Skip to content
In the news TRM Labs × Finray — audit-ready crypto transaction monitoring for banking
Finray

Agentic accounting under four eyes: AI prepares, a person approves, the journal proves it

AI can draft journal entries, match documents and propose classifications. The accounting records only stay trustworthy if a second person approves what posts, and the journal can show who did what. A control pattern and a worked example.

By
Finray
Published
Updated
Reading time
9 min

Short answer: how should AI be used in accounting?

AI should prepare and a person should approve: a model can read a document, suggest an account, draft a journal entry or flag an anomaly, but it should carry no authority to change the books. A second, accountable person approves what posts, and the journal records who prepared and who approved each entry.

This note sets out that pattern, explains why each part matters, walks through an illustrative month-end example in Swiss francs, and states plainly what Corebanq Accounting does today and what it does not do.

Why is accounting a good place for AI, and a risky one?

Accounting is full of structured, repetitive judgement: reading a supplier bill, choosing an expense account, matching a bank line to an invoice, proposing a foreign-currency revaluation. These tasks suit language and document models, which can draft quickly and surface items a person might miss.

The risk is that the output of the work is the record itself. The Swiss federal SME portal describes proper accounting as resting on an inventory, a balance sheet and an income statement prepared with supporting documents, and states that accounting documents must be retained for at least ten years. A wrong entry that nobody can trace is worse than a slow entry that anyone can trace. So the question is not whether AI can draft, but who is accountable for what ends up in the books and how that can be demonstrated.

What is the “AI prepares, a person approves” pattern?

The pattern has four parts.

  1. Preparation. AI produces a proposal: a draft entry, a suggested mapping, a flagged item. The proposal is stored as a proposal, clearly distinct from a posted entry.
  2. Review. A person who is not the preparer sees the proposal together with its source document and the reasoning or fields the AI used.
  3. Approval. The reviewer approves, edits or rejects. Only approval turns the proposal into a posting.
  4. Proof. The journal records the posting, the identity of the approver, the preparer, and the link to the source document.

Two properties make this a control rather than a habit. The approval must be technically required, not merely recommended. And the approver must differ from the preparer, whether the preparer is a person or a model operated by a person.

How does this relate to four-eyes and human oversight rules?

Four-eyes, or maker-checker, is an old control: no single person both creates and releases a change. It has long been applied to payments, payroll and journal entries. AI does not change the logic; it changes who the “maker” might be.

Neither FINMA’s guidance nor the EU AI Act sets rules specifically for AI-prepared bookkeeping, but their general expectations point the same way. FINMA’s Guidance 08/2024, which is addressed to supervised financial institutions rather than to companies generally, says they should identify, assess, manage and monitor the risks of AI, and lists operational and model risks such as robustness, correctness, explainability and bias among them. The EU AI Act, for systems that are classified as high-risk, requires that they can be effectively overseen by natural persons while in use, that overseers can remain aware of the tendency to over-rely on outputs, that they can decide not to use the system or override its output, and that the system supports automatic recording of events. Whether an accounting assistant is in scope depends on its use and classification, so treat these as design guidance and confirm applicability with counsel. Application dates are staged, so check the current consolidated text before relying on a date.

The over-reliance point deserves emphasis. If approvers click through hundreds of AI proposals, four eyes becomes two eyes and a habit. Design the review screen so that the source document and the decisive fields are visible, and sample approvals afterwards.

What must sit outside the model?

RiskControl outside the modelWho owns it
AI proposes a wrong account or amountReviewer compares proposal to the source document before approvalAccounting lead
AI-prepared entry posts without reviewPosting requires approval by a second person as a system rule, not a procedureFinance management; software configuration under change control
Preparer approves their own workMaker and approver must be different identities, enforced in the data layerFinance management; engineering
Policy quietly weakenedChanging the approval policy is itself an approved changeFinance management
Document content manipulates the modelModel output treated as untrusted input; no posting authority; reviewer sees the documentSecurity; accounting lead
Reviewer rubber-stampsSampling of approved proposals; metrics on edit and rejection ratesInternal audit or independent reviewer
Entries altered after approvalAppend-only journal; periodic verification of the journal’s integrityFinance management; engineering
Cannot show why an entry existsDrill-down from statement figure to account, line, entry and source documentAccounting lead; auditor
Sensitive documents leave the companyModel hosted where the company chooses; cloud processing only by explicit opt-inSecurity; data protection officer

Several of these rely on the same idea: the model is one input, not an authority. The MCP specification expresses a similar principle for tool use in general, stating that hosts must obtain explicit user consent before invoking a tool and that tool descriptions should be treated as untrusted unless obtained from a trusted server. If an AI system can reach accounting functions through any integration protocol, the consent and authorisation checks must sit in the accounting system and not in the model’s goodwill.

Why does the journal have to prove it?

Approval is an event; a journal is the lasting record of it. If the journal can be edited after the fact, an approval trail is only as strong as the people with database access. Two design choices help:

  • Append-only, hash-chained records. Each journal record includes a cryptographic hash that depends on the previous record. Changing a record without recomputing every later hash breaks the chain, and a verification check finds the break. Someone who can rewrite the whole chain could hide a change, so hash-chaining makes ordinary tampering detectable rather than impossible, and it supplements access control rather than replacing it.
  • Drill-to-source. A reader should be able to start at a figure in a statement and walk down to the accounts, the lines, the entry and the source document, and see who prepared and who approved at the entry.

