How you waste money on tokens (and how routing fixes it)

Quick bit of jargon, explained plainly, because I promised myself I'd never use a term on this site without doing that: a "token" is roughly a chunk of a word. AI models charge by how many of them go in and come out. That's it. That's the whole mystery the pricing pages don't want to make obvious.

Here's the bit almost nobody selling "AI automation" to local businesses tells you: not every job needs the expensive model. Most don't. And running everything through the most capable (and most expensive) model available is the single most common way small businesses quietly overspend on AI without noticing.

The grunt work doesn't need a genius

Summarising a call transcript. Drafting a first-pass caption for a reel. Checking whether an email address bounces. None of that needs the equivalent of hiring a senior consultant for twenty seconds. It needs someone competent and fast — a cheaper, sometimes local, model does this perfectly well, at a fraction of the cost.

Save the expensive model for the decision, not the drafting

Where the good model earns its keep is the bit that actually matters: deciding which hook to lead with, whether a customer email needs a careful reply, whether an angle is worth pursuing at all. That's the "specialist staff" — the everyday admin work is the junior team, and you don't put your most expensive person on data entry.

I run this myself across several model providers rather than being locked into one vendor's tooling — not because switching providers is fun (it isn't, particularly), but because being able to route the cheap stuff to the cheap option, and only escalate to the expensive one when it's worth it, is the actual saving. Multi-model routing sounds technical. In practice it's closer to "don't pay a specialist to file paperwork."

Why this matters to you, not just to me

If someone's pitching you "AI automation" and can't explain, in plain terms, why their approach doesn't burn money on every single request — that's worth asking about directly. It's a fair, simple question, and a vague answer tells you something useful.