Using 'generalist' intelligent agents can cost more than an employee

Three hundred dollars a day: the figure that sparked the debate. Jason Calacanis, a well-known tech investor and co-host of the popular business and technology podcast "All In", has publicly recounted - as reported by CIO.com - that his organisation has reached that level of daily spending by using Claude's Bees to power artificial intelligence agents and only to replace a marginal fraction of a developer's work. Chamath Palihapitiya, Ceo of venture capital fund Social Capital, added during the podcast that his company imposes token budgets on the best developers, but that unlimited usage can quickly become unsustainable. "If you aggregate the data on all people, a trend clearly emerges: at this point they have to be at least twice as productive as another employee," he explained.

The debate, which started at the podcast's microphones and then amplified on X and LinkedIn, quickly clarified a point that many IT departments still struggle to accept: the high cost of AI agents is not an inherent feature of the technology, but a direct consequence of the way it is used. "It's not an argument against AI agents," Ayesha Khanna, CEO of Addo AI, wrote on LinkedIn, quoted by CIO.com. "The question is whether you are using them in a way that makes economic sense. Most teams don't do that yet." Vygandas Pliasas, Ceo of Solidmatics, explained that letting expensive custom agents run freely via APIs - instead of using dedicated coding tools - is a recipe for cost drift. "Sometimes I get up to $500 in a week, and that's with a human-driven approach," he said. "If you let agents delete and rewrite code blindly, you're not really controlling the work anymore, and I would have serious doubts about the quality." His mantra: "AI is a tool, not a replacement for the brain. If you trust it blindly and let it run, you will achieve that. And the output won't be worth it."

At the heart of the problem is the economic structure of large language models: each interaction consumes tokens, and the bill goes up according to the length of the context, the number of steps, and possible retries. One solution, put forward by the experts, is to use AI agents with a circumscribed perimeter of action, combined with a smaller model running locally: a significantly cheaper solution. "The difference between an intelligent use of AI agents and an approximate one can easily be ten times the operational cost," clarified Kateryna Babenko, analyst at Katico. Babenko's operational recommendation, reported by CIO.com, is clear: model the volume of tokens for each task, benchmark the cost of inference at the level of the model used, and compare those costs with the cost of human labour in the flows you intend to automate. "The agents that justify their cost are those applied to jobs that nobody wants to do manually, where the task is bounded, the output can be verified, and a human can intercept errors before they propagate."

On a practical level, experts recommend setting hard caps: leading AI providers allow you to configure hard caps per API key or per organisation. Separate keys can be created per team or individual engineer, with a monthly limit. For managed tools - Cursor, GitHub Copilot, Claude's subscription plans - it is up to IT to decide who gets a licence and who does what. With one important caveat: if an engineer is producing high-impact features, perhaps his output is equivalent to what a team of ten would produce in a year. In that case, the cost is justified.

The message is clear: AI agents are not a fixed cost. They are an operational variable cost, with budgets, controls, responsibilities and ownership to be defined from day one. Those who do not understand these issues risk paying a hefty bill.