AI can draft parts of the close: bank matches, accrual proposals, unusual-balance flags and first-draft variance commentary. It can't make judgment calls, post entries without approval or sign off. The setup that works keeps the math in formulas, ties every number to the ledger and puts a person on every sign-off.
What Does "AI in the Close" Actually Mean?
Vendors sell three different jobs under one label. Split them and most decisions get easier.
- Calculation. Sums, variances, amortization schedules and tie-outs belong in formulas or code. The same input must give the same output every time.
- Drafting and matching. A language model reads messy text, such as bank memos and invoice lines. It proposes a match, an accrual or a sentence of commentary.
- Judgment and sign-off. Someone with authority decides if the proposal is right, edits it and approves it.
Ledge, a close-software vendor, lists "using AI reasoning for every close task" as a mistake. Rule-based steps belong in fixed logic, it says, or results get "harder to explain during review or audit" (Ledge).
Our rule: if a step has one right answer, it's a formula. AI earns its place where the input is messy text.
Which Close Tasks Can AI Handle Today?
This assumes a team of 3 to 25 on QuickBooks, Xero, NetSuite or Sage Intacct.
| Close task | What AI does well | What it can't do | Verdict |
|---|---|---|---|
| Bank and card reconciliation | Matches lines with messy memos, suggests rules for repeat items | Decide what an unexplained deposit is | Draft now, person clears exceptions |
| Accruals for unbilled costs | Spots vendors that usually bill but haven't, proposes amounts | Know whether a service was received | Draft now, person approves each entry |
| Prepaids and amortization | Reads contract terms into a schedule | Beat a formula once terms are set | Formula, AI only for intake |
| Intercompany and consolidation | Matches intercompany balances across entities and flags mismatches | Decide how to eliminate, translate FX or settle a difference | Match now; eliminations by formula, person approves |
| Flux and variance review | Ranks accounts by change, drafts first-pass explanations | Know a cause that isn't in the data | Draft now with strict prompts |
| Board or management commentary | Turns a variance table into prose | Judge what the board needs to hear | Draft now, person rewrites tone |
| Revenue recognition | Summarizes contract clauses | Make the ASC 606 call | Keep with a person |
| Judgmental reserves (bad debt, inventory obsolescence, warranty) | Pulls history and aging into a starting view | Set the assumption | Keep with a person |
| Posting journal entries | Prepares the entry and its support | Approve its own work | Never without approval |
Intercompany is the sleeper, and often the slowest step on a three-entity NetSuite setup. The model can spot that Entity A booked a management fee in March and Entity B in April. A person decides which side to fix.
The expensive part of a flux review isn't the nine lines the data explains. It's the three that need a call to a department head. AI shortens the first part and hands you the call list for the second.
How Should You Roll It Out?
Start with one workflow. A broad launch makes errors hard to trace.
- Pick one workflow. Variance commentary and unbilled-cost accruals both have clear inputs and a visible output.
- Run it on a period you already closed. You know the right answer, so you can compare line by line.
- Run it in parallel for one live close. Your team closes the normal way while the AI draft runs alongside.
- Write down the review step. Record who checks the draft, what they check and where they sign.
- Measure two things. Track reviewer time and how many lines needed a material edit.
- Expand after a clean cycle. Add the next workflow once the first needs little correction.
Preparers told PCAOB staff, who oversee US public-company audits, much the same. Some govern GenAI like other technology, for example by "running technology in a parallel environment prior to deployment." Others are "testing individual use cases" before wider use (PCAOB staff Spotlight, July 2024).
A Monthly Variance Review With AI Drafts
All figures below are sample numbers for illustration, not client data.
Sample Co. has US$18M revenue, a three-person finance team and QuickBooks Online. Its policy flags any P&L account that moved more than US$10,000 and more than 10% against the prior month.
Step 1: flag by formula. A spreadsheet compares 140 P&L accounts, and twelve cross the threshold. No AI is involved yet.
