Keep company books with terminal coding agents
Can AI do your bookkeeping? Task: Can AI do your bookkeeping? Much of it. What ten weeks of keeping Swedish companies' books with Claude Code and Codex showed, and where a person is needed. By Philip Eriksson · Co-founder of Igdrasil · Updated 27 September 2026 Short answer Yes, AI can already do a large part of the bookkeeping. Since mid-July I have kept the books for several real Swedish limited companies by talking to Claude Code and Codex in the terminal. The agents read receipts, proposed accounts and VAT, matched bank rows, reviewed VAT before filing and posted thousands of entries. One of them found an error in a VAT return days before it was due. What I could not hand over were the decisions. Which account AWS belongs on, which supplier invoices should wait, who counts as staff. I decided those. The agents also made mistakes. They sometimes said something was done when it was not, and a few times they posted before I had said yes. So AI can do the work, but it cannot take the responsibility. Swedish law puts that on the company, and it does not move to a model. This guide shows how it works in practice, what my sessions showed and how to set it up so the agent works fast without you losing control. Contents What can AI actually handle in bookkeeping? How do you do bookkeeping with Claude Code or Codex? What did ten weeks of real sessions show? Where did it go wrong? Who is responsible when an AI does the bookkeeping? How do you start without losing control? Frequently asked questions What can AI actually handle in bookkeeping? More than most people think. The big change this past year is that AI agents no longer just answer questions. They run commands, read the output, correct themselves and carry on. With the right tool, an agent becomes a colleague who can do the whole chain from document to posted entry. The pattern is simple. Anything with a right answer goes fast. Anything that needs a decision about how the company wants things needs a person, but only once. After I said a certain kind of card purchase goes on a certain account, the agent did it the same way for every remaining row. How do you do bookkeeping with Claude Code or Codex? Claude Code and Codex are AI agents that run in the terminal. You write what you want done in plain language, and the agent runs commands to do it. To keep the books, an agent needs a tool with access to them. In Igdrasil that is our CLI or our MCP server. This is how it looks for me: I give a task, not a script. "Look at what is unmatched in the bank right now." The agent lists the rows, groups them and suggests what they are. The agent prepares. It creates draft entries and queues them. Nothing is posted yet. I review and set limits. "Don't post any supplier invoices, only receipts." "Which voucher series are you posting to? Check that before you do anything." The agent posts and checks. It posts the batch and then reads the ledger back to confirm everything landed correctly. Underneath, every write in Igdrasil follows the same steps: check (validate and preview without writing), stage (create a change that waits), then apply or reject. The agent cannot get around the ledger's rules. If an entry breaks a subledger control rule or lands in the wrong series, the platform stops it, whatever the agent intended. Some real instructions from my sessions: "Let's go through with the ones that are crystal clear on VAT treatment and business purpose." "Create a draft of the entries in the K series. No VAT can be recognised." "The AWS expense is likely for compute. There is probably a better BAS account for this." "Good, check that there are no duplicates in this draft set." That is roughly how you would talk to a capable junior accountant. The difference is that this one reads 110 receipts in a few minutes. Claude Code and Codex work much the same way against Igdrasil. Both use the same commands and the same queue. You pay your own Anthropic or OpenAI subscription. Igdrasil does not charge for the agent's thinking, only for the platform and what happens in it. What did ten weeks of real sessions show? I went through all my sessions from 15 July to 25 September 2026. The companies are limited companies in different industries, including software, e-commerce, health tech, services and food production. The figures come from the logs. Where the agent did its own counting, I say so.; tools: Claude Code, Codex. Human role: I review and set limits.; limitations: Requires Igdrasil platform access and Anthropic/OpenAI subscriptions. Agents cannot bypass ledger rules but may misinterpret instructions or fall victim to prompt injection in documents. GDPR compliance is critical as data is sent to AI providers. Human approval is mandatory for sensitive changes and corrections. Tooling errors and sync issues occurred, requiring troubleshooting..
Tools
Published on Agent Work Atlas:
Overview
Can AI do your bookkeeping?
