One afternoon we flipped the order of how our chat gets answered, three times in a single sitting, and by evening the whole thing was running backwards from how everyone tells you to build it. Where before each question went straight to the big famous expensive brain, now it went somewhere humbler first. A small, modest helper that lives in our own house and runs on little more than electricity. Only when that helper stumbled did the question get handed up to the pricey one. We had taken the part that decides who answers and quietly turned it on its head.
If you read the brochures, this is the wrong thing to do. The frontier brains are smarter. The little local ones are, well, little. The obvious move is the expensive one, and the expensive one even comes with a nice story for customers. We use the best. So why send the everyday chatter to the modest helper running in the back room first?
Because best is a measurement of the ceiling, not the floor. Almost none of what people actually ask is a ceiling question. Most of it is what does this button do, can you summarize this, say that again but shorter. A modest helper handles those without breaking a sweat. The expensive brain's real superiority only shows up on the hard sliver of questions, not the easy mountain of them. We had been paying a brain surgeon to hand out plasters.
The shift in our heads was this. Deciding who answers is not pick the best one. It is start with the cheapest thing that actually works and climb up only when you must. You stop asking which helper is best for this and start asking what is the cheapest one that succeeds here, and how do I spot a failure fast enough to climb the ladder before the customer ever notices.
A hospital already works this way at the front desk. They do not put every single person through the big expensive scanner. They start with the nurse out front. Most folks are sorted right there. The scanner is for the cases the nurse cannot handle, and its value is that it exists for those, not that it sees everyone who walks in.
So we built a ladder. The modest helper first, then the famous brain, then the cheaper backups below that, each waiting its turn. A quiet check runs in the background to make sure the modest helper is healthy, and if it ever falters the next rung takes over on its own without us lifting a finger. The next morning we added one bit of common sense. Any question too big for the small helper to hold in its head skips it entirely and goes straight to the bigger one, because we would rather pay than give someone a half-answer. None of this is a clever trick. It is just a short, sensible list of who goes first. The whole change came down to this. Most of our chat now costs us essentially nothing per answer, and the ladder quietly catches the handful of cases where nothing is not enough.
For a small business stitched together out of paid services, the same idea applies in corners that have nothing to do with AI. Every service in your stack has tiers. Every tier has a price per use. And almost everything you do is a routine case the cheapest tier handles fine.
The trap is paying top-shelf prices for ground-floor work. You buy the highest tier so it is there when you genuinely need it, and then most of the monthly bill turns out to be the boring middle, the cases a far cheaper option could have served at a fraction of the price. It all hides inside one tidy round-number invoice, and you stop seeing the waste.
The move is to write down the ladder. For each thing you pay for, what is the cheapest version of it you control? A free tier, something simple running on a small box you already own, a feature buried in a tool you already pay for. Try that first. Catch failure cheaply. Climb to the paid option only on an actual miss, not on a hunch about quality.
You will not get rid of the expensive option. You need it standing by for the hard cases, the way you keep a senior person on call even when most of the day is password resets. But you should be paying them for the hard cases, not the easy ones. The bill should match how hard the work actually was, not how good the brochure looked.