Step 2: draft explanations. The model gets the 12 flagged lines plus vendor detail, invoice memos and headcount. It drafts one explanation per line and must cite a source for each cause.
Step 3: flag what it can't explain. For 9 lines a cause shows up in the data, and for 3 lines nothing does. Those 3 go on the list as questions, not guesses.
One drafted line read: "Hosting up US$14,200 (22%). Two usage invoices from the hosting vendor, dated March 3 and March 17."
Step 4: a person reviews. The controller confirms 8 of the 9 drafts, edits 1, and takes the 3 questions to department heads.
The edited line was hosting, because two bills from a monthly vendor in one month is a cutoff flag, not a cause. The March 3 invoice was February usage that February never accrued. The controller rewrote the line to show part of the US$14,200 as a cutoff miss and added hosting to the monthly accrual list.
| Stage | Sample result | Who does it |
|---|---|---|
| Accounts compared | 140 | Formula |
| Lines over threshold | 12 | Formula |
| Explanations drafted with a cited cause | 9 | AI |
| Lines flagged as open questions | 3 | AI |
| Drafts accepted without edit | 8 | Controller |
| Drafts edited | 1 | Controller |
AI didn't do the variance review. The controller started from a sourced draft and a three-item call list. That one edit is why the reviewer exists.
What to Automate vs Keep With a Person
Automate (draft or calculate):
- Variance math, thresholds and tie-outs (formulas, not AI)
- Bank, card and intercompany match proposals
- Accrual proposals from POs, receipts and vendor history
- First-draft commentary that cites the data, plus the open-question list
Keep with a person:
- Whether a service was received or an estimate is reasonable
- Revenue recognition, reserves, impairment and other judgments
- Approving and posting every journal entry
- What the board, lender and investors are told, and the final sign-off
What Will Your Auditor, Lender or Buyer Want to See?
Four outside readers check your close, and each asks something different.
- Your annual audit or review. Private-company auditors follow AICPA standards (AU-C) in the US, CAS in Canada and ISAs in the UK. Under SAS 145, US auditors must understand your journal entry controls and general IT controls. They must also evaluate how those controls are designed (Journal of Accountancy, March 2025).
- Your lender. Covenant certificates pull from the same close, so a wrong accrual can move EBITDA across a covenant line.
- Your board and investors. A fluent wrong cause costs more trust than a blank.
- A buyer's diligence team. A quality-of-earnings review tests cutoff, accrual true-ups and one-off items. A clean trail of drafts, edits and true-ups shortens that conversation.
Your auditor will likely treat the sign-off on an AI draft as a review control. Expect questions on what the reviewer checked and at what threshold. Approving a fluent draft without checking it is the risk they'll probe.
COSO's Achieving Effective Internal Control Over Generative AI applies its framework to GenAI (COSO). It sorts GenAI uses into eight capability types, such as ingestion, posting and judgment. Each comes with minimum control expectations across all five COSO components (Journal of Accountancy, Feb 2026).
For each AI step, keep the input data, the prompt version, the raw draft and the reviewer's edits. If a reviewer can recompute every number from the ledger, and the file shows what they checked and changed, you're in good shape. If you're a US-listed company, these steps also sit inside your SOX controls (see our checklist).
Where to Go Next
- How to Use AI for Accruals Without Losing the Audit Trail
- How to Estimate Accruals for Invoices That Haven't Arrived
- How to Get AI Variance Commentary That Only Cites Your Numbers
FAQ
Can AI close the books on its own? No. It can draft matches, accruals and commentary, but a person still judges estimates, approves entries and signs off.
Should AI calculate our variances? No. Use formulas or code so the results repeat and tie to the ledger. Use AI to explain the numbers, not to produce them.
Will our auditor accept AI-drafted work? They'll test the review, not the tool. Keep the input, prompt version, draft and reviewer edits so they can trace and recompute each number.
Flux is your embedded finance team. We run one workflow free on a period you've already closed, so you can compare it line by line.
Contact us