---
title: Reconciliation with AI: bank, payments and inventory | Efidis
description: How to reconcile your bank, payment processor and inventory with AI: what it does well and badly, a five-step workflow, the key controls and common gaps.
url: https://efidis.co/en/blog/conciliacion-con-ia
lang: en
---

# Reconciliation with AI: bank, payments and inventory

*Applied AI · Efidis team · Updated November 4, 2026*

**In short:** AI is good at matching your bank statement, your payment processor report and your own records by amount, date and reference, and at suggesting why each discrepancy exists. Closing a discrepancy, adjusting inventory or approving a journal entry is still a person's call. Used well, it saves time: in a survey of nearly 600 accounting professionals, those using AI saved an average of 60 minutes per employee per day (Karbon, 2026). But only 21% of firms had an AI policy, and reconciliation is exactly the control that needs rules most.

Illustrative example: a business gets 420 bank transactions a month, 900 payments through its payment processor and 150 invoices on credit terms. One person spends the first three days of every month lining them up in a spreadsheet and highlighting what matches. Almost everything matches. The real work is in the 30 or 40 lines that don't.

That's where AI fits. Not to "do the reconciliation," but to clear away the 1,400 lines that match and leave you only the ones that don't, each with a likely explanation next to it.

## The three keys: amount, date and reference

Every reconciliation, whether it's bank, processor or inventory, is the same job: two lists that should say the same thing and don't. To decide whether a line on one list is the same as a line on the other, you use three keys.

- Amount: the exact value, or the net value if a fee or withholding was taken along the way.
- Date: the transaction date, with a tolerance window. A Friday payment can show up at the bank on Monday.
- Reference: invoice number, order number, transaction ID, customer name. It's the strongest key and the one most often missing.

The rule: AI can match on three keys without asking, on two it proposes and you confirm, and with one or none the line goes to the discrepancy list. Writing that rule down at the start of every reconciliation is what separates a fast match from a match nobody can review.

> If customers pay without citing the invoice, no AI will guess with certainty. The first fix is almost always in how the data gets recorded, not in how it gets matched.

## What AI does well and what it doesn't

An AI assistant (ChatGPT, Claude, Copilot in Excel, Gemini in Sheets) is good at reading messy files and spotting similarities. It's bad at guaranteeing a total is right, and it knows nothing about your business you haven't told it. Split the work like this:

| Task | AI? | Why |
|---|---|---|
| Reading PDF or CSV statements in different formats | Yes | Converts columns and dates to a common format in minutes |
| Matching by amount, date and reference | Yes, with the three-key rule | Finds matches even when the reference is written differently ("INV-0012" and "invoice 12") |
| Suggesting why a discrepancy exists | Yes, as a hypothesis | Recognizes patterns: fee, batched payment, refund, duplicate charge |
| Adding up and balancing totals | Not on its own | Language models can get arithmetic wrong. Recalculate totals with formulas in the spreadsheet |
| Deciding a discrepancy is resolved | No | It's a judgment with accounting and control consequences; a person signs off |
| Posting adjustments to the books or inventory | No | Whoever adjusts shouldn't be whoever reconciles, let alone a tool with no defined permissions |

Human error doesn't go away either. If your records were keyed in by hand, they already carry errors before they reach the match. In a lab study, entering data once with no automatic check left errors in about 1% of values (Barchard and Pace, 2011). AI doesn't fix that 1%; it turns it into discrepancies someone has to look at. The full math on that cost is in [what manual data entry costs](https://efidis.co/en/blog/cuanto-cuesta-copiar-datos-a-mano).

## Step by step: bank reconciliation with AI

This workflow works for your bank and, with small changes, for your payment processor. Run it on copies of the files, never the originals.

1. **Export both lists for the same period.** The bank statement as CSV or Excel (better than PDF) and your record of receipts and payments for the same month. Write down the opening and closing balance of each: they're your check at the end.
2. **Strip out what AI doesn't need to see.** Remove full account numbers, Social Security or tax ID numbers and individuals' names if you don't need them to match. Use a tool on a business plan whose policy says it doesn't train on your data.
3. **Ask for the match with the three keys.** Give it both files and the prompt below. Ask for a table, not a paragraph: each line with its pair, which keys matched and a status column.
4. **Recalculate in the spreadsheet, not in the chat.** Paste the table into your spreadsheet and total it with formulas. Bank balance plus deposits in transit minus outstanding checks or payments should equal your book balance. If it doesn't, the match has an error, even if the chat says everything ties out.
5. **Review only the discrepancies, then sign off.** One person reviews every line marked "proposed" or "unmatched," decides what to do and leaves their name and the date. Someone else posts the adjustments, or at least there's a record of who did.

