SGM Lektra: automating bank reconciliation for an industrial manufacturer
A second project for SGM Lektra, this time outside marketing: bank reconciliation that took days of spreadsheet work, sometimes a full week, now takes about half an hour.
The challenge
SGM Lektra manufactures instrumentation for the measurement and control of industrial processes. The multi-domain website migration is a separate case study; this project was a different kind of work for the same company.
Bank reconciliation means matching every bank transaction against the records in the accounts and chasing down whatever does not match. At SGM Lektra it was done by hand, in spreadsheets. Each reconciliation took days, sometimes a full week, and because every match was made and checked manually, it was prone to errors.
That is the typical shape of work worth automating: mechanical, repeated on a schedule, slow, and risky precisely because people are doing by hand what a system could do consistently.
The approach
No tool was chosen up front. The automation was shaped around how SGM Lektra already worked, in three steps:
-
Mapping the process with the team
The team walked me through how reconciliation ran today, the projects it had to account for, and where the bottleneck was: the days of manual matching in spreadsheets.
-
Proposing and testing an approach
From that map I proposed how the reconciliation could be automated, and we tested the approach together against their real way of working rather than a generic template.
-
Refining it on their feedback
The team’s feedback from testing went back into the build until the automation fitted their process, so the people who run reconciliation were part of designing it.
What I built
Riconcilia is a small web app that reconciles a company’s monthly bank statement against the bank account in its accounting system. It has three parts.
-
Getting the data in
The user creates a period (a month, a quarter or a free interval) and uploads two files: the statement exported from the bank portal, as Excel or as the original PDF, and the ledger printout from their ERP, DAFOWIN. The app recognises both formats from their content, so there is no column mapping to do. Unknown layouts fall back to a manual mapper that remembers the mapping per file type, and files that cannot be reconciled, such as scanned PDFs or exports without dates, are rejected with a message that says what to download instead.
-
Matching: deterministic first, AI last
Three passes run first: exact matches (same amount, date within a few days, one candidate), reference matches (invoice number or counterparty in the description, with a small tolerance for fees), and combinations, where one bank line equals the sum of several ledger entries or the reverse, as with RiBa, SDD and multi-invoice transfers. Learned rules handle recurring fees and counterparty aliases. Only the lines still open after that go to Claude, pseudonymised and with the closest candidates for each, and it returns a proposal with a confidence and a reason, or “no match”. Every proposal, deterministic or AI, stays a proposal until a person accepts it.
-
Reviewing what is left
Unmatched items are split into “bank only” and “ledger only”, each with the reason it stayed open. The user accepts, rejects, matches by hand, or sets a line aside with a category such as fees, stamp duty or interest. Open lines carry forward into the next period, so a payment booked in one month and credited the next closes itself. A CSV export lists every line with its status for the month-end file, and nothing is ever written back to the bank or the ERP.
The stack is deliberately plain: a React and TypeScript front end where parsing and matching run in the browser, Supabase hosted in the EU for data, sign-in and shared company workspaces, and a single server function that calls Claude Haiku with a fixed output schema and logs every request and response. There is no API connection to the bank or the ERP, by design: the finance team keeps full control of what goes in.
The outcome
The headline
Bank reconciliation went from days, up to a week, to about half an hour.
The spreadsheet work that tied up the team for days now runs in about thirty minutes, and the manual matching steps where errors used to creep in are no longer done by hand.
On the first real month, 129 bank movements, the deterministic passes proposed 128 matches and the team accepted 127. What was left were foreign transfer fees and end-of-month bank charges booked in the following month: exactly the items a person should look at.
Next on the list is support for several bank accounts per company, since SGM Lektra works with four banks.
What this case shows
Most of my work is in marketing and SEO, but the method does not depend on the department. Find the work a team repeats by hand, separate the mechanical steps from the ones that need judgement, automate the first and give people better tools for the second. Bank reconciliation is a finance version of the same problem as briefing, reporting or refreshing content.
Is your team doing the same work by hand every week?
The automation audit maps where the hours go and which automations are worth building, in what order, and what each costs to run. It starts in marketing and SEO, and covers any repetitive process.