Your AI Agent Has No ROI. That's Why You Want a Kill Switch
Nobody caps a Facebook campaign that returns $1.50 on the dollar.
Somewhere, out there on our little blue dot, there’s an AI agent running, providing a great service to the organisation that commissioned it.
Its token bill is climbing, and since it was put into production, someone has “ownership” of what it’s doing. That same someone is starting to have responsible-sounding voices in their head asking if a kill switch is a good idea. Something to flip when it burns through too many tokens.
They know the feeling. I know because I see the emails all the time.
A cap, a limit. A way to stop it running away with a budget.
It feels like prudence, the grown-up doing their job. I’ll tell you what that instinct actually is, and it’s not a criticism. It’s a confession.
And what it’s confessing has nothing to do with tokens.
We’re told that a rising LLM token bill is a cost problem, and cost problems are solved by control - a cap, a throttle, maybe even some prompt optimising. The fear that’s felt when the bill climbs isn’t really fear of the cost. It’s fear of the cost with nothing on the other side of it.
And that missing other side has a name: a business case.
This is where you need to think about the money you’re not frightened of spending
If you ran a campaign that reliably turned every dollar in into a dollar-fifty out, you wouldn’t cap it - you’d feed it. You’d spend half a million a month and lie awake wishing you could spend more, because every dollar going in was a machine for making more come back.
Nobody bolts a nervous little kill switch onto a profitable ad campaign. Or think about Black Friday: when a retailer lets Amazon auto-scale its infrastructure through the busiest weekend of the year, the cloud bill goes vertical — and nobody trips a breaker when the invoice looks frightening, because killing the servers on your biggest sales day would be madness. The spend is enormous and nobody’s afraid of it, because everyone can see the revenue pouring in behind it.
Alex Hormozi says the quiet part out loud better than anyone: the business that can afford to spend the most to acquire a customer wins. Build a machine that turns a dollar into two, and you don’t ration the dollars — you spend every one you can find. Cost stops being a threat and becomes the most obvious move in the world, the moment you know what it returns.
So here’s the question that turns the whole thing over. Why do you treat the AI agent differently? Why is the token bill the one number that makes you reach for the switch?
Because with the ad campaign and the Black Friday servers, you know the return - and with the agent, you don’t. That’s the whole of it. The kill switch isn’t a cost control; it’s a monument to a business case you never built. You want to be able to shut the agent down when it spends “too much” because, if you’re honest, you can’t say what “too much” even means — because you’ve never worked out what one run of it is worth. You don’t know if it turns a dollar into two or a dollar into nothing, so every pound it spends lands as a risk instead of an investment. It isn’t the tokens frightening you. It’s the silence where the return should be.
Nobody fears the cost of a machine they know makes money. A token bill is only frightening in the absence of a business case and the fear is the tell.
There's a legitimate version of the cap, and it isn't this one: a ceiling that catches a runaway loop, a leaked key, a bug that turns one query into ten thousand. That's blast-radius control, and every production system should have it. What I'm describing is the other kind. The cap you want because you can't say what a single run gives back. One is engineering. The other is a business case with a switch where the number should be.
If this sounds familiar, here’s what it means, and I’ll say the reassuring part first, because it matters: the fear is not irrational, and you are not bad at your job for feeling it. You’re feeling exactly what a sensible person feels when they’re spending real money with no proven return - the discomfort is doing its job. The mistake isn’t the fear. The mistake is what we’ve all been trained to do with it: point it at the tokens. Capping the spend doesn’t fix what’s actually wrong; it just makes a machine you can’t measure slightly cheaper to keep not measuring.
The answer unfortunately (because there aren't any shortcuts here) is to do the harder, better thing. Before you build the kill switch, build the number the kill switch is standing in for. Take one agent, one job it does, and work out what a single run of it truly returns - the hour of someone’s afternoon it handed back, the query it answered that would’ve cost a specialist a morning, the escalation it stopped, the deal it kept warm. Put it in dollars. Set it against what the run cost.
If the return is bigger (and for a well-aimed agent it usually is, by a lot) take your hand off the switch, because you’ve just discovered you own the campaign that prints, and the right response to that is to feed it, not fence it. And if you sit down to do it and find there’s nothing solid to put on the page, then you’ve found your real project, and it’s a far better one than “make the tokens cheaper” - because it’s the one that decides whether this thing should exist at all. Either way the fear lifts, because the fear was only ever the missing number made physical.
So my view is don’t reach for the switch. Reach for the number behind it. Pick an agent or agent use case whose bill scares you most, write down what one run of it is worth in pounds, and let that single figure tell you whether you’re holding an investment to feed or a business case to go and build. One of those is the answer. The kill switch never was.
Until next time,
Chris
P.S. — The money gate is only the first of three, and most projects that survive it die at another door. How AI Projects Die walks all five — two screens, three gates, ten real engagements, the numbers left in, one move at each. Free, twenty pages: take a copy.