AI at Workfrom Acuity AI Advisory

Case studies · Law firm, Ireland

Using Claude to reconcile scanned bank statements

By Ger Perdisatt · Reviewed

A law firm in Ireland reconciled its bank accounts by hand in Excel, from statements that existed only as scanned pages ticked in pen. Claude, alongside two other AI models, read every statement page independently, and code applying the firm's own matching rules cross-referenced the results against thousands of lines in the firm's ledger. The output is a reconciliation workbook for each account: matched items showing the rule that matched them, and a short exceptions list for staff to resolve, each pointing to the scanned page it came from. The work is in use and the full-year reconciliation is being completed. Time savings are estimates, and the firm's 20-minute daily reconciliation is a target, not yet a result.

In use In use for the firm's reconciliations. The full-year reconciliation is being completed.

Client identity withheld to respect confidentiality.

The starting point

The firm's bank statements existed only as scanned images, many of them ticked and annotated in pen from earlier manual checks. The ledger sat in the firm's practice management system and could be exported.

Reconciling meant reading the scans, finding each line in the ledger and matching them up in Excel, with older unexplained items carried forward by hand. Finding a single year-end difference had taken the financial controller weeks.

What we implemented

Each statement page was read three times, independently: once by Claude and once each by two AI models from other providers. Every transaction had to agree across all three readings and tie to the running balance on the page. Where the readings disagreed, a person checked the page image. Pen ticks on a scan can be read as extra digits, which is why one reading was not enough.

All arithmetic was done by ordinary code working in whole cents, not by an AI model, so totals are exact and repeatable.

The statement lines were then matched to the ledger export by code that applies the firm's own matching rules, agreed with its finance staff: same amount and date, cheque number, regular payees, and both sides of a transfer between accounts. The rules run the same way every time, unlike spreadsheet lookups that break when a column moves.

Claude Code and other AI coding tools built the matching code and ran it in batches across the year's data.

Using it, step by step

  1. 1What the person supplies

    The scanned statement pages for an account and period, and an export of the ledger for the same account.

  2. 2What happens next

    The pages are read, checked against each other and against the page balances, then matched to the ledger. Staff receive a workbook with the matched lines and a list of exceptions.

  3. 3What they check before using it

    Staff work through the exceptions: they tick matches in their own system, post anything missing, and answer the questions only they can answer, such as what a lodgement was for. Each exception names the statement page to look at.

Example output

One month in the workbook for one account. Matched lines show the ledger posting and the rule that matched them. Exceptions say why the line needs a person and which scanned page to check.

Reconciliation workbook for March on a fictional office account. Matched lines, 212 of 231, list the statement date, line, amount, ledger posting and the rule used, such as same amount and same day or cheque number. Below, 19 exceptions for staff list the reason, such as no receipt found in the ledger or two possible postings of the same amount, with the scanned page number for each.
Illustrative example using fictional data. A month's matched lines and the exceptions left for staff, each linked to its statement page.

Human checks

  • Where the three readings of a page disagree, a person checks the page image before the line is used.
  • Staff approve matches in the firm's own system. We do not post, write off or sign off anything.
  • Exceptions are resolved by the finance team, who know what each payment was for.
  • Progress is reported by the value of money explained, not by the number of lines matched, because a few large batch payments can hold most of the value even when most lines are matched.

Results

In use In use for the firm's reconciliations. The full-year reconciliation is being completed.

  • Measured

    Hundreds of pages of scanned statements turned into checked data.

  • Measured

    Thousands of ledger lines cross-referenced by code applying the firm's own rules.

  • Estimate

    Days of senior time saved on each difference traced. Differences that took weeks by hand are now explained from evidence the team can check. Not timed.

  • Target

    20 minutes a day is the daily reconciliation the firm is working towards. It has not been reached yet.

  • Observed

    Claude in finance The financial controller now uses Claude himself on his own finance questions.

Measured: counted or timed. Estimate: worked out from the old routine, not timed. Target: what the client is aiming for. Observed: seen in use, not a number.

Limitations

  • The full-year reconciliation is still being completed. Some differences remain open and nothing has been signed off.
  • Batch payments and transfers between accounts that the records do not break down line by line still need staff knowledge to explain.
  • Statement images were read by AI models from three providers, Claude among them. A firm doing the same needs to be content with that, and with each provider's data terms.
  • The time savings are estimates, not measured results.

Which Claude product did what

Claude API
One of three AI models that read every statement page independently. Claude called directly from our own code, for one specific step in a larger process.
Claude Code
Built and ran the matching work in batches, alongside other AI coding tools. Anthropic's coding and automation tool. It runs on a computer, can read and write files, write software and carry out multi-step jobs that a person has set up and approved.
Claude app
The financial controller now uses Claude directly on his own finance questions. Claude used directly by a member of staff, in the browser or desktop app.

How Claude Team, Claude Code and software built with Claude differ

Other case studies

Interested in bank reconciliation for your business?

Tell us how the work is done now and who does it. We'll suggest a time for a short call. See also ways to start.

Talk to us about bank reconciliationhello@acuityai.co