The AP2 payments protocol applies a related idea to payments: a chain of signed records from intent to cart to payment forms an audit trail meant to answer who authorised what. The accounting analogue is the chain from source document to proposal to approval to posting.

Worked example: a month-end close with AI assistance (illustrative)

The numbers are synthetic. The example illustrates the control pattern; it is not a description of any one product’s features. A small Swiss company closes January. The assistant, hosted on a model the company runs itself, prepares proposals for the controller to review.

  • Supplier bill. A cloud-hosting invoice for CHF 2,160.00. The assistant proposes an expense account and a VAT mapping from the invoice text and attaches the document. The controller notices that the invoice covers February as well as January and edits the proposal to defer CHF 1,080.00. The finance manager, who did not prepare or edit the proposal, approves it.
  • Foreign-currency revaluation. A EUR 12,000.00 receivable was booked at an exchange rate of 0.9400 and the month-end rate is 0.9350. The proposal books an unrealised loss of CHF 60.00 (12,000 times 0.0050). The controller confirms the proposal after checking the rate source, and the finance manager approves it.
  • Suspicious item. The assistant flags a bill for CHF 4,990.00 from a supplier whose bank details differ from the last three bills. It does not post anything; it adds a note for the reviewer, who calls the supplier.

At close, the journal contains, for each posted entry, the approver and the source document. The auditor selects the hosting entry, drills from the income statement to the entry, sees the edit the controller made to the AI’s proposal and who approved it. The AI proposal that was edited is part of the story, not hidden from it. The flagged bill did not post until the supplier confirmed the change.

How should a team start?

  • Keep AI off by default and turn it on per company, with an owner who decides what it may read.
  • Start with preparation tasks that have an easy check, such as matching and classification, before more consequential ones.
  • Measure the review. Track how often proposals are edited or rejected; a rate of zero is a warning sign about review quality, not a sign of accuracy.
  • Put approval rules under change control. If an approval requirement can be switched off quietly, it is not a control.
  • Decide where the model runs before the first document is read, and record the decision.

What does Corebanq Accounting do today, and what does it not do?

Today. Corebanq Accounting is accounting software for Swiss companies. Postings and bills post only after a second person approves, by default; this is a company policy, and changing it is itself an approved change. A person cannot approve their own change, and the database enforces this. The journal, audit log and access log are append-only and hash-chained, with a verify-ledger check. A reader can drill from a statement figure to accounts, lines, entry and source document. Each company’s data is isolated by row-level security. Month-end close runs use a checklist and a period lock, and foreign-currency revaluation is proposed as draft entries at month close. The AI assistant is off by default, runs on a model the customer hosts locally, and has no posting authority. AI proposals need a person’s confirmation. Cloud document reading is off unless the company turns it on. Corebanq Accounting offers chat access from Claude or ChatGPT. Questions are read-only. Documents arrive as drafts; a person approves before anything posts. Chat access is off by default and activated per company.

Not claimed. Corebanq Accounting does not use autonomous agents, and Finray does not claim agent integrations through the Model Context Protocol. The software does not claim that every change always needs four eyes; the default is a company policy. The approach described here is a control design pattern, not a statement that any AI proposal is correct.

Platform pages: Corebanq Accounting and Ordinis for institution-level control evidence. Related glossary entries: human-in-the-loop approval, agent mandate, agentic payments and the Model Context Protocol.

Frequently asked questions

Can AI post journal entries on its own?
Software can be built to let AI post journal entries on its own, but it is a poor control design. Posting changes the books, and the books are the record an auditor and a tax authority rely on. A safer pattern lets AI prepare a proposed entry and leaves the decision to post with an accountable person who is not the preparer.
What does four eyes mean in accounting software?
In accounting software, four eyes means a change takes effect only after a second person approves it, and the person who made the change cannot be the approver. It is strongest when the software enforces the rule rather than leaving it to procedure.
Why does the journal matter if a person approves everything?
Approval is a moment; the journal is the evidence. If the journal is append-only and records who proposed, who approved and from which source document, a reviewer can later verify that approval actually happened and see what the AI contributed.
Does Corebanq Accounting use AI agents?
No agents. Corebanq Accounting has an AI assistant that is off by default, runs on a model the customer hosts locally, and has no posting authority. AI proposals need a person's confirmation. Cloud document reading is off unless the company turns it on. Corebanq Accounting offers chat access from Claude or ChatGPT. Questions are read-only. Documents arrive as drafts; a person approves before anything posts. Chat access is off by default and activated per company.
How long must accounting records be kept in Switzerland?
The Swiss Federal portal for SMEs states that accounting documents and related reports must be retained for at least ten years under Article 958f of the Code of Obligations. Electronic retention is subject to separate rules, so confirm the current requirements with your accountant.

Sources

Discuss how this applies to your institution with the Finray team.

Book a briefing

All notes

Certificate of Registration NQA · UKAS Management Systems
ISO/IEC 27001:2022 Certificate of Registration issued by NQA to Finray Technologies Ltd, certificate number 215646, valid 21 October 2025 to 21 October 2028
Search
Type to search across Finray, products, company, and journal.

    Press Esc to close · ↵ to open the highlighted result.

    Book a briefing Step 1 of 3

    Your details

    Who is requesting the briefing.