Draft and review journal entries
An AI agent drafts recurring and adjusting journal entries from your source data and prior periods, attaches the support, and flags anomalies for a human to approve, cutting the manual grind out of month-end close.
Tools you'll use
Drafting and reviewing journal entries is the work of recording financial transactions into the general ledger — accruals, prepaids, depreciation, payroll, intercompany, and reclassifications — and then having a second person check the accounts, amounts, dates, and supporting documentation before the entry is posted. It is the repetitive core of every month-end close, and it is where small mistakes (a transposed digit, a wrong cost center, a missing accrual) quietly distort the financials.
It is also slow. According to a 2025 benchmark of finance professionals reported by CFO.com (citing Ledge), 50% of finance teams still take six or more business days to close the books, and 27% regularly take more than seven. A big chunk of that time is people preparing routine entries by hand and reviewers racing through a stack of them. The downstream cost is trust in the numbers: in a BlackLine survey of more than 1,300 C-level executives and finance professionals, 47% said they were concerned about making decisions based on outdated or inaccurate information.
This is a strong fit for an agent because most journal entries are not creative work. They follow the same logic every month: pull this report, calculate this amount, code it to these accounts, write this description, attach this support. An agent can do the drafting and the first-pass checking from your own files and templates, leaving your team to do the judgment and the sign-off.
The result is something your finance team owns: a documented, repeatable routine that drafts the predictable entries, surfaces the exceptions, and never posts anything without a human approving it.
Moriva's take
This clears Gate 1 cleanly: drafting and reviewing entries is real, recurring weekly-and-monthly work, not a demo. It clears Gate 2 because Claude Code or Codex builds a routine your accountants can read, run, fix, and extend in plain files — no black box. The reason it lands CAREFUL rather than GO is Gate 3 risk, not value: this touches the general ledger and audited financials, so the agent drafts and proposes but a qualified person must review and post every entry. Stand it up on the five or six highest-volume recurring entries first, measure the hours saved against your current close calendar, then widen.
How do you draft and review journal entries?
- 1
Pick the 5-10 entries that eat the most time
Don't start with the whole close. List your recurring and adjusting entries and rank them by how often they recur and how long they take to prepare by hand — prepaids amortization, accruals, depreciation, payroll allocations, intercompany, recurring reclasses. Choose the handful that are high-volume and rules-based. These are where an agent pays off fastest and carries the least judgment risk.
- 2
Write down the rules for each entry, in plain English
For each chosen entry, document the source (which report or export), the calculation or formula, the accounts and cost centers it hits, the description template, the support that must be attached, and any threshold that requires extra review. This is good practice even without AI — and it becomes the spec the agent follows. If you can't write the rule down, the entry isn't ready to automate yet.
- 3
Point Claude Code or Codex at sample data and have it build the drafting routine
Give the tool read access to anonymized or sandboxed copies of your source files (trial balance exports, sub-ledger reports, prior-period entries) and the rules from step 2. Describe the task in plain English: 'read this prepaid schedule, calculate this month's amortization, produce a draft journal entry in this format with the support referenced.' It builds a script or small agent that produces draft entries — a CSV or your ERP's import template — that your team can read and edit.
- 4
Add the review layer: anomaly checks and a reconciliation back to source
Have the tool extend the routine so each draft is checked before a human sees it: does it tie back to the source report, is the amount within the expected range versus prior periods, are the accounts valid, does it balance. Flag anything outside tolerance for closer review. This is the agent doing the tedious first-pass review so your senior reviewer spends time on the exceptions, not the clean entries.
- 5
Keep a human in the loop for approval and posting
The agent drafts and flags; it does not post. Build the workflow so a qualified preparer or reviewer sees every draft, its support, and any flags, then approves and posts through your normal controls. Preserve segregation of duties: the agent (and whoever runs it) is not the approver. Nothing reaches the ledger without a named person signing off.
- 6
Run it alongside the manual process for one close
For the first month, run the agent's drafts in parallel with how you do it today and compare. Where they match, you've validated the rule. Where they differ, you've either found a bug in the routine or an error in the old manual entry — both are wins. Tune the rules until the drafts are reliably correct, then retire the manual version for those entries.
