Reviewing AI accounting output: sources and human control
An accounting result is easier to review when the person approving it can see the source document, the extracted values, the related documents, and the workflow status in one place. That is a practical product requirement, not a claim that the software replaces the company’s statutory records or compliance controls.
This distinction matters because terms such as “auditability,” “traceability,” and “explainability” are often used as if they meant the same thing. They do not.
Three different requirements
Source visibility
The reviewer can open the original invoice and any related purchase order or delivery note. The product keeps those documents connected to the structured data and proposed accounting result.
Review context
The reviewer can see extracted fields, confidence, matching differences, the proposed category, subaccount, and cost centre, plus the approval or exception status. This gives the person enough context to validate the work or send it back.
Regulated records and audit logs
Retention, formal audit trails, e-invoicing records, and tax evidence are governed by the relevant legal and accounting systems. A product may support some of those requirements, but a document review screen is not proof of a compliant audit log.
What to ask in a demo
Can I move from a value to its source?
Select a total, supplier reference, or line item and ask to see where it appears in the original document. Repeat the test with a custom field and a low-confidence value.
Can I see the complete source set?
For a purchase-based invoice, ask to open the invoice, purchase order, and delivery notes together. The differences should remain visible without recreating the comparison manually.
How does human review work?
Ask what sends a field or invoice to review, who receives it, what that person can change, and how the validated result continues. The workflow should remain understandable without relying on colour alone.
What records stay in the ERP or compliance system?
Ask the vendor to name the boundary. Which documents and accounting data are submitted to the ERP? Which audit, retention, filing, and access records remain there? A precise answer is more useful than a broad promise of “full auditability.”
Useful governance without inflated claims
Confidence, source visibility, human review, permissions, and clear system boundaries are useful controls for AI-assisted accounting. They help a team detect uncertainty and understand prepared work before it is posted.
They do not remove professional responsibility. They also do not make ordinary accounting software a high-risk AI system under the EU AI Act merely because it handles invoices or accounting data. Classification depends on the system’s intended purpose and the specific categories defined in the law.
How Calitem approaches review
Calitem presents the source documents, extracted information, accounting proposal, and workflow status for human review before validated data is sent to the ERP. Field confidence can direct uncertain values to a person, while matching and approval exceptions remain visible in the same accounts-payable flow.
Calitem is not positioned as a customer-facing compliance audit-log product or as a replacement for the record retention, access logging, and statutory controls maintained in the ERP and compliance stack.