It depends on the journal. Standard journals with a defined source, such as payroll postings or bank charges from the feed, can be posted by an AI agent at depth 3. Manual, unusual and top-side journals are drafted at depth 2 and approved by a named person who did not prepare them. The agent is never both preparer and approver.
Why are journals treated so carefully?
A journal entry can move any amount between any accounts. Most of the ledger is fed by sub-systems: sales invoices, supplier invoices, payroll, fixed assets. Journals are how everything else gets in, which makes them the main route for errors and for deliberate misstatement. Auditors test them for that reason, looking especially at entries made late in the period, by senior people, to unusual accounts, or with round amounts. The basics of what an entry contains are in what a journal entry is.
So the question for an agent is not whether it can create a balanced entry. It can. It is which entries it may put into the ledger without someone else looking first.
Which journals can the agent post, and which does it draft?
Post at depth 3, when all conditions hold: - recurring journals from an approved template, such as the payroll summary posting or a fixed monthly management charge; - entries whose amounts come straight from a sealed source, such as bank charges and interest from the bank feed; - releases from an approved schedule, such as prepayments and deferred income; - evidence-backed accruals and prepayments with auto-reversal.
Draft at depth 2: - any manual or one-off journal; - reclassifications and corrections of past entries; - estimates and provisions; - top-side or consolidation adjustments; - entries to sensitive accounts: cash, suspense, intercompany, equity, tax; - anything posted after the period close timetable's cut-off.
The conditions for depth 3 are the same every time: the period is open, the template or source is approved, the amount is within the customer's limit for agent-posted entries, and the supporting evidence is attached.
Where does the named approver sign off?
On every drafted journal, before it reaches the ledger. The approver must be a different person from whoever asked for the journal, and must have the authority for its value. Most ledgers can enforce this, with preparer and approver as separate fields and the system refusing a match.
For the agent, this means its identity is recorded as preparer and it cannot be recorded as approver. If two agents are involved, one drafting and one checking, the check is useful but it is not an approval. Approval of a manual journal is a person's act. The process is set out in how to draft journal entries for a person to post.
An illustrative example
Take a professional services firm at month end. The agent posts 26 standard journals: payroll, bank charges, rent from the lease schedule, prepayment releases and depreciation from the asset register. Each has a template or source, and each is under the agent's posting limit.
The finance manager then asks it to reclassify 45,000 of software subscriptions coded to office costs. That is a correction, so the agent drafts it, attaching the invoices and the coding rule. The financial controller approves and the agent posts it, recording both names. A partner separately emails asking for "a 200,000 accrual for the client project we expect to win". The agent drafts nothing: an accrual for revenue not yet earned has no support, and the request goes to the controller with that explanation.
What controls make it safe?
| Control | Evidence it leaves |
|---|---|
| Separate preparer and approver fields, enforced by the ledger | Both identities on every journal |
| Approved templates for depth 3 posting | Template version and approver on file |
| Value limit for agent-posted journals | Limit set by the customer, checked per entry |
| Supporting evidence attached | Source documents linked to the entry, using the journal entry support template |
| Period lock | Refusals recorded for attempts after close |
| Every verdict filed | Allowed, held and refused journals all on record |
Mistakes still happen, and the answer is a reversing entry, not an edit. The rules for that are in whether an AI agent can reverse a journal entry. The wider controls are in AI agents for finance: books you can audit.
Questions people ask
Do auditors accept journals posted by an agent?
They accept them as automated entries if they can test the controls around them: who set up the template, what limits apply and whether the record shows each posting. Manual journals still need the human approval they would always need.
Should the agent post journals across all entities?
Only where the conditions are met in each. Entity-specific limits and templates keep a mistake in one ledger from spreading.
Can the agent spot unusual journals people post?
Yes. Reviewing all manual journals for late timing, round amounts, unusual accounts and preparer patterns is a good depth 1 job, and the output goes to the controller.