The monthly bookkeeping now does itself
We got tired of the work, so we built something to do it. A full year of our accounts — 221 payments, 196 invoices — for ₹48. Now we build agents like it for other companies.
It was not hard work. It just never ended.
Every month the same thing. Download the bank statement. Two hundred rows of things like UPI-4021xxxxxx-PAYMENT FROM PHONE.
Work out what each one was. Find the invoice for it — in Slack, in email, in someone’s downloads folder. Match them up. Chase whoever forgot to send theirs. Then do it again next month, having forgotten everything you worked out last time.
Nobody enjoyed it. So we stopped doing it by hand.
You give it the statement. It does the rest.
One file, once a month. That is your whole job.
- 01
It works out what each payment was for
Teach it once — “anything with SWIGGY is Meals” — and it never asks again. Whatever your rules don’t cover, it asks an AI model. Anything it still isn’t sure about, it marks Needs Review and leaves for you. It never guesses.
- 02
It collects the invoices by itself
From your Slack channel. It notices when someone posts the same invoice twice, and quietly ignores the copy.
- 03
It reads them
Vendor, date, amount, GST number, tax breakdown. It reads a normal PDF for free. Only a photo or a scan costs anything.
- 04
It matches each invoice to its payment
On amount and date. If two invoices could fit the same payment, it doesn’t pick — it shows you both and asks.
- 05
It chases the missing ones
Every day at 4pm it posts one message in Slack: who still owes which invoice, tagged by name. It stops as soon as the invoice arrives.
Ledger
Every transaction in the book of record — click a row for its full story, invoice included.
Every payment, categorised.
Ledger
Every transaction in the book of record — click a row for its full story, invoice included.
CLAUDE.AI SUBSCRIPTION
Open a row and the invoice is sitting there, with what decided the category.
It chases people so you don’t have to.
“Which payments still have no invoice?” It answers, and shows why.
There is a chat inside it. You ask in ordinary English, and it answers — and then shows you the actual rows it used to work that out, printed underneath.
That last part matters. You never have to take a number on trust.
It can only read. Ask it to change a category and it says so plainly, and points you at the button that does it.
Ask
Questions about your books, answered from the actual rows — every figure comes with the payments behind it. Reading only: nothing here changes the books.
We spent Rs 288,197.89 on AI Tools this year across 36 payments.
Total spend for 2025-04-02 to 2026-03-31: Rs 288197.89 across 36 payment(s). (The row list shows 36 example payment(s).)
Not a demo. Our actual books.
What it did on our own accounts, April 2025 to March 2026.
This is not a demo built for a website. It is the system our company’s books actually run on, every month — and we are still improving it, because we use it ourselves.
- 221
- bank payments handled
- 196
- invoices collected and read
- 69%
- read free, with no AI needed
- ₹48
- total AI cost, the whole year
AutoLedger running cost (all time)
deepseek-v4-flash
calls : 443
tokens in : 373,062
tokens out : 563,305
cost : $0.3525 (approx Rs 33.75)
claude-haiku-4-5
calls : 58
tokens in : 92,740
tokens out : 11,428
cost : $0.1499 (approx Rs 14.35)
TOTAL: 501 calls -> $0.5024 (approx Rs 48.11)A year of bookkeeping, for less than a cup of coffee.
Three things mattered. None of them was the AI.
These are not bookkeeping ideas. They are how we think any AI agent touching real money, or real decisions, ought to behave.
Never guess
An AI that is 95% right sounds great until you realise 5% of your books are quietly wrong and nothing tells you which 5%. Ours would rather stop and ask. Everything it isn’t sure about goes in one place.
Never pay twice
Every step can be run again safely. Import the same statement twice and nothing duplicates. Re-run a month and it doesn’t re-buy work it already did. That is why a year costs ₹48.
Show your working
Every category in our books records what decided it — your own rule, or which AI model. Months later you can ask “why is this a salary?” and get a real answer.
Your problem isn’t bookkeeping. It may be the same shape.
Ours looked like this:
- it came back every month
- it was mechanical, but it needed judgement
- the information arrived messy, from several places
- getting it wrong was expensive — and quiet, so nobody would notice for months
- and part of it was chasing people
We have only built this for bookkeeping. But other problems have the same shape — you may recognise it in insurance claims, in purchase orders checked against what actually arrived, in compliance filings, in stock counts.
Our own bookkeeping is done by an AI agent we built. An agent, in plain words: not a chatbot. Software that does a whole job on its own — and stops to ask you when it isn’t sure.
Now we build agents like it for other companies. We build products too — ChatPlotDB lets you chat with your database in plain English.
What this looks like for you
The monthly job nobody wants
Matching things up, making reports, chasing people, filing.
Reading documents
Invoices, bills, statements, forms. Getting facts out of PDFs and photos, reliably, without paying more than you need to.
Chat over your own data
Ask a question, get an answer with the rows behind it.
Agents where your team already is
AutoLedger lives in our Slack; nobody had to learn a new tool. Putting one where your team works is a small piece of work.
We offer to build AI agents like this for your company.
How we work: we start with one job that’s costing you time every month, and build it properly — tested, and safe to run again. We agree who owns what before we start, in writing.
