The Invoice That AI Didn't Warn You About

Andrew Avanessian, CEO at Haiilo.

​A piece published by a16z made a claim worth sitting with. George Sivulka’s “You Just Hired a Million Bad Employees” opened with a provocation that cuts directly to the economics of enterprise AI: “For the first time in history, humans are cheaper than software.”

AI productivity gains are real, but the costs of sustained usage are real, too. While your teams celebrate gains and your board asks about returns, invoices are quietly building in the background.​

Productivity illusions mask cost problems.

Pilots work. A team deploys an AI tool. A day-long process takes hours. The case study writes itself. I have seen this at Haiilo. Request for proposal responses once consumed days of specialist resources. They now take hours. I am certainly not questioning the gains, but what a pilot rarely surfaces is how cost structures scale with usage.

Traditional SaaS licensing is predictable. You pay-per-seat, and finance can model it. Token-based AI consumption is a different beast. It scales with activity. It’s what makes it so powerful, and it’s what makes it dangerous when underlying workflows are poorly designed.

The moment you embed an AI system into a high-frequency workflow, you are no longer buying a tool. You are commissioning a workforce that bills hourly, with no visibility into whether those hours are being spent productively.​

Agentic loops are the AI version of ghost spend.

Sivulka explores a concept that anyone running agentic workflows should understand: the loop.

Loops occur when an AI agent calls itself repeatedly to compensate for an inadequately specified task. The agent cannot complete what was required, as the task wasn’t articulated clearly enough for it to succeed. So it retries and retries, each retry consuming tokens and each token costing money.

The financial implications quickly become significant. If 80% of tokens are consumed unproductively, the economics of that deployment look nothing like the pilot stage models.​