A finance agent should spend its first months drafting journals that a person reviews and posts. Starting at depth 2 builds a record of how often its drafts are accepted unchanged, edited or rejected, and why. That record is the evidence any later posting authority must rest on. Without it, posting rights are a guess.
What does drafting at depth 2 involve?
The agent does everything except the final step. For each journal, it identifies the need, gathers the supporting evidence, calculates the amounts, selects the accounts and cost centres, writes the narrative and attaches the support. The draft then waits for a person, who checks it and either posts it, edits and posts it, or rejects it.
The person's role is real review, not a rubber stamp. They are checking that the journal is needed, the amounts are right, the accounts are correct and the support justifies the entry. The approach to doing this well is set out in how to review manual journals before they are posted with an AI agent.
To prepare good drafts, the agent needs to see a lot: the ledger, the sub-ledgers, the documents behind each entry. That is why finance agents need read access to more than they can change. Wide reading supports narrow drafting.
What evidence does the drafting period produce?
Every reviewed draft is a labelled case. The reviewer's decision says whether the agent got it right. Over a few close cycles, those decisions add up to a picture of the agent's readiness, journal type by journal type.
| Signal | What it shows |
|---|---|
| Drafts posted unchanged | The agent handles this journal type correctly |
| Drafts edited before posting | Close, but something is off: amounts, accounts, narrative or support |
| Drafts rejected | The journal was not needed, or the approach was wrong |
| Reason for each edit or rejection | Where the agent's sources, rules or understanding fall short |
| Reviewer time per draft | Whether the drafts are saving effort or creating it |
| Same journal type across periods | Whether performance is stable or varies with circumstances |
The reasons matter most. A reviewer who simply edits and posts leaves no trace of why. Asking reviewers to pick a short reason for each change, from a fixed list, turns their routine work into the evidence base for certification.
Which journals might earn depth 3?
Routine, rule-based journals with clear sources and predictable amounts are the natural candidates. Examples include monthly depreciation from the fixed asset register, release of prepayments on a fixed schedule, recurring allocations based on a set key and standard accruals from approved purchase orders.
For these, the drafting period can show sustained accuracy, and certification on a labelled set of past journals can confirm it gate by gate. If the agent clears every gate, it can earn depth 3 for that journal type, posting within the firm's limits, such as a maximum amount and a list of permitted accounts. Whether an AI agent can post accruals and prepayments is answered this way: for some, on evidence, within bounds.
The next step, where an agent posts on its own, is a significant change in control design. The considerations are covered in how to let an AI agent post journals in your general ledger, and the control question in segregation of duties when an agent posts journals.
Which journals should stay drafted?
Some journals should stay at depth 2 however good the evidence looks, because they involve judgement or carry high risk:
- Judgemental provisions: bad debt, inventory obsolescence, warranty, legal claims.
- Manual adjustments outside the normal process.
- Top-side and consolidation adjustments at group level.
- Reversals and corrections of prior entries.
- Entries to sensitive accounts: equity, related parties, suspense, intercompany settlements.
- Anything above a materiality threshold set by the controller.
The firm can enforce this through its own limits, so these journals are held for a named person regardless of the depth the agent has earned elsewhere.
How long should the drafting period be?
Long enough to see each journal type enough times, across enough different circumstances, to trust the pattern. That is a matter of case count, not calendar. A journal posted daily builds evidence quickly. A quarterly accrual may need a year of drafts before the evidence means much.
Seasonal and unusual periods matter. A drafting period that covers only quiet months will not show how the agent handles year-end, a reorganisation or a new revenue stream. If possible, include at least one period-end with unusual activity before granting posting authority.
There is a trade-off. Drafting costs reviewer time, and finance teams under pressure will want to move to posting quickly. The cost of moving too early is an agent posting errors that nobody catches until reconciliation, or worse, audit. A few extra weeks of drafting is cheap by comparison.
What happens if posted journals go wrong later?
Then the agent drops back. A certification run that clears no depth returns the agent to depth 1 the same day, and posting stops. The drafting record from before, combined with the investigation into what went wrong, shows what needs fixing. After the fix, the agent drafts again and re-earns depth on fresh evidence.
What should you do differently?
If your finance agent can already post journals without a drafting period behind it, consider moving it back to depth 2 until you have the evidence. If you are about to deploy one, start at drafting, ask reviewers for reasons, and decide in advance which journal types might earn posting and what evidence they need. This is how AI agents for finance: books you can audit works in practice.
Questions people ask
Does drafting at depth 2 save any time?
Yes. Preparation is most of the effort in journal work. Reviewers spend minutes on a good draft rather than the longer time needed to build the entry themselves.
Who should review the agent's drafts?
The people who would normally prepare or review those journals. Their knowledge makes the review meaningful and their reasons make it useful as evidence.
Can the agent post corrections to its own drafts after review?
No. Changes after review should be made by the reviewer or sent back as a new draft, so the record shows clearly who decided the final entry.