### The prompt

Copy it and adjust the column names: "I'm giving you two files for the same month: A is the bank statement and B is our record of receipts and payments. Match each line in A to one in B using three keys: amount (exact), date (3 business day tolerance) and reference (invoice number, order number or customer name, even if written differently). Return a table with these columns: line in A, line in B, matching keys, status (Matched if all three match, Proposed if two match, Unmatched if one or none match) and likely cause only for lines that aren't matched. Don't add up totals. If you don't have the information to explain a discrepancy, write 'no data' instead of guessing."

The last two sentences are the ones that matter: one makes you do the math in the spreadsheet, the other keeps AI from inventing an explanation that sounds right.

## Payment processors: the deposit is never the sale

Reconciling a payment processor (the one that takes cards in your online store or through payment links) has a catch: money doesn't arrive sale by sale. The processor batches one or more days of charges, takes out fees, refunds and chargebacks, and deposits a single net amount a few days later. That's why your bank never matches your sales, even when everything is fine.

Illustrative example: on Tuesday you sell 12 orders for $1,840. The processor takes a $58 fee, deducts a $95 refund from the week before and on Thursday deposits $1,687 as a single line. If you look for $1,840 in the bank, you won't find it.

The fix is to reconcile in two legs. First, orders against the processor's charge report (every sale has its charge). Then, the processor's payout report against the bank (every payout has its bank line). Almost every processor offers a report that breaks each payout into the charges, fees and refunds inside it; that's the file you give the AI, not the dashboard summary.

- [ ] Every paid order in the store has a charge at the processor, and every charge has an order.
- [ ] Processor refunds match returns recorded in the store and in inventory.
- [ ] The month's fees match your agreed rate.
- [ ] Open chargebacks are on a separate list, with the response deadline.

## Inventory: what the system says versus what's on the shelf

Reconciling inventory means comparing the physical count with the system balance. The gap is usually bigger than owners think. In a study of nearly 370,000 inventory records across 37 stores of one U.S. retailer, 65% didn't match what was actually in the store (DeHoratius and Raman, 2008).

Here AI helps little with the count and a lot with the analysis. Give it the count and the system balance, and ask it to group the discrepancies by product, supplier, location and transaction type. If 8 out of 10 discrepancies come from the same supplier or the same items with size and color variants, you know where to look. If the gaps come from online sales, see [how to stop selling what you don't have](https://efidis.co/en/blog/vender-sin-inventario).

- [ ] Freeze transactions while you count, or note the exact cutoff time.
- [ ] Count blind: whoever counts doesn't see the system balance.
- [ ] Check pending receipts and unbilled shipments before writing off a discrepancy as lost.
- [ ] Have someone other than the counter approve adjustments, and write down the reason.

## Table of common discrepancies

Put this table next to the prompt. When you give AI the list of possible causes, its hypotheses get better and your review goes faster.

| Discrepancy | What it looks like | Usual cause | What to do |
|---|---|---|---|
| Deposit in transit | In your books, not in the bank | Payment on the last day that the bank posts later | Flag it and confirm it shows up in the first days of next month |
| Bank fee or charge | In the bank, not in your books | Monthly maintenance fee, wire fee, overdraft charge | Book it with the bank's support |
| Batched payment | One deposit against several invoices | The customer paid three invoices in one transfer | Match one to many; ask the customer for the remittance detail if it's missing |
| Partial payment | Amount is less than the invoice | Partial payment, withholding or unrecorded discount | Apply the payment and leave the balance open in receivables |
| Processor net payout | Deposit doesn't match any sale | Processor took out fees and refunds | Reconcile in two legs with the payout report |
| Chargeback | A processor debit with no refund from you | The customer disputed the payment with their bank | Respond with proof of delivery before the deadline |
| Keying error | Similar amounts that don't tie (1,250 vs. 1,520) | Transposed digits when entered by hand | Fix it at the source; if the difference is divisible by 9, suspect a transposition |
| Inventory item mix-up | One size is over and another is short for the same product | The wrong variant was deducted at sale | Adjust both and review how sales get recorded |

## The controls you can't skip

Reconciliation is a control against both errors and fraud. More than half of the 2,402 occupational fraud cases in the ACFE's 2026 study happened because controls were missing or someone overrode them, and the typical scheme ran 12 months before anyone caught it (ACFE, 2026). If AI speeds up the match but wipes out the control, you've lost.

