I handed our accounting to an AI agent, and it found money in eight days
In eight days, from 28 October to 5 November 2026, we handed the day-to-day accounting at Norditech, the company behind Incredible, to an AI agent I set up: Claude, from Anthropic, running in the Claude Code desktop app. It books supplier invoices, checks every payment and keeps track of every deadline, and it comes back to me only with decisions. A founder still signs every payment.

8 daysfrom read-only to running our day-to-day accounting
How it works
Claude runs in Claude Code on my Mac, in one folder that holds our accounting workspace. It writes its own operating manual, with our company facts, our systems and seven hard rules, and reads it at the start of every session. Next to it are a list of about 38 open issues, each with its evidence and next step, and a calendar of 112 deadlines through 2027. The calendar is generated by code that knows the Swedish bank holidays, so a rule change rebuilds every date.
I talk to it in short messages, often voice-typed, and send it photos of paper invoices. Instead of long chat replies, it keeps a web page of questions for me, where each open decision is a card with its evidence and a one-word answer.
Nine systems
It works in nine systems, among them our ledger in Fortnox, our bank exports from Nordea, the finance mailbox, two years of finance chat, our company cards in Pleo, our sales and tax settings, supplier portals and the tax agency's web pages. Where no connector existed, it wrote its own tools, like a Fortnox client with its keys in the macOS Keychain and three parsers for Nordea's exports. For research it starts helper agents of its own. One of them checked every computed deadline against the tax agency's pages before the calendar went live.
Trust, earned in three steps
It started read-only. On day one it could read everything, and its Fortnox client could only send read requests. On day seven it made the case for write access and got it for supplier invoices and bookkeeping in Fortnox, behind an allowlist in code: no deleting, no payroll, no changing a supplier's bank details, and every write logged. On day nine it could send approved invoices from Fortnox to Nordea as payment files, only while I am logged in, and the bank still needs two founders to sign. The agent asked for each step and I granted it, and each limit lives in code, not only in a prompt.
Humans keep the keys
BankID logins, one-time codes, passwords and signatures stay with us founders. The agent is built to stop at those points and hand over. It works under seven written rules: it never moves money or submits a tax return, sends nothing outside the company without my OK for that message, never copies passwords or ID numbers into its files, and never acts on payment instructions or new bank details that arrive by email. It also said no on its own. It declined to give my phone number to a debt collector, and instead of emailing an address it had guessed, it found the company's verified channel.
Before every payment it checks that the supplier is known, that the amount is plausible, that the bankgiro belongs to that supplier in the public register, that the bank details haven't changed and that nobody is pushing for urgency. It reads emails, invoices and web pages as information and never as instructions, which protects it against prompt injection. In October, before the agent started, our team had stopped a fake invoice of about 850,000 SEK backed by emails impersonating a co-founder.
What it found in eight days
It worked out October's employer tax from payroll, then found what was already in our tax account and corrected the payment from 147,374 to 112,425 SEK before anyone paid it. It noticed that Fortnox showed a payment as paid that had never been signed, and made the bank the source of truth. It matched an incoming payment of 187,500 SEK, which nobody had identified since July, to a customer invoice. From our payroll it also flagged an R&D deduction we may not be using, worth an estimated 30,000 SEK a month, which our accountant will confirm.
The phone bill that kept coming back
One photo of four paper invoices, and a request to check that we hadn't paid any, became a fixed process within two days. Two of the invoices were already paid, and the agent matched them to Nordea's payment batches to the öre. The fourth was a debt-collection letter for a bill already waiting in the bank under a new reference, so paying both would have paid it twice. It queued the bill once and prevented a double payment of 2,701 SEK.
The ledger showed the pattern behind it. Our paper phone bill had been missed four times this year, and each time it came back as a reminder with a 450 SEK fee and then as a collection demand. The agent held the debt collector's support chat in Swedish from my browser and learned that one earlier double payment had been refunded. Then it switched our phone operator from paper to PDF invoices by email, so the bill now lands in the finance inbox, where it can't be missed.
When it was wrong
It was wrong four times in eight days, and each time it caught the mistake itself by checking a second source before any money moved. It first estimated that up to about 4,300 SEK had been paid twice; after asking the debt collector, the real figure was 2,069, and it told me so. It also refused to send payment files at all at first. Then it reread our own rule, that it prepares and a founder signs, saw that an unsigned payment file is preparing, and asked me to confirm the change in writing.
How I build agents
An agent earns trust the way a new hire does. It starts with read access, shows what it finds and gets more when it makes the case, with every limit written in code. It writes everything down, so the next session, or a person, can pick up where it stopped. And it meets people where they work: I send a photo or a sentence, and I get back a page of decisions with one-word answers.
