The great AI cost collapse, and how a small business rides it

The most important AI story for a small business isn't any single model release. It's a curve: the price of getting frontier-grade work out of an AI system has been falling by roughly an order of magnitude a year, and it hasn't stopped. Tasks that cost a dollar to automate two years ago cost cents now. Tasks that were laughably uneconomic — reading every supplier email, drafting every quote, summarizing every call — quietly crossed the line into "cheaper than not doing it."

Why the curve matters more than the headlines

Big companies respond to this curve with strategy decks. A fifteen-person business gets a sharper deal: you can simply re-ask an old question every quarter. The question is, "what did we decide wasn't worth automating last time we looked?" At the old prices, the answer was probably correct. At today's prices, some of those answers have flipped — and nobody sends you a memo when they do.

The pattern we see across small companies is consistent: the first automation that pays for itself isn't the flashy one. It's a boring, high-volume, low-stakes task — categorizing inbound email, first drafts of routine replies, turning call notes into CRM entries. High volume is what lets a falling per-task price compound into real money.

A rule of thumb you can run on a napkin

Take any repetitive task in the business and estimate three numbers: how many times it happens a month, how many minutes a person spends on it, and what an error costs when one slips through. If the monthly minutes are in the thousands and an error is annoying rather than dangerous, it's a candidate now. If an error is expensive — money moves, a customer commitment is made, something legal is signed — the task isn't a candidate for full automation at any price. It's a candidate for a draft-and-review setup, where AI does the typing and a person does the deciding.

The move worth making this quarter

Pick one high-volume, low-stakes task and run a two-week trial with a person still in the loop. Measure minutes saved and errors caught. If the math works at today's prices, ship it. If it almost works, put it on the calendar to re-check in six months — on this curve, "almost" is a timing problem, not a verdict.