In one 48-hour stretch this month, the price of running AI dropped in public. Meta shipped its first paid model at $1.25 and $4.25 per million tokens — the chunks of text these models read and write. GPT-5.6 landed with a bottom tier of $1 and $6. Grok 4.5 came in at $2 and $6 and beat Anthropic's flagship on the benchmark its own maker chose to headline. Production intelligence now has a price sheet, it keeps falling, and it's quietly moving the money in accounting from the monthly close toward AI accounting setup.
That's the engine behind the $49 close. When categorizing a month of transactions costs pennies, someone will sell the month-end close for the price of a streaming subscription. The Squeeze series spent three parts on what that does to your fees, and its final part — on why you should stop selling compliance on its own — closed on the one thing the compression can't reach: getting the system set up right in the first place. This is that piece, the payoff to that promise.
Here's the part the $49 headline hides. The cheap close assumes a clean, well-built system underneath it — and almost no client's books are that.
The vendors won't hold a flat price
Look at what the people selling the automated close actually charge. Pilot spent nine years running its own bookkeeping firm and built an internal engine to close the books. It's now selling that engine to other firms as Meridian, running on top of QuickBooks. The price isn't a flat number — it's per client, billed monthly, "depending on the complexity of the business," and the spread between their simplest client and their most complex runs about sixfold.
They can't even quote it up front. They run an easy file, a medium file, and a hard file, take the average, and price from there. Their own summary of the ceiling: the automation is "as good as the information provided," and anything missing or inconsistent gets flagged back to a human.
The thing that varies from $49 to six times $49 was never the intelligence. It's the state of the system underneath — how the chart's built, whether the feeds are clean, how much of the month is genuine exceptions. A flat price for the close is a fiction the sellers themselves refuse to sign.
Three ways a client lands on your desk without a working system
Clients arrive needing the build for one of three reasons. Some were never set up properly — a ledger that was never really a system, no chart design, no app stack, no rule for what to do with the odd transaction. Some were set up once and then left to rot — usually the owner doing their own books, bank recs never finished, a bank feed that quietly dropped months ago and nobody noticed, coding that means something different every quarter, year-end adjustments never posted. And now a third: the client whose AI did the books, at speed, unsupervised.
The middle one predates AI by decades. The third is the new accelerant — as "my AI handles the books" spreads, more unsupervised files pour into the same funnel, faster than their quality can keep up. You won't know which kind of file you've got until you've looked. All three land in the same place: nobody can run a cheap close on a system that doesn't exist yet.
AI does the labour — not the design, not the name on the file
Be honest about what AI can already do here, because the profession keeps picking losing fights. AI does the mechanical cleanup: recategorizing, deduping, reconstructing. Intuit and Digits are shipping onboarding automation right now. Meridian itself, by its makers' own description, is deterministic software with a person watching the exceptions — "a glorified staff accountant" doing last-mile categorization on a system somebody configured. Deterministic just means it either does the step or flags it as unknown; it isn't guessing its way through your client's books.
The labour is going to get cheaper, like everything else. What doesn't get cheaper is deciding what this specific business's system should be — the chart, the app stack, what data you scope in, which exceptions a human has to see — and putting your name on that foundation. That's design, and it's accountability. It's the same argument the Squeeze made about who does the work — that when the mechanical labour commoditizes, the accountable human is the product that survives — pushed one layer down, to the foundation itself. Neither the design nor the accountability is on anyone's price sheet.
Everyone's automating the close. Almost no one sells the build
Watch where the money is racing. The native AI ledgers — Digits, Kick, Puzzle — and the productized closes like Meridian are all fighting for the run-time, the monthly close itself. The tools will automate pieces of the setup along the way. But no tool decides what a given business's system should look like, because that's judgment, and judgment doesn't productize.
That makes the build the least-contested ground in the whole market — and the most durable. The open lane and the safe lane are the same lane.
So stop pricing the setup as throat-clearing before the "real" monthly work begins. The build is the real work — a distinct, front-loaded, project-priced engagement, not a discount you eat to win the recurring fee. And before you quote a cleanup blind, look at the file first. Cleaning up after a system that failed — or one an AI quietly broke — is a different, harder job than the original bookkeeping ever was, and it's the one this series prices next. What's the honest price of the system your next client doesn't have yet?
If you're looking at a file right now and you can't tell whether it's a clean system or a cleanup, that's the conversation worth having before you quote. Book a free consultation at theaiaccountant.ai/consultation, and we'll scope where your setup and cleanup work should actually be priced.