrew Ferguson said: “If someone tells a tool to do something, and the tool does it, I don’t think we would say, ‘Oh, what do we do about the tool?’”

He also said that when AI systems were described as acting on their own, reviews of their records showed they were following the instructions they had been given.

Put plainly: if your automation does something, the question will be who told it to. The answer lives in the records.

A made-up example from a shop’s day

Here is a made-up example. A heating and cooling company uses an automation to answer website forms after hours. It reads the form, checks the price list and drafts a reply.

In March, someone updated the price list but left an old tune-up special in a second tab. The automation found the old special and offered it to a customer. Nothing malicious. Just an old tab.

If the shop has good records, the fix takes ten minutes. They can see the form the customer sent, the price list tab the automation read, the reply it sent, and the rule that allowed it to send without a person checking. They honor the price, delete the old tab, and change the rule so price quotes wait for approval.

Without records, the same shop spends a day arguing about what happened, and it may happen again next week.

Four things you should be able to show

For any automation that talks to your customers or touches your money, you should be able to pull up four things on a normal day, without calling anyone:

  1. What it saw. The email, form or record it was working from.
  2. What it did. The message it sent, the record it changed, the file it moved.
  3. Which rule or instruction it followed. The setting, template or step that told it to act.
  4. Who approved it, if anyone. Whether a person checked it first, and who.

If any of those four is missing, that is the gap to close first.

Five questions to ask whoever runs your automations

You do not need to read code to ask these. You only need the answers in writing.

  1. Where do the records live, and are they in an account you own? Records that sit only inside a vendor’s account can disappear when the vendor does.
  2. How long are they kept? Thirty days is not long if a customer complains about something from last quarter.
  3. Can you export them? You should be able to download them yourself, in a format you can open.
  4. Who gets told if something goes wrong, and how fast? Name one person on your side who receives that message and can act on it.
  5. How fast can it be switched off? Ask them to show you, not tell you.

A vendor who answers these clearly has built for the day something goes wrong. A vendor who talks only about how smart the AI is has not.

Check what you promise, too

One more thing worth ten minutes this week. Read what your own website and your own emails say about how your business handles customer information and automated replies. If you promise a person reviews every message, make sure a person does. A promise you cannot back up is a problem of its own, whoever built the tool.

Your one step this week: pick the one automation that talks to your customers most. Try to pull up those four things for its last message. Time how long it takes. If you cannot find them, you now know exactly what to ask for.

If you want help sorting it out, here is what we build and how it is handed over to you.

Sources

Kush AI Automation builds in accounts you own, and nothing that moves money happens without a person approving it first. Bring the automation you are least sure about and book a free call.