The obvious way to build an AI accountant is to let it post. It is also the fastest way to lose the trust of an audit committee. When we designed execution into MFT Intelligence, we started from the opposite end: what would a careful reviewer need to see before letting a machine touch the ledger at all?

Three rules, enforced in code

  • Agents only draft. A correcting journal entry or a stock adjustment is a proposal with its evidence attached, never a posting.
  • Approval is by name and by tier. A registered person approves below your ISA 320 materiality threshold; at or above it, a second registered reviewer or manager must approve too. Unregistered names are refused.
  • Nothing is deleted. A wrong entry is corrected by a new reversing entry that references the original, approved by the same tiers.

Evidence or silence

Every finding an agent produces must cite the record ids it actually retrieved in that run. A finding that cites nothing, or quotes a number that is not in the cited records, is rejected by the evidence gate and shown to the reviewer struck through, never as a fact. The agent is allowed to say it found nothing. It is not allowed to guess.

The log is the product

Every run records which model processed the data and whether it ran in the cloud, the tokens used, the client and the task, and every approval, execution and reversal with the resulting entry id. The log is kept for seven years by default, in line with working-paper conventions. When a client’s auditor asks how a number got there, the answer is a line in a file, not a meeting.

Governance that lives in a document gets waived. Governance that lives in code gets a pull request.

Where this is today

The Auditor and Virtual CFO agents are available as a managed service with MFT accountants reviewing every output. The execution path is built and tested; live write-back to Odoo is configured per client once their connector is in place, and until then approved drafts are handed to the team as ready-to-post entries. The same two agents are being built into the seg-audit report pipeline so every report is AI-reviewed before a person reviews it. That integration switches on only after the AI-processing disclosure is published in the platform’s terms.