What It Actually Costs to Run a Company on Agents

Most writing about AI agents shows you a demo. A prompt goes in, something impressive comes out, and the cost, the failure rate, and the human still sitting behind it are left off screen. So it is worth paying attention when a company opens its books.

SaaStr, the SaaS community and media company, runs its operation with three humans and more than 20 AI agents, and it has started publishing the actual back end: commit counts, API stacks, monthly costs, and what each agent does. The picture that emerges is not the demo. It is the invoice, and the invoice is the more useful artifact.

The demo shows you what an agent can do. The invoice shows you what it costs to keep one running.

The stack, with numbers

SaaStr says it has put more than $500,000 into AI infrastructure to get where it is. It runs more than 20 agents across go-to-market work, and in its latest accounting it walked through the real back ends: the commit counts behind each agent, the API stacks they call, and the monthly run costs.

This is the part the demos skip. An agent is not a one-time build. It is a standing line item that calls paid APIs every time it works, and across 20 agents running more or less continuously, those calls compound into a real operating cost. The leverage is genuine. So is the meter.

The meter never stops

A piece of software you write once and run costs you mostly at the start. An agent inverts that. The build is the cheap part now, and the running is where the spend lives, because every task an agent completes is a fresh set of paid model and tool calls.

That changes how the cost behaves. It scales with usage rather than with headcount, which is the upside, but it also means a busy month is an expensive month, and a runaway agent is a runaway invoice. A loop that should have stopped, a retry that should have failed fast, an agent that calls an expensive model when a cheap one would do: each is a quiet line on a bill that nobody approved in advance.

This is why the teams getting real leverage are the ones putting controls around the spend before they scale it, with per-task budgets, caps, and a kill switch, rather than discovering the number at the end of the month. The agent that works is not the problem. The agent that works too much, on the wrong model, with no ceiling, is.

The 90/10 rule

The most transferable thing in SaaStr's account is a rule, not a number. It buys 90 percent of what it needs and builds only the 10 percent where no product exists. It describes vibe-coding an "AI VP of Marketing" not because it wanted to write software, but because nothing on the market did the specific thing it needed.

That is the opposite of the instinct the current hype encourages, which is to build everything yourself now that building feels cheap. SaaStr's experience says the cheap part is the trap. Creation is easy. Maintaining 20 bespoke agents is not. The discipline is to buy the commodity and build only the edge.

What still needs the three humans

The number that matters as much as the spend is the headcount that did not go to zero. Three humans remain, and they are not there to watch the agents succeed. They are there for the cases the agents cannot close: judgment, escalation, the relationships, and the moments where being wrong is expensive.

This tracks with what we found when we costed agentic commerce. The real cost of AI commerce is rarely the headline price of a model call. It is the integration, the oversight, and the failure handling the demo never shows.

The economics are not what the pitch implies

Put the pieces together and the agent economy looks less like free labor and more like a new category of infrastructure spend. There is a large up-front investment, a continuous metered cost, a maintenance burden that grows with every bespoke build, and a human floor that does not disappear.

None of this argues against agents. SaaStr clearly gets leverage from running a company this lean. The point is narrower: agents are paid infrastructure, not free employees, and like all infrastructure they reward the teams that treat the spend as a budget to manage rather than a magic trick.

What to take from it

Three lessons carry over from SaaStr's bill for anyone building the same way. Treat agent spend as a metered budget, not a one-time cost, because it recurs every time the agent runs. Buy the commodity and build only the edge, because maintenance, not creation, is where bespoke agents get expensive. And keep the humans for the expensive-to-be-wrong cases, because the floor is real.

We mapped the broader version of this in our state of the agentic commerce stack. SaaStr's contribution is to put real numbers on one company's version of it, which is rarer and more useful than another forecast.

The companies that win with agents will not be the ones that build the most. They will be the ones that read the invoice early and learn to manage it.

If your agents are paid infrastructure rather than free employees, what is the first budget line you would put them on, and who owns it when it overruns?

Charlie Major is a Product Development Manager at Mastercard. The views and opinions expressed in Major Matters are his own and do not represent those of Mastercard.