How to automate bookkeeping with Jev
A startup founder gave Jev 34 months of books that a bookkeeping firm had charged over $20,000 for, and had it redo the work and check the firm’s. It took 20 seconds. Here’s what he built, where to find it, and a prompt that builds the same review on your own books.
- months of books
- 34
- start to finish
- 20s
- total cost
- 32¢
From @andywang’s public demo
The code: not published yet
There’s no repo for this one. Andy Wang, who founded the accounting startup Finta, built it over a weekend. In his replies he offered to open-source it if people bring their own Jev key, but he hadn’t by September 28. If he does, I’ll put the link right here.
Until then there are two ways to get it. Build your own with the prompt below, or use Finta, his paid bookkeeping product, which he says now uses Jev for its first pass at categorizing transactions.
What the demo did
It read every table in a QuickBooks account, put together its own set of books, then compared its work with what the human bookkeepers had done, using the audit log. In his replies he lists the jobs it ran side by side:
- the category and department of every transaction
- journal entries for amortizations and accruals
- invoices that were never collected and should be written off
- reconciliations
Among the mistakes it caught: a SAFE investment booked as revenue, and a sales-coaching contractor filed under the wrong category.
“Did a better job” is his claim, not a benchmark. The $20,000 and 32 cents are in his caption; on screen, a counter shows over 10,000 decisions reviewed in 19 seconds.
The part worth copying: it knew when to stop
It didn’t try to force an answer on the messy transactions. When there wasn’t enough documentation to be sure, it marked them “should’ve asked more questions” and moved on.
The same pattern with Jev
Jev takes a situation and answers small typed questions about it, with a probability on each. For bookkeeping, one call covers one transaction:
- The state is one transaction written out plainly: date, amount, vendor, memo.
- The questions are yes/no, all asked in the same call: one per candidate category (“This belongs in Software”), plus “this is an investment, not revenue” and “there’s enough here to be sure.”
The highest category wins. If the “enough here to be sure” score is low, or two categories are nearly tied, the transaction goes to a person. It’s the same shape as the prospecting guide, with a transaction in place of a post.
New to Jev? Start with the setup guide: key, first call, working response. Then come back.
Let an agent build it
Paste this at Claude Code or Codex. It reviews your books and your bookkeeper’s choices, and it never changes a thing. The thresholds are starting points I haven’t run yet, so check them against a month you already know is right.
Build me a bookkeeping review job using Jev through the Vercel AI Gateway. Read-only: it never writes to my accounting software.
1. Ask me where my books live (a QuickBooks or Xero export, or a CSV from my bank) and pull every transaction: date, amount, vendor, memo, and the category and department my bookkeeper chose.
2. Ask me for my chart of accounts, then pick the 10 or so categories that cover most of my spend.
3. For every transaction make ONE call: experimental_evaluate from the ai package, model typesafe-ai/jev. The state is the transaction as plain text (date, amount, vendor, memo). Ask boolean questions together in that one call:
- one per candidate category: "This transaction belongs in <category>."
- investment: "This money is an investment or loan (a SAFE, equity, a note), not revenue."
- documented: "There is enough information here to categorize this transaction with confidence."
4. The top category is the pick. If "documented" is under 0.50, or the top two categories are within 0.15 of each other, mark it NEEDS A HUMAN instead of guessing.
5. Compare each pick with what my bookkeeper chose. Write review.csv (date, vendor, amount, bookkeeper category, Jev category, confidence, verdict), with disagreements first.
6. Run 3 calls at a time; bursts get rate-limited. Tell me the total cost when it finishes.
Show me the disagreements before anything else. Never change a transaction yourself. I decide what to fix, with my accountant.The honest part
I haven’t run this one. The numbers are from @andywang’s post on September 22, 2026, and there’s no code on this page because I only publish code I’ve run. The prompt carries his design: judge everything, and flag what isn’t documented.
Use it as a second opinion, not as your bookkeeper. It reads what’s in the export. It doesn’t know about the handshake deal or the receipt in someone’s inbox. Take the disagreements to whoever signs your taxes before you change anything.
Don’t have Jev running yet?
The setup guide is one key and a few lines of code, and you’ll have a working response in a few minutes.
Get started with Jev →More field guides
All guides →
Jev · GuideHow to set up Jev and make your first API call →Step by step, about 5 minutes
Jev · GuideHow to automate SEO internal linking with Jev →The script, the prompt and the threshold rule
Jev · GuideHow to find sales prospects ready to buy with Jev →Two scripts and the three filtering questions
Jev · GuideHow to filter Amazon products fast with Jev →The agent browser, the rules as questions, the raceJev · GuideHow to search Zillow by what the filters can’t see with Jev →699 listings judged in under 18 secondsJev · GuideHow to check legal citations and research state law with Jev →LegalJev (free) and Monid, side by sideJev · GuideHow to build a Clay-style lead list with Jev →Treg on GitHub: judge first, pay secondJev · Directory130+ Jev demos, repos and docs by team →Everything worth watching or running
Higgsfield · GuideHow to swap the people in any video with Higgsfield Genjutsu →The clip, the photos, the prompt and the cost