AI-Assisted Bank and Account Reconciliations

Reconciliations are the part of the close that consumes a disproportionate share of time, especially when there are unmatched items, foreign currency wrinkles, or volume that has outgrown the original spreadsheet. AI is not a replacement for your matching engine, but it is excellent at the parts you currently do by hand: classifying unmatched items, drafting the reconciliation narrative, explaining variances to reviewers, and turning a messy bank statement into a clean schedule.

This lesson gives you four concrete reconciliation workflows and the prompts that drive them.

What You'll Learn

• The four reconciliation tasks where AI saves the most time
• A copy-paste workflow for unmatched items
• How to draft a reconciliation summary that holds up under review
• Where AI helps and where it actively hurts in reconciliations

Workflow 1: Triaging Unmatched Items

After your matching engine has done its work, you typically have a list of unmatched bank or general ledger items. Some are timing differences, some are coding errors, some are missing transactions, and a few are genuinely unknown. Triaging them is a chunk of work that AI can compress to minutes.

The setup: copy the unmatched list into your AI scratch tab. Strip any sensitive descriptions. Paste this prompt into ChatGPT Business or Claude Team:

"Act as a senior accountant. Below is a list of unmatched bank reconciliation items with description, amount, and date. For each item, categorise as: (a) likely timing difference, (b) likely coding error, (c) likely missing GL entry, (d) requires investigation. Then group by category and list the items that need the most urgent follow-up first. End with a 3-bullet summary I can paste into my close memo."

Then paste 20 to 200 rows of unmatched items. The AI returns a categorised list and a draft summary. You spend your time investigating the items that actually need investigation, not staring at obvious timing differences.

A real example output you might get:

"Likely timing differences (no action): Stripe payouts dated 30 March cleared 1 April (3 items, total $24,300). Likely missing GL entries: 2 unmatched bank credits from a customer that does not appear in the AR ledger (Customer A, $8,200 and $3,150) — recommend confirming with AR team. Requires investigation: 1 unmatched debit of $1,425 with description 'WIRE FEE ADJUSTMENT' — no corresponding GL entry; recommend asking the treasury team. Summary: 3 timing differences, 2 likely missing AR receipts, 1 unexplained fee."

That summary, lightly edited, becomes part of your reconciliation working paper.

Workflow 2: Bank Statement to Schedule

The second high-time-saving workflow is converting an unstructured bank statement PDF or CSV into a clean schedule. The trick is that AI is not reliable for absolute number extraction at scale — but it is excellent at restructuring data you have already pasted as text.

Workflow:

1. Open the bank statement as a PDF or download as CSV
2. If PDF, copy the relevant pages as text
3. Paste into Claude or ChatGPT with this prompt:

"Below is text copied from a bank statement. Restructure as a clean table with these columns: Date, Description, Debit, Credit, Running Balance. Flag any rows where the running balance does not reconcile to debit/credit movement against the prior row. Output as a markdown table."

The AI will give you a clean schedule and flag arithmetic anomalies. You always — always — tie the closing balance back to the source statement. AI can be confidently wrong by a digit or two. The flagging step in the prompt is the one that catches its own mistakes.

Workflow 3: Drafting the Reconciliation Narrative

Most reconciliation working papers need a written summary: what the account is, what was reconciled, what the unreconciled balance is, and why it is acceptable. AI is excellent at this.

The prompt:

"Act as a finance manager preparing a reconciliation working paper. Draft a 120-word reconciliation summary for [Account Name] as of [date]. GL balance: [X]. Subledger or bank balance: [Y]. Reconciling items: [list]. Net unreconciled balance: [Z]. Explain in plain English: the purpose of the account, the reconciliation process applied this month, the nature of reconciling items, and whether the unreconciled balance is within tolerance of [tolerance amount]. Tone: professional, audit-defensible, no marketing language."

This produces a tight summary that fits at the top of the reconciliation tab and reads identically to what a senior accountant would write — because it has been trained on exactly that style of writing.

Workflow 4: Intercompany and Multi-Currency Reconciliations

Intercompany reconciliations are notorious time-sinks because of FX timing and rounding. Here AI helps in a slightly different way — by explaining the differences rather than fixing them.

Prompt pattern:

"I have an intercompany imbalance of [amount] between Entity A (functional currency USD) and Entity B (functional currency GBP). Entity A booked the transaction on [date 1] at FX rate [rate 1]. Entity B booked it on [date 2] at FX rate [rate 2]. The underlying invoice was for GBP [amount]. Break down the imbalance into FX timing difference, rounding, and any other component. Show the math step by step."

The AI will walk you through the reconciliation arithmetic. This is gold for training junior staff and for explaining the variance to a non-finance reviewer. You still book the adjustment yourself; you have just saved 30 minutes of explanation.

Where AI Hurts Reconciliations

Three failure modes you should know.

Failure 1: Trusting AI extraction of numbers from images. Photos of bank statements, scanned PDFs, and screenshots are not reliable inputs. The error rate on OCR-then-AI pipelines is too high for reconciliations. If you do not have machine-readable data, do not use AI to extract numbers. Use proper OCR software and verify against the source.

Failure 2: Asking AI to "do the reconciliation". This is the prompt that goes wrong. AI does not know your tolerance thresholds, your account substance, or your accounting policy. Ask it to triage, draft, and explain — not to perform the reconciliation.

Failure 3: Pasting raw bank data into a consumer-tier tool. Bank statements contain account numbers, vendor names, and patterns that reveal your business. Use Business or Team tier minimum. Redact account numbers and large customer names.

A 30-Minute Pilot

If you have not used AI on a reconciliation yet, try this 30-minute pilot.

• Pick one moderately painful reconciliation from last month (10 to 30 unmatched items is ideal)
• Open a Business or Team account chat
• Run Workflow 1 (triage) and Workflow 3 (narrative)
• Compare the AI summary to what you would have written
• Time how long the manual version took versus the AI-assisted version

For most teams the first pilot saves 30 to 60 percent of the time. That gap widens as your prompts get sharper.

Saving Your Reconciliation Prompts

Build a section in your prompt library called "Reconciliations". Drop in the four prompts above as starting templates. Mark which subledger or account each one was tuned for: AR aging, AP, payroll clearing, intercompany, bank, fixed-asset register. Within six months you will have a personal library that turns a 4-hour reconciliation review into a 90-minute one.

Key Takeaways

• AI is excellent at triaging unmatched items, drafting narratives, and explaining FX differences
• Always tie AI-restructured numbers back to the source — AI can be confidently wrong on digits
• Do not ask AI to perform a reconciliation; ask it to triage, draft, and explain
• Use Business or Team tier minimum, and redact account numbers and large customer names
• A first pilot typically saves 30 to 60 percent of time on a moderate reconciliation