How to Automate Bank Reconciliation with AI (Step-by-Step)

Bank reconciliation is boring. I don’t mean that as a complaint — I mean it as a technical observation. It’s repetitive, pattern-based, and follows predictable rules. That makes it one of the most automatable tasks in accounting.

And yet most accountants still do it by hand, line by line, staring at two lists and trying to match them with tired eyes at 4 PM on a Friday.

I stopped doing that eight months ago. Here’s the workflow I use now, step by step, with the exact prompts and the exact verification checks that keep it safe.

The Reality Check: What AI Can and Can’t Do Here

Before we start, let’s be clear about scope. AI doesn’t log into your accounting software and press the “reconcile” button. It can’t access your bank feed directly. It doesn’t replace the actual posting of reconciled entries.

What AI does is the analysis layer — the matching, the investigation, the exception identification. The part that eats 70% of your reconciliation time.

You still export data. You still review results. You still post entries. But the middle part — the actual brain work of comparing two lists — goes from an hour to 15 minutes.

Step 1: Prepare Your Data (5 Minutes)

Clean data in, clean results out. This step is boring but non-negotiable.

From your bank: Export the monthly statement as CSV. You need: date, description, amount (positive for deposits, negative for withdrawals or vice versa — just be consistent).

From your ledger: Export the bank account transactions for the same period. Same columns: date, description/memo, amount.

Clean up before pasting:

• Remove header rows and summary rows
• Make sure dates are in the same format (MM/DD/YYYY or YYYY-MM-DD, pick one)
• Make sure amounts use the same sign convention
• Remove any client-identifying information if using a consumer AI plan (client name, account numbers)

This prep takes 5 minutes and prevents 30 minutes of AI confusion downstream.

Step 2: Run the Matching Prompt (2 Minutes)

Here’s the prompt I use:

“I have two lists for [MONTH YEAR]: bank statement transactions and general ledger entries for the same bank account. Compare them and produce four tables:

Table 1 — MATCHED: Transactions that appear in both lists (matched by amount and approximate date, within 3 business days). Columns: Bank Date, Bank Description, Bank Amount, Ledger Date, Ledger Description, Ledger Amount, Match Confidence (High/Medium).

Table 2 — UNMATCHED BANK: Bank transactions with no matching ledger entry. These could be outstanding items or missing entries. Columns: Date, Description, Amount, Possible Explanation.

Table 3 — UNMATCHED LEDGER: Ledger entries with no matching bank transaction. Columns: Date, Description, Amount, Possible Explanation.

Table 4 — DISCREPANCIES: Matched items where amounts are close but not exact (within $5). Columns: Bank Amount, Ledger Amount, Difference, Possible Cause.

At the end, provide a summary: total bank balance, total ledger balance, difference, and number of items in each table.

Bank transactions:
[PASTE BANK CSV]

Ledger entries:
[PASTE LEDGER CSV]”

Copy, paste, run. You’ll get four structured tables back in about 30-60 seconds.

Step 3: Review the Matched Items (5 Minutes)

Don’t skip this. AI matching is good but not perfect.

Quick review protocol:

1. Check the total. Add up all matched amounts on the bank side and the ledger side. Do they balance? If yes, you’re probably fine. If not, there’s a mismatch hiding in the details.
2. Spot-check 5-10 items. Pick a few at random and verify the match makes sense. Does the description align? Is the date reasonable?
3. Look at Medium confidence matches. These are the ones where AI wasn’t sure. They’re usually right, but review each one. Common issues: split transactions matched to a lump sum, or two similar-amount transactions swapped.

For a typical month with 100-200 transactions, this review takes 5 minutes. For 500+, budget 10-15 minutes.

Step 4: Investigate Unmatched Items (10-15 Minutes)

This is where your expertise matters. The unmatched items fall into predictable categories:

Unmatched bank items (on bank, not in ledger):

• Bank fees and interest — often not entered until reconciliation. Book them now.
• Deposits in transit — entered in ledger before month-end, cleared at bank after. Document and carry forward.
• Errors — something was missed during the month. Enter it now.
• Unauthorized transactions — rare but important to catch. Flag and investigate.

