Your Automation Needs a Speed Limit

Paul Birrell went looking for a file and found out his business had sent roughly 400,000 emails over three days.

He runs a food and catering delivery service in New Jersey, plus a couple of other ventures, with about 30 AI agents and a handful of human contractors. An automation he had set up months earlier had started looping. A second loop in his messaging system burned through $1,200 of text messages in a single day before he caught it. Forbes reported the story in October, and his reaction is the part I keep coming back to: "I don't have 400,000 customers."

Nobody hacked him. Nothing was broken in the usual sense. A workflow did exactly what it was built to do, over and over, with nobody watching the meter.

The new tools don't wait for you to ask

A year ago, most AI in a small business was a chat window. You typed, it answered, you decided what to do with it. If it got something wrong, the damage stopped at one bad answer.

That is changing fast. On November 4, Perplexity launched Teammates, agents built for long-running work that, in the company's words, "monitor ongoing work and follow up as things change" without waiting for a fresh request. Meta, OpenAI and Anthropic are all pitching agents to small business owners too. The pitch is an assistant that keeps working while you are on a job site or asleep.

That is useful. It is also a different kind of risk. When the AI only talked, a mistake happened once. When it acts on a schedule, a mistake can repeat itself thousands of times before anyone notices.

The question used to be whether the answer was right. Now it is also how many times the thing can run before a person looks.

Why small businesses get hit hardest

A big company has someone whose job is to watch dashboards. When something spikes at 2am, an alert goes off and a person wakes up.

A ten-person company has the owner. And the owner is not checking the email platform's send count on a Tuesday afternoon. Birrell found his problem by accident, while looking for something else.

There is also a quieter cost. When Birrell finally set up spend reporting, it turned up a billing detail he had been paying for without knowing it. His text message provider counted one message as 70 characters, and he had written dozens of system messages assuming 120 or 140. He was paying twice for every send.

That is the pattern with automation. The expensive mistakes are rarely dramatic. They are small, repeated, and invisible until you count.

Put a ceiling on everything that bills by the unit

You can't watch every workflow yourself. You can make sure none of them can run forever. A few rules cover most of it.

1. Set a hard cap on anything that charges per use. Text messages, emails, AI credits, ad spend, API calls. Many platforms let you set a daily or monthly limit. If yours doesn't, that tells you something about the platform.
2. Add a volume tripwire. Pick a number that would be strange for your business, like 500 emails in an hour when you normally send 50, and have the tool pause and notify you when it gets there.
3. Get a regular spend report. Birrell now has a job that reports his messaging spend every four hours, with a standing instruction at the end of it: "Put the brakes on it. Stop it." You do not need every four hours. A daily text with yesterday's counts would likely have flagged his loop on day one.
4. Know where the off switch is before you turn it on. For every automation, write down one line: how to stop it, and who besides you can do that.

None of this needs a technical background. It needs ten minutes per workflow and the habit of asking one question: what is the most this thing could do in a day if it went wrong?

Separate "can" from "may"

One detail from the Birrell story stuck with me. He has an AI agent he calls his chief financial officer. It reconciles his books, checks his banking, and sends him a daily expense report. It has the access it would need to move money.

It has never made a transaction, because he never gave it permission to.

That distinction matters more as these tools grow. Access is what the tool can technically reach. Permission is what you have actually told it to do. Most runaway problems happen in the gap, when a workflow has the access to keep going and nothing that says when to stop.

So when you connect a new agent to email, payments, or your customer list, ask two separate questions. What can it touch? And what is it allowed to do on its own without checking with you first? If you can't answer the second one in a sentence, it isn't ready to run unattended.

Start with the one that would hurt most

You don't need to audit everything this week. Pick the single automation that could cost you the most if it looped: usually anything that messages customers or spends money. Find its limit settings. Set a cap. Turn on a daily count. Write down the off switch.

Then do the next one.

The goal isn't to slow AI down. Birrell still runs about 30 agents. The difference is that a spend report now reaches him every few hours, so a loop gets caught the same day instead of found by accident.

Where we come in

At Nexera, setting limits and off switches is part of every automation we build for clients, not something added after the first surprise bill. If you have workflows running and you are not sure what they would do on a bad day, we can help you find out at nexeraintelligence.com.