- [ ] Whoever reconciles doesn't record receipts or approve payments.
- [ ] Opening and closing balances tie out with formulas in the spreadsheet, not with what the chat says.
- [ ] Every closed discrepancy has an owner, a date and support.
- [ ] Open discrepancies roll to next month with their age; they don't get deleted.
- [ ] There's a written rule on which data can go into which AI tool.
- [ ] The owner or manager reviews a sample of automatic matches every month.

## When AI for reconciliation isn't worth it

- When you have few transactions. With 40 lines a month, a spreadsheet with XLOOKUP and an hour of review is faster and easier to audit.
- When your accounting software already reconciles with bank rules. Use those before adding another tool.
- When you can't strip sensitive data out of the files and don't have a tool with clear privacy terms.
- When nobody will review the discrepancies. A match nobody looks at is a control that only exists on paper.
- When the problem is upstream. If references are missing or payments get recorded late, AI will hand you a long discrepancy list every month. Fix where the data is born first.

## When the same discrepancies come back every month

Orders with no payment, payments with no invoice, sales that don't deduct inventory: if they keep coming back, the problem isn't the reconciliation. It's that sales, receivables and inventory live in files that don't talk to each other. When the order, the payment and the inventory movement live in the same system with the same reference, the match is done the moment you sell. What's left comes from outside, the bank and the processor, and for that the three keys and a person who signs off are enough. If you reconcile on behalf of clients, there's a guide to [AI for accounting firms](https://efidis.co/en/blog/ia-para-firmas-contables).

## Frequently asked questions

### Can I do a bank reconciliation with ChatGPT or another AI?

Yes, to match your statement against your records and suggest why discrepancies exist. Don't ask it to add up totals or call the reconciliation done: recalculate totals with formulas in your spreadsheet and have a person review and sign off on every discrepancy. Use a business plan that doesn't train on your data and strip sensitive information before you upload files.

### What is the three-key rule?

A way to decide when two lines are the same transaction: amount, date (with a tolerance of a few days) and reference. If all three match, they're matched automatically; if two match, AI proposes and a person confirms; if one or none match, the line goes to the discrepancy list.

### Why doesn't my payment processor deposit match my sales?

Because the processor batches one or more days of charges, takes out fees, refunds and chargebacks, and deposits the net a few days later. Reconcile in two legs: orders against the processor's charge report, and the payout report against the bank.

## Sources

1. [Karbon, "State of AI in Accounting 2026" (survey of nearly 600 accounting professionals), 2026](https://karbonhq.com/resources/state-of-ai-accounting-2026/)
2. [ACFE, "Key Findings from Occupational Fraud 2026: A Report to the Nations," 2026](https://www.acfe.com/about-the-acfe/newsroom-for-media/press-releases/press-release-detail?s=report-to-the-)
3. [Nicole DeHoratius and Ananth Raman, "Inventory Record Inaccuracy: An Empirical Analysis", Management Science 54(4), 2008](https://doi.org/10.1287/mnsc.1070.0789)
4. [Barchard, K. A. and Pace, L. A., "Preventing human error: The impact of data entry methods on data accuracy and statistical results," Computers in Human Behavior 27(5), 2011](https://doi.org/10.1016/j.chb.2011.04.004)

## Close the month without chasing data

Efidis builds your company's app so every sale is recorded with the same reference from the start. On the call we'll go over your last reconciliation and where the data gets lost.

- [Book my call](https://wa.me/573214646372?text=Hi%2C%20I%20read%20an%20article%20on%20the%20Efidis%20blog%20and%20want%20to%20see%20how%20to%20apply%20it%20in%20my%20company.)
