Your AI agent doesn't know when to stop.
A client recently got a $2,500 surprise bill from a single tool his AI was using. Half a million calls in a month, most of them asking the same questions over and over. Here's what actually went wrong, and the four guardrails that turn an expensive habit into a predictable line item.
A client called this week with a familiar problem. His monthly bill from one of the services his business runs on had climbed past $2,500. That number was not a typo. It was not malice, fraud, or a price hike either. It was an AI agent he had built, doing exactly what he had told it to do, just a lot more often than anyone realized.
That agent had quietly made roughly half a million calls to the service in a single month. The included plan covered five thousand. Every call past that was billable. The bill arrived without warning, the way these bills always do.
What the agent was actually doing
The setup was reasonable on paper. The agent watched a community platform for new questions from customers and helped draft answers. Most operators would call that smart automation. The trouble was how the agent went looking for new questions.
Every few minutes, it asked the platform a simple question: anything new? Then it asked again. And again. In between those asks, it pulled down a list of every recent post, scanned them all, then scanned them again on the next pass to see if anything had changed. By the end of the day, it had asked about the same posts hundreds of times.
When we pulled the numbers, the vast majority of those half a million calls were the same handful of questions, repeated. The agent never wrote down what it had already seen. So every run was a fresh start, as if it had amnesia.
The agent was working hard. It was just answering the same question half a million times.
This is not an "AI gone wild" story
It is tempting to file this under runaway AI. That framing is wrong, and it points people at the wrong fix. The agent did not go rogue. It did exactly what it was built to do. The miss was that nobody put limits on it. There was no daily ceiling, no memory of what it had already done, no off switch the owner could flip from his phone.
AI without guardrails is like a sprinkler with no timer. It will keep running. The cost will not show up for a while, and when it does, it will not be subtle.
The four guardrails every AI workflow needs
When we walked the client through the fix, it came down to four small pieces. None of them are exotic. All of them are missing in most setups we audit.
1. A kill switch you can flip in seconds
Every AI workflow should have a single on or off control that anyone with access can toggle without writing code, calling a developer, or restarting anything. One flag in a database, one row, two values: on or off. The workflow checks it before every action. If it is off, nothing happens.
This sounds obvious. It is almost never built. When something goes wrong at three in the afternoon, the founder should not have to wake up a developer to stop the bleeding.
2. A daily budget
A second, equally simple piece of math: a counter that says "the workflow may do at most this many calls in a 24 hour window." Once the counter hits the limit, the work pauses until tomorrow. The cap is set by what you are willing to spend, not by what the underlying service technically allows.
This converts the worst case from "unbounded" to "known." A surprise bill is mostly a function of unbounded. Replace that with a number you chose, and the surprise goes away.
3. A memory of what is already done
This is the one that almost always collapses the bill. Most runaway costs come from an agent doing the same work twice, or two hundred times. The fix is unromantic: a small list of things the agent has already handled. Before doing any piece of work, the agent checks the list. If it is on the list, it skips.
For the client above, this single change would have cut the bill by something like 99 percent. The work that actually mattered was a few hundred items a month, not five hundred thousand.
4. Do not ask. Get told.
The most common shape of an expensive AI bill is what engineers call polling: asking a service "is there anything new?" over and over. Polling is the most expensive way to find out about new things, and almost every modern service offers a free alternative.
That alternative is called a webhook. Instead of your agent constantly checking the platform, the platform calls your agent the moment something happens. Same information. A fraction of the cost. Often, no cost at all.
Polling is asking the mailbox every two minutes if a letter arrived. A webhook is the mail carrier ringing the doorbell.
Most platforms support this. Most agents are not built to use it. That gap explains almost every expensive automation story you have heard in the last two years.
What this looked like once it was fixed
For the client, the fix was not a rewrite. It was a single small piece of code sitting between the agent and the outside world, doing four things at once: checking the kill switch, counting against the daily budget, looking up what had already been done, and pacing the work so it could not burst.
The expected new bill is closer to zero than to $2,500. The agent is doing the same useful work it was doing before. It is just no longer doing the wasteful work it was also doing.
Why this keeps happening
Almost every operator we audit has at least one of these gaps somewhere. The reason is the same every time: the people who set up the automation were focused on getting it working. Nobody was asked the second question, which is what happens when it works too well, at three in the morning, on a Saturday, with no one watching.
That second question is the one that separates an AI experiment from an AI system. It is also the one most shops will quietly skip, because it does not produce a flashy demo. It just keeps you out of trouble.
What we do about this
Most of the work we do in an AI Operations Sprint is not building a smarter AI. The smart AI already exists, you can rent it by the month. The work is putting it inside a frame that knows when to stop, what to skip, how much to spend, and how to be turned off.
Frame the AI right and the bill is small and predictable. Frame it wrong and the bill shows up as a phone call from a vendor you were not expecting to hear from.
If you are running anything in your business that calls an outside service on a schedule, and nobody in the room can tell you the maximum it could cost next month, that gap is what we are talking about. It is a one afternoon problem to fix. The other option is letting the bill teach you the lesson the harder way.
Writing about AI systems for founder-led businesses across NWA, the River Valley, and Eastern Oklahoma.
Context is the bottleneck, not the model.
Almost every business I audit has the same problem. It isn't capability. It isn't tooling. It's that the AI doesn't know how the business actually thinks.
I charge less than agencies — on purpose.
Pricing in this industry is mostly a function of labor cost. I removed most of the labor, so I removed most of the cost. Here's why I pass the savings on.