McKinsey puts one agent task's run-to-run cost spread at 30 times

Lari Hämäläinen's rule of thumb sets the bar low. Say a task takes a human an hour and the agent's work can be verified in six minutes. In that case, he said, an agent with a success rate above 10 percent could already begin to create value ^[8]. One success in ten attempts means ten rounds of six-minute checking, which is sixty minutes, the same hour the human would have spent ^[14]. The threshold sits at break-even in human minutes, and the tokens burned by the nine failed attempts are not in that comparison ^[14].

Those tokens are the variable part. Hämäläinen said agents can take different paths to the same result, making costs highly variable, and that the same task can cost up to 30 times more from one run to another ^[5]. Much of that cost comes from the reasoning and repeated refinement behind the final output, he said ^[5].

The per-unit price is doing what the vendors say it is doing. "Intelligence at a certain capability level is getting a lot more affordable," said Hämäläinen, a leader in McKinsey Digital ^[2][13]. Models performing at roughly GPT-4's level on established benchmarks can be run for a fraction of what they cost when it launched in 2023, and companies are still spending more and more ^[3]. Hämäläinen's account of the gap has two halves: enterprises ask models to perform vastly more reasoning and work, especially through autonomous agents, and AI vendors are capturing some of the efficiency gains through higher margins ^[4]. In software development, he said, agents can repeatedly inspect, modify and rewrite entire codebases, generating far more code than a human developer would typically touch ^[6].

"The truth is, there is no single cost lever," said Tanguy Catlin, a director of the McKinsey Global Institute ^[10][13]. He named three areas. The first is visibility into which use cases, business units, agents, models and users drive spend. The second is workflow optimization, including matching model complexity to the task, routing requests, caching reusable context and limiting unnecessary tool calls and agent loops. The third is sourcing discipline, meaning removing unused licenses, managing quotas, negotiating provider terms and avoiding excessive dependence on one model or vendor ^[11]. All three are internal work. The budget line they grow is engineering and procurement time.

The case against planning for growth rests on that second lever. Caching, routing and shorter agent loops all cut tokens per task, and if they compound faster than task volume rises the total invoice can flatten while usage climbs. Testing that needs spend data, and the Fortune account of the session does not include a dollar figure for enterprise AI spending or a direction for AI budgets ^[15]. In my view the growth case holds for the next few budget cycles. Verification capacity is what limits how much agent work a team can absorb, and Hämäläinen said the bigger challenge is redesigning the surrounding workflow to take advantage of the capacity an agent frees up ^[9]. Catlin cautioned that the goal should not be indiscriminate cost-cutting, and that companies should determine where AI spending delivers the highest returns and optimize toward those returns ^[12].