Il Sole 24 Ore puts its finger on something that anyone building AI workflows inside a company already knows: a generalist agent in production can cost more than a human resource. It isn't consultant hyperbole, it's the sum of line items nobody puts in the BOM when signing the pilot. API calls that explode the moment the agent goes into a loop on ambiguous tasks. Multi-agent orchestration that multiplies the calls for every step. No reliability metrics, so nobody knows how much of the cost is value produced and how much is noise.
The piece doesn't put a single number on it, and rightly so: it depends on the workload, the model, the degree of supervision. But it closes a window many CFOs were keeping open for convenience, the price-per-token as a proxy for annual cost. The TCO of an agent in production includes refactoring when the provider changes its pricing or retires a model, and a dedicated person monitoring it. On my desk, on real workflows, these line items weigh more than the API bill.
Why this matters for anyone building enterprise AI: without spend guardrails and reliability metrics, the agent isn't a saving, it's an opaque cost center.