McKinsey's math on AI agents: the token meter shows your bill, not your return

A new McKinsey QuantumBlack analysis this week takes apart the economics of AI agents and lands on a blunt conclusion: counting tokens has stopped being a useful way to understand what enterprises actually pay. An agentic task can consume on the order of a thousand times more tokens than a simple chat query, and roughly sixty percent of that spend goes not into the first answer but into the retries and refinement behind it. Two identical requests can ring up wildly different bills depending on the path the agent takes.

The deeper point is that tokens are the invoice, not the payoff. Volume tells you how hard the machine worked; it says nothing about whether the work was worth doing. As agents scale, the firm argues, the discipline that matters shifts from experimenting with the technology to proving it is financially sustainable.

That split - govern the cost on one side, prove the value on the other - is the whole design of Uprovd. We track what agentic workloads spend and attribute it to the outcomes they delivered, so unpredictable token bills become a managed line with a return attached rather than a surprise on the monthly statement.