AI for account reconciliation: start with one account

Account reconciliation means comparing a ledger balance with the records that support it and explaining every difference. AI can do most of the preparation: pull the records, apply your matching rules and list what didn't match. People still decide what each difference means, and an independent reviewer approves any adjustment.

Across the organizations we interviewed, reconciliation and data movement were the most common workflows we coded and the main source of manual effort. One accounting team told us it uses AI for two parts of its month-end reconciliation: matching vendor records between systems, and building the schedule that supports the balance. A reviewer checks the uncertain matches and signs off before anything posts.

Start the same way: one account, tested on past months, then run next to your normal close. Expand only when it passes the checks below.

Biggest takeaways

• Start with one recurring account, and leave estimates for later. Accounts payable, a prepaid schedule or a straightforward accrual makes a good first test. Accounts that rest on estimates, like reserves, can wait.
• Give AI the preparation, not the sign-off. It can retrieve the records, apply your matching rules and list what didn't match. A reviewer who didn't prepare the work approves every adjustment.
• Check the population, not just the total. A schedule can tie out while records are missing from it. Count the source records before you trust the balance.
• Make every match show its method. “Matched on invoice ID” and “similar vendor name” need different levels of review.
• Test on past months, then run it beside your close. Compare what each process caught, and stop if the checks fail.

Plan the pilot

1. Measure today. Record how long preparation and review take now, and what the review usually catches.
2. Rerun two or three past months. Use months your team already reviewed, and compare the AI-prepared reconciliation with what the reviewer found then.
3. Run next to the live close. Prepare the account both ways for at least one close, and compare what each caught.
4. Expand or stop, using the checks in “Decide whether to expand” below.

Each run follows the five steps in the next section.

How each run works, step by step

The example reconciles a made-up accounts payable balance in the general ledger to the open bills in the AP system. Names, counts and amounts are illustrative.

Step 1: Pick the account and the reviewer

Choose an account you reconcile every month, with support you can reach easily and an owner who knows it well. Name a reviewer who isn't the person who prepares it or sets the matching rules.

Write down six things before you start: the entity, the period, its cutoff, where the source records come from, the matching rules, and what “done” means.

Step 2: Pull the source records and count them

The AI retrieves the ledger detail and the supporting records for the period, and records how it got them: the report, the filters and the time it ran.

Then count. If the AP report lists 214 open bills and the schedule has 211, explain the gap before you look at totals: missing bills, duplicates or split lines. A filtered export can quietly drop cancelled invoices, a business unit or records outside a date range.

Step 3: Match, and label how each match was made

Exact matches use rules you approved, run as ordinary formulas or code. The AI proposes candidates only where records are ambiguous. Every row says which method it used. The team we interviewed reported working this way: AI matches vendor identifiers across systems, and a person reviews and signs off on the uncertain ones.

Step 4: Work the exceptions, then spot-check the matches

The preparer works the exception list: accepts or rejects proposed matches, investigates unmatched items and proposes the accounting treatment. Agree in advance which differences must be investigated, which can be cleared under a documented threshold, and who can approve a write-off.

The reviewer then checks a sample of the accepted matches. If you only review what the system flagged, you depend on it to find its own mistakes.

In the example, the reviewer accepts the Northwind match after checking the vendor ID. The Fabrikam bill is dated after period end, so it belongs to the next period and stays out of this reconciliation, with the reason recorded.

Step 5: Approve, post and keep the record

The preparer drafts any adjustment. The reviewer named in Step 1 approves it before it posts, then checks the posted entry against the schedule and signs off the reconciliation. In the example, the schedule now ties to the ledger, and the sign-off records both decisions. Keep the source references, each row's match method and the reviewer's decisions, so the next close and your auditor can retrace the work.

What goes wrong, and what the reviewer does

Records go missing between the source and the entry. A practitioner at a second organization told us about a journal entry prepared with AI from a source report. At review, cash didn't tie out, and records in the source weren't in the entry. Nobody could tell whether a person or the AI dropped them. Step 2 catches this: count the source records before trusting the total.

A record changes after sign-off. In the example, the vendor sends a corrected invoice after the schedule is approved. The review should show the original and the corrected invoice side by side. Usually the difference is recorded in the open period. Reopening a closed period depends on materiality and whether statements were issued, under your close policy. Rerunning the job must keep that decision and must not post the same adjustment twice. The first team described source records that changed after close without the change reaching data already exported for accounting.

A confident match is wrong. Test with lookalike vendor names, subsidiaries, duplicates and missing IDs, using cases your team has already reviewed. Track wrong matches that were accepted, not just items left open. Treat a model's confidence score as a cue for review until you've tested how it relates to being right.

Decide whether to expand

How do you know it worked?

Agree these with the controller before the first run. There is no universal accuracy threshold, so set yours in advance.

• Every source record is matched, listed as an exception, or excluded with a reason.
• A sample of accepted matches has no errors you wouldn't accept from a preparer.
• Differences above your threshold are explained.
• Approvals and record versions are kept, and a rerun on the same inputs gives the same matches, or the differences are explained.
• Preparation plus review takes less time than the baseline, measured the same way.
• The manual process still works if you need it.

When should you stop?

• You can't check the records against their source.
• The approved matching rules aren't available or keep changing.
• Changed inputs make the reviewed result unreliable.
• Review takes as long as preparing the account by hand.

When you stop, record which condition failed and go back to the manual process.

If the reconciliation is a key control, such as a SOX key control, changing how it's prepared changes the control. Document the new design, keep access and change controls over the tool, and talk to internal audit and your external auditors before relying on it.

Improve, build or buy

• Improve your current tools if you know the process and the burden is repetitive preparation.
• Build a narrow workflow if your sources or rules are specific and someone can maintain it.
• Buy close software if several accounts or entities need shared tasks, approvals and evidence.

Whichever you choose, test it on the same account with edited and deleted records, ambiguous matches and a failed job. Airframe's Financial Close Management research describes close software as covering checklists, reconciliation and matching, journal entries, flux analysis and sign-off trails. In every option, finance still owns policy, exception review and sign-off.

Questions about account reconciliation

What is account reconciliation?

Comparing a general ledger balance with the records that support it, such as subledgers, bank statements and payment processor data, then investigating the differences, documenting the support and signing off.

How do you reconcile an account?

1. Pull every record for the period and count them against the source.
2. Match the records to the ledger using approved rules.
3. Investigate and explain each unmatched or ambiguous item.
4. Have a reviewer approve the reconciliation and any adjustment.
5. Check the posted adjustment and keep the support.

What is month-end close automation?

Software that runs repeatable close steps, such as task tracking, matching and journal-entry preparation, while people review the exceptions and sign off.