What does it actually cost to run AI bookkeeping bots?

Last month my firm's total AI bill was about $2,700 — and that covered AI agents working real client books across 77 ledgers, plus building an entire software platform on the side. The floor is much lower: one $200/month subscription is enough to put AI to work on a real book of clients. The trap most bookkeepers fall into is paying per-task API pricing for work a flat subscription would cover.

I run a bookkeeping firm built for real estate investors, and this isn't a review of tools I demoed once. These are the numbers off my own bill, because when I showed this setup to a group of fellow bookkeepers, nine of the first thirty-six questions were about cost. So here's the whole answer.

The two ways you pay for AI

Everything below is about Claude (Anthropic), which is what I use, but the shape is the same everywhere:

Subscriptions (flat monthly). You pay $20–200/month and get a usage allowance that resets on a schedule. The $200 Max tier is the workhorse: it covers an enormous amount of real work, and your worst-case bill is known before the month starts.

API (pay per use). You pay for exactly what you consume, metered by the token. It's how software talks to AI directly, it's what most "AI bookkeeping" products are quietly built on, and it's the meter that runs up bills when nobody's watching.

My rule after months of running both: do your working on subscriptions, and reserve the metered API for overflow and special jobs. The same workload that fits comfortably inside a $200 subscription can cost multiples of that if you pay for it by the token.

What my real bill looks like

• Claude subscriptions (the fleet's engine): $200 each — I run three for different parts of the operation
• API usage (overflow + batch jobs): variable — the swing line
• Website + dashboard hosting (Vercel): $0
• DNS + data infrastructure (Cloudflare): ~$5
• Last month, all-in: ~$2,700

Two honest notes on that number. First, it was a heavy month — the agents were doing year-end catch-up work *and* building software. A steady month runs far less. Second, scale is what makes it small: spread across the books my fleet touches, it's a fraction of what one part-time human doing the same volume would cost.

One more number, because it surprised me: when I tested batch processing (sending transaction categorization to the AI in bulk, overnight, at half price), the cost came out to about $2.86 per 1,000 transactions — with the model flagging the genuine judgment calls for a human instead of guessing. That's not a typo. Categorization is no longer the expensive part of bookkeeping.

Where the money actually goes

The spend divides into three jobs, and they cost very differently:

1. Reading and thinking — reviewing a bank feed, researching a client question, planning a cleanup. Cheap, and getting cheaper.
2. Doing — the bot clicking through QuickBooks Online, pulling statements, attaching receipts. This is where volume lives, and it's the reason subscriptions beat the API meter.
3. Judgment calls — is this an owner draw or a loan repayment? The best models cost more here, and it's worth every penny, because this is where errors get expensive. We don't guess, and we don't let the AI guess either.

For scale on what that buys: one client — a 130-property, four-entity portfolio — handed us a 30,514-document archive. The AI didn't read all of it, and that's the point: the books only need the slice that maps to actual transactions, and the AI knows which 3% matters. In the same portfolio, 1,683 transactions were rebuilt from 26 bank statements with every month tying to the penny. The AI cost of that work was noise compared to the human hours it replaced.

What it costs to start

You don't start at $2,700, and you shouldn't.

• Prove it to yourself: $20/month. One entry-level subscription, one messy workflow — receipt renaming is a great first victim.
• Put it to work: $200/month. One Max subscription runs real daily work on a real client list. Most solo firms can live here for a long time.
• Run a fleet: $400–700/month fixed, plus usage. Multiple subscriptions, scheduled jobs, agents working while you sleep. This is where I live, and the bill still costs less than one part-time hire.

FAQ

Isn't the API cheaper since you only pay for what you use?

Only at very low volume. Once AI is doing daily work, a flat subscription covers what would be hundreds or thousands of dollars of metered usage. Measure one week of API spend and compare it to $200 — the answer will be obvious.

Can I really do this with one subscription?

Yes. I ran real client work on one Max plan before I added more. You add subscriptions when the work outgrows the allowance — a good problem, arriving later than you'd think.

Do the costs grow with every client I add?

Not linearly, and that's the whole point. My fixed cost is the same whether the fleet touches ten books or seventy-seven; only heavy overflow moves the needle. Software scales; hourly labor doesn't.

What about all the other tools — dashboards, hosting, automation?

Nearly free. My client dashboard and website are AI-built and host for $0 on Vercel; Cloudflare runs about $5/month. The AI subscriptions are the only bill that matters.

I teach this whole setup — the cost routing included — inside the community: founding membership, locked for life.