- 7
Document it and hand it to the team to own
Store the routine, the rules, and a short run-book in your repo or shared drive so any accountant can run it, understand it, and change it when a rate, account, or schedule changes. Because Claude Code and Codex produce readable scripts and plain-English logic — not a sealed product — your team can extend it to new entries themselves and keep it working as the business changes.
What could go wrong (and how to handle it)
A wrong entry gets posted to the live general ledger and distorts the financials.
The agent never posts. It produces drafts only; a qualified person reviews the entry and its support and posts through your normal controls. Start with low-judgment, rules-based entries before anything requiring estimation.
The agent quietly produces a plausible-looking but incorrect amount (a hallucinated or miscalculated figure).
Build in deterministic checks: the draft must tie back to the source report, balance, and fall within a tolerance band versus prior periods. Anything outside tolerance is flagged for human review rather than passed through.
Sensitive financial data is exposed during development or operation.
Develop against anonymized or sandboxed copies, not production. Run the routine in your own environment with least-privilege, read-only access to source files and no posting rights. Treat ledger data as high-sensitivity throughout.
Segregation of duties breaks down — the same actor prepares and approves.
Keep the approver a different, named human from whoever runs the agent. Document the approval matrix and route every draft above a materiality threshold to a senior reviewer or controller, exactly as you would a manually prepared entry.
Auditors challenge AI-prepared entries that lack a clear, evidenced trail.
Make the routine attach or reference support for every draft and log who reviewed and posted it. The plain-English rules and readable script become your documentation of how the entry is derived — often a cleaner audit trail than an undocumented spreadsheet.
The routine silently breaks when a rate, account, or report format changes and keeps producing stale drafts.
Because the team owns readable code, they can fix it. Add a review/expiration date to each entry's rules, validate that source files match the expected format before drafting, and review the routine at a set cadence rather than assuming it still holds.
Prompts to get started
FAQ
Will the AI post entries directly to our ledger without anyone checking?
No, and it shouldn't. The setup we recommend has the agent draft entries and run first-pass checks only. A qualified person reviews each draft and its support and posts it through your normal controls. The agent removes the typing and the tedium, not the judgment or the sign-off.
How is this different from the journal entry automation already in our ERP or close tool?
ERP rules and recurring-entry templates are great for entries that never change. The difference here is that you own a flexible routine you can point at any source file and adapt in plain English, including entries that involve a calculation or pulling from a report your ERP doesn't natively read. It complements your existing tools rather than replacing them — and the logic is readable, not locked inside a vendor's product.
What happens when our accounts, rates, or report formats change?
Because Claude Code or Codex produces readable scripts and plain-English rules, your team can update them — that's the whole point of owning it. We recommend giving each entry a documented review date and having the routine check that source files match the expected format before it drafts, so a format change surfaces as a flag instead of a silently wrong entry.
Is our financial data safe?
You develop against anonymized or sandboxed data and run the routine in your own environment with read-only access to source files and no posting rights. The ledger data never has to leave your control. We treat this as high-sensitivity work throughout.
Will auditors accept AI-drafted entries?
Auditors care about evidence and controls, not who typed the entry. A routine that attaches support to every draft, reconciles back to source, and logs who reviewed and posted it typically produces a cleaner trail than an undocumented spreadsheet. The documented rules are your explanation of how each entry is derived.
Sources
- 50% of finance teams still take six or more business days to close the books, and 27% regularly take more than seven (2025). — CFO.com (citing Ledge month-end close benchmarks)
- 47% of more than 1,300 C-level executives and finance professionals surveyed were concerned about making decisions based on outdated or inaccurate information. — BlackLine
More from Finance & Accounting
Reconcile accounts faster at close
Use an agentic tool to match transactions, surface only the real exceptions, and draft reconciliation write-ups every close cycle, so your team shrinks a multi-day chore into a few reviewed hours.
Expense categorization and anomaly flags
Build an owned script that reads your card and expense exports, assigns a GL category to each line with a confidence score, and flags the odd ones (duplicates, out-of-policy amounts, unusual vendors) for a human to check.
Month-end variance commentary and board-pack drafts
Turn your closed actuals into a first-draft variance narrative and board-pack pages your team owns and edits, instead of writing every line from a blank page each month.
Want help shipping this?
We'll build it with your team on your real work — and leave you owning it, not renting it.