Unmatched ledger items (in ledger, not on bank):

• Outstanding checks — written but not yet cashed. Normal. Document and carry forward.
• Timing differences — ACH payments initiated near month-end. Will clear next month.
• Voided or reversed entries — check if the original bank transaction is also absent. If so, clean.
• Posting errors — entry posted to wrong account. Correct and re-run.

For each unmatched item, decide: book it, document it as outstanding, or investigate further. This is the judgment work that AI can’t do and you get paid for.

Step 5: Reconciliation Summary Prompt (1 Minute)

Once you’ve resolved all items, use AI to draft your reconciliation summary:

“Prepare a bank reconciliation summary for [ACCOUNT] as of [DATE]:

Bank balance per statement: [AMOUNT]
Add: Deposits in transit: [LIST]
Less: Outstanding checks: [LIST]
Adjusted bank balance: [AMOUNT]

Book balance per ledger: [AMOUNT]
Add: [ADJUSTMENTS]
Less: [ADJUSTMENTS]
Adjusted book balance: [AMOUNT]

Confirm these balance. If they don’t, identify the remaining difference and suggest where to look.”

This gives you a clean reconciliation document you can put in your workpapers.

Step 6: Final Verification (3 Minutes)

Before you sign off:

1. Do the adjusted balances match? If not, you have unresolved items. Go back to Step 4.
2. Are all outstanding items documented? Outstanding checks and deposits in transit need to be tracked for next month.
3. Are all adjusting entries posted? Bank fees, interest, and corrections should be in the ledger before you close.
4. Does the reconciliation tie to last month? Prior month’s outstanding items should either have cleared or still be outstanding with explanation.

Time Comparison

For a typical small business client with 150 transactions per month:

Manual reconciliation:

• Data prep: 10 minutes
• Line-by-line matching: 45 minutes
• Investigating exceptions: 20 minutes
• Documentation: 15 minutes
• Total: ~90 minutes

AI-assisted reconciliation:

• Data prep: 5 minutes
• AI matching + review: 7 minutes
• Investigating exceptions: 12 minutes
• AI summary + verification: 4 minutes
• Total: ~28 minutes

That’s 62 minutes saved per client per month. For a bookkeeper with 10 clients, that’s over 10 hours a month. Just from reconciliation.

Batch Size Matters

One thing I learned the hard way: AI accuracy drops when you paste too many transactions at once.

For reconciliation, my limits:

• Under 100 transactions: Paste everything at once. AI handles it cleanly.
• 100-300 transactions: Split into two batches. Run the prompt twice, then combine results.
• 300+ transactions: Split into weekly batches or by transaction type. More work up front, but far fewer matching errors.

Pushing 500 transactions into one prompt is tempting but produces noticeably worse matching. The extra 5 minutes to split is worth it.

When This Doesn’t Work

AI-assisted reconciliation works well for most small and mid-size business accounts. It doesn’t work as well for:

• High-volume accounts (1000+ transactions/month) — you’re better off with dedicated reconciliation software that integrates directly with your bank feed.
• Multi-currency accounts — AI handles the matching but struggles with exchange rate differences. You’ll need to manually verify any FX-related discrepancies.
• Accounts with heavy internal transfers — transfers between company accounts create matching confusion because similar amounts appear on both sides. Flag these separately before running the prompt.

For everything else — the standard operating account, the payroll account, the savings account — this workflow handles it cleanly.

Start With One Account

If you haven’t tried AI-assisted reconciliation before, pick one client’s operating account next month. Run the workflow alongside your normal process the first time — compare AI’s matching against your own.

You’ll find that AI catches 90%+ of matches correctly, flags the same exceptions you would have found, and does it in a fraction of the time. After one month of parallel testing, you’ll trust it enough to lead with AI and review manually.

That’s when the time savings become real.

The reconciliation prompt and 9 other bookkeeping-specific prompts are in the free PDF, plus a verification checklist for every workflow. Download it here.