What does AI actually cost a small business?

The question

What we told them

You've met both halves of AI pricing, and only one of them was on the quote.

Half one is seats — a fixed price per user per month, like your office suite. Predictable, budgetable, boring in the good way. Half two is usage — the work the AI actually does is metered in tokens, and billed like electricity. Every document processed, every automation run, every agent step draws from the meter. That's the line that surprised you.

Usage pricing isn't a trick — it's often the cheaper model, because you pay for work done instead of licenses parked on desks. But it misbehaves in ways subscriptions can't: a busy month costs more than a quiet one, and a badly supervised automation can burn money silently. The failure mode we warn every client about is the retry loop — an agent that hits an error, tries again, and again, at full token price, all night. That's how an AI assistant ends up costing more than the employee it was meant to help.

So budget AI the way the industry learned to budget cloud a decade ago:

• Hard caps at the provider. A spending ceiling that stops the meter, not just a notification.
• Alerts well below the caps. You want the warning at 60%, not the outage at 100%.
• Attribution by workflow. Know which automation earns its cost and which one just eats. If you can't see spend per use case, you can't manage it.
• A monthly review. Ten minutes with the usage report. The tooling market has noticed this gap — spend-tracking products for AI are now launching — which tells you how common surprise invoices have become.