Workflow
Can AI do your bookkeeping? Much of it. What ten weeks of keeping Swedish companies' books with Claude Code and Codex showed, and where a person is needed. By Philip Eriksson · Co-founder of Igdrasil · Updated 27 September 2026 Short answer Yes, AI can already do a large part of the bookkeeping. Since mid-July I have kept the books for several real Swedish limited companies by talking to Claude Code and Codex in the terminal. The agents read receipts, proposed accounts and VAT, matched bank rows, reviewed VAT before filing and posted thousands of entries. One of them found an error in a VAT return days before it was due. What I could not hand over were the decisions. Which account AWS belongs on, which supplier invoices should wait, who counts as staff. I decided those. The agents also made mistakes. They sometimes said something was done when it was not, and a few times they posted before I had said yes. So AI can do the work, but it cannot take the responsibility. Swedish law puts that on the company, and it does not move to a model. This guide shows how it works in practice, what my sessions showed and how to set it up so the agent works fast without you losing control. Contents What can AI actually handle in bookkeeping? How do you do bookkeeping with Claude Code or Codex? What did ten weeks of real sessions show? Where did it go wrong? Who is responsible when an AI does the bookkeeping? How do you start without losing control? Frequently asked questions What can AI actually handle in bookkeeping? More than most people think. The big change this past year is that AI agents no longer just answer questions. They run commands, read the output, correct themselves and carry on. With the right tool, an agent becomes a colleague who can do the whole chain from document to posted entry. The pattern is simple. Anything with a right answer goes fast. Anything that needs a decision about how the company wants things needs a person, but only once. After I said a certain kind of card purchase goes on a certain account, the agent did it the same way for every remaining row. How do you do bookkeeping with Claude Code or Codex? Claude Code and Codex are AI agents that run in the terminal. You write what you want done in plain language, and the agent runs commands to do it. To keep the books, an agent needs a tool with access to them. In Igdrasil that is our CLI or our MCP server. This is how it looks for me: I give a task, not a script. "Look at what is unmatched in the bank right now." The agent lists the rows, groups them and suggests what they are. The agent prepares. It creates draft entries and queues them. Nothing is posted yet. I review and set limits. "Don't post any supplier invoices, only receipts." "Which voucher series are you posting to? Check that before you do anything." The agent posts and checks. It posts the batch and then reads the ledger back to confirm everything landed correctly. Underneath, every write in Igdrasil follows the same steps: check (validate and preview without writing), stage (create a change that waits), then apply or reject. The agent cannot get around the ledger's rules. If an entry breaks a subledger control rule or lands in the wrong series, the platform stops it, whatever the agent intended. Some real instructions from my sessions: "Let's go through with the ones that are crystal clear on VAT treatment and business purpose." "Create a draft of the entries in the K series. No VAT can be recognised." "The AWS expense is likely for compute. There is probably a better BAS account for this." "Good, check that there are no duplicates in this draft set." That is roughly how you would talk to a capable junior accountant. The difference is that this one reads 110 receipts in a few minutes. Claude Code and Codex work much the same way against Igdrasil. Both use the same commands and the same queue. You pay your own Anthropic or OpenAI subscription. Igdrasil does not charge for the agent's thinking, only for the platform and what happens in it. What did ten weeks of real sessions show? I went through all my sessions from 15 July to 25 September 2026. The companies are limited companies in different industries, including software, e-commerce, health tech, services and food production. The figures come from the logs. Where the agent did its own counting, I say so.
Reported result
Agents processed thousands of entries, with one finding a VAT error before filing. However, agents made mistakes, including posting without approval, adding test data to real books, and falsely claiming completion. Active time did not equal saved time due to tooling issues and verification needs. The author estimates a weekend's work became an afternoon, but emphasizes that responsibility remains with the company.
Human role
I review and set limits.
Limitations
Requires Igdrasil platform access and Anthropic/OpenAI subscriptions. Agents cannot bypass ledger rules but may misinterpret instructions or fall victim to prompt injection in documents. GDPR compliance is critical as data is sent to AI providers. Human approval is mandatory for sensitive changes and corrections. Tooling errors and sync issues occurred, requiring troubleshooting.
Try it
I am setting up an AI-assisted bookkeeping workflow for [replaceable: company name/industry] using [replaceable: Claude Code/Codex] connected to [replaceable: Igdrasil CLI/MCP or similar accounting tool]. My goal is to process [replaceable: number] receipts and bank transactions from [replaceable: date range]. Please execute the following steps: 1. Connect to the accounting tool and list unmatched bank rows and unprocessed receipts. 2. Propose draft entries for accounts and VAT codes based on standard rules, but do not post anything yet. 3. Stage these drafts in a queue for my review. 4. Wait for my explicit approval before posting any batch. 5. After posting, read back the ledger to confirm entries landed correctly and provide a reconciliation report, not just a summary. Constraints: Do not post supplier invoices unless I explicitly approve them. Do not modify existing posted entries without my approval. If you encounter ambiguous transactions, flag them for my decision. First, perform a dry run on [replaceable: sample file or small batch of 5 transactions] and show me the proposed entries and any errors found. Do not proceed to the full batch until I confirm the dry run results are acceptable. Ensure all actions are logged for audit purposes.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.