What they do and how
The agents read supporting documents, propose an account and VAT treatment, match bank transactions to documents, review VAT before filing and post entries in batches. The author gives tasks in natural language (“see what is still unmatched in the bank”), the agent prepares drafts in a queue, he reviews and sets limits (“do not post supplier invoices, only receipts”), and then the agent posts and rereads the ledger to check. Accounting policy decisions, such as where a cloud expense goes, are made by a person once and the agent applies them to later rows.
Reported result
Stated by the author, with data from his sessions between 15 July and 25 September 2026: 53 sessions with Codex (about 111 active hours) and 16 with Claude Code (about 25 hours), with batches of 195, 291 and 582 entries and about 1,570 in total in Claude Code. One agent caught an error in a VAT return days before the deadline and another discovered bank rows imported twice. His estimate, which he himself calls approximate, is that a weekend of work became an afternoon. Part of the volume was old backlog, and many entries were simple.
Stated figures. An identified person or organization states that they have used it. The results are stated by the source; we have not audited them.
What has really been tested
The author did it himself with real companies, according to the logs he cites. Some figures were counted by the agent. The text is published by the vendor of the tool the agents use, so it should be read as an interested testimonial. He acknowledges failures: an agent posted 14 invoices to the wrong account, added test entries to the live books, treated rows that were only linked to drafts as posted and called a payroll import “verified” that landed in the wrong month. We have tested none of this.
How to replicate it
The author recommends starting in a test company, asking first only for read-only access and proposals, setting the rules before posting anything (series, recurring expense accounts, what to expect), letting the agent prepare and a person approve, asking for a reconciliation read from the ledger instead of a summary and widening the scope once a type of row has gone well several times. Include a duplicate check as a fixed step before posting and verify yourself everything that goes to the Swedish tax authority. The regulatory part is specific to Sweden.
What to keep in mind
- Responsibility for the books being correct stays with the company, not the agent. The text frames this for Sweden; in other countries you will need to check local rules.
- The data the agent reads is sent to its AI provider (Anthropic or OpenAI). If it includes personal data, review data-protection rules such as the GDPR.
- A document can contain hidden instructions to divert the agent (prompt injection); that is why the author prefers the agent to prepare in a queue and a person to approve.
Sources
We wrote this entry in our own words from the sources above and checked that the links worked on October 17, 2026. Is anything incorrect? Tell us.