Without proper controls, AI agents can cost more than what outputs are worth according to Jason Calacanis and Chamath Palihapitiya, two IT experts and cohosts of popular podcast, All In podcast.
During a recent episode, long-time tech investor Calacanis noted that agent costs quickly rose to $300 a day while using the Claude API at one of his organizations. At the same time, these $100,000-a-year agents were replacing only a fraction of an employee’s work.
“When do tokens outpace the salary of the employee?” he asked. “Because you’re about to hit it. I’m about to hit it.”
Palihapitiya, CEO at VC firm Social Capital, said his organization sets token budgets for its best developers, but unfettered agent use can get expensive.
“If you aggregate it across all people, you can clearly see a trend where you’re like, ‘Well, now they need to be at least twice as productive as another employee,’” he said.
Limits are necessary, otherwise he’d run out of money, he added.
The podcast raised some eyebrows on X and LinkedIn, with some posters suggesting that AI company profits are too high. Others said the cohosts’ experiences show why organizations need to keep tight controls on agent spending.
“This isn’t an argument against AI agents,” Ayesha Khanna, CEO of AI solutions vendor Addo AI, wrote on LinkedIn. “The question is whether you’re running them in a way that actually makes economic sense. Most teams aren’t yet.”
Several other AI experts say that the $300 daily cost can be realistic if IT leaders don’t limit spending. But agents don’t have to cost nearly that much with the right controls in place, they say.
The cost of customized agents
Without controls, the agent costs can skyrocket when organizations use expensive and customized agent systems for coding, and run them through APIs instead of using coding tools directly, says Vygandas Pliasas, CEO of fractional CTO services provider Solidmatics.
“I can sometimes hit $500 in a week, and that’s with a deliberate, human-led approach,” he says. “The thing is, if you let agents wipe and rewrite code blindly, you’re not really doing the work yourself anymore, and I’d have serious doubts about the quality.”
Allowing an agent to take over coding and other tasks can lead to security and other risks, he adds.
“I always say AI is a tool, not a replacement for your brain,” he says. “You lead with your brain. You use AI to get answers faster, but if you blindly trust it and let it run, then you’ll hit that number. And the output won’t be worth it.”
Also, a $300-per-day cost can be misleading without context, adds Shahram Anver, CEO and data scientist at AI-driven site reliability engineering vendor Cleric. “That happens when you give a general-purpose model a broad task and let it run,” he says. “It’s like hiring a contractor without a clear scope and being shocked when you get the invoice.”
Cleric has built a site reliability engineering (SRE) agent for a fraction of the $300-per-day cost. It’s built for a specific purpose, and knows what questions to ask and when to stop.
In many cases, agents can be more cost effective than hiring more IT workers, despite the potential cost overruns, Anver says. “A senior SRE costs $200,000-plus per year fully loaded,” he adds. “If an agent can do 80% of their incident investigation work at a fraction of that cost, what matters is whether the output is good enough to trust and not what it actually costs.”
Still, it’s easy for organizations to get carried away if IT leaders don’t set limits, he says.
“Most companies are learning this the hard way,” Anver adds. “Someone on the team spins up an agent for a task, it works, they tell their teammates, and suddenly, you have 15 agents running without visibility into what they’re doing.”
The cheaper alternative
There are expensive ways to deploy agents, and there are cheaper ways, adds Kateryna Babenko, CX/CS software analyst at AI-focused chatbot and support-desk optimization provider Katico.
“If you’re running a persistent agent against a frontier model API, with high token consumption, long context windows, multi-step reasoning, and heavy output, the economics can get ugly fast,” she says. “In some cases, the cost per task can end up worse than just having a person do the work.”
However, a tightly scoped agent paired with a smaller, fine-tuned AI model running locally or on a controlled inference layer can cost a lot less, Babenko adds.
The two ends of the agent spectrum are completely different operating models, but many companies don’t distinguish between them, she says. “They hear ‘agent’ and assume one category of spend, but it isn’t,” she says. “It’s a wide spectrum, and the difference between a smart deployment and a sloppy one can easily be 10 times the operating cost.”
So IT leaders should model the token volume per task, benchmark inference cost at the model tier they’re using, and compare those costs against the labor cost of the workflows they want to replace or augment, she recommends.
“The agents that justify their cost are the ones applied to work nobody wants to do manually, where the task is bounded, the output can be checked, and a human can catch mistakes before they cascade,” Babenko adds.
IT leaders need to think about the economics of agents as well as the technology implications. “The organizations that run them well will be the ones that treat them like any other variable-cost operational resource with budgets, controls, ownership, and accountability built in from day one,” she says.
Capping agent use
Solidmatics’ Pliasas suggests budget caps on agent use. The major AI providers allow customers to set hard spending caps per API key or organization, he notes. IT leaders can create separate API keys per team or engineer, and set a monthly dollar limit.
For managed tools such as Cursor, GitHub Copilot, or Claude’s subscription plans, the organization’s IT team controls the licenses centrally and decides who gets a seat, what tier they’re on, and whether they get access to the basics or to premium models.
However, Pliasas also advices IT leaders to be flexible with budget caps.
“It depends on the value being delivered,” he says. “If an engineer is shipping high-quality, high-impact features, maybe the output is equivalent to what a 10-person team would produce over a year. In that case, the cost is justified and you really can’t compare it to a single person’s salary.”