Customers
The close got shorter. That was never the interesting part.
What operators tell us about, months in, is different from what they bought. They bought a match rate. What they talk about is the break they found in parallel running, the acquirer conversation they had four days early, and the audit finding that closed.
Case study 01
Kestrel Pay
Payments platform · 340k transactions/month
Ireland
- 11 → 4
- working days to close
- 98.6%
- of lines matched without a human
- 4
- carried-forward breaks found in parallel running
- 5 weeks
- parallel run before cutover
Where they started
Three acquirers, two banks and an internal ledger that booked card captures gross on authorisation day. Settlements arrived T+2, net of merchant discount, with references that never survived the acquirer's own file format. The team reconciled in a spreadsheet that had grown to eleven tabs, and the monthly close had settled at eleven working days.
What we did
Finlecy ingested all six sources in their native formats without changes upstream. Fee bands were configured per acquirer from the actual contracts rather than a single global tolerance, and the value-date window was set per channel. The old spreadsheet ran in parallel for five weeks while the two were compared line by line.
What changed
The parallel run surfaced four breaks the spreadsheet had been carrying forward since March, including a duplicated payout from a retried API call. Close moved to four working days. The eleven-tab spreadsheet was retired and the two analysts who maintained it now work on merchant disputes.
The thing that actually changed was not the match rate. It was that every break arrived with a reason code and a drafted entry.
Aitor Zubeldia
Head of Finance Operations, Kestrel Pay
Case study 02
Rueda Marketplace
Marketplace · 41k sellers across four countries
Spain
- 212
- instructions matched to a single bank line
- 9 days
- first file to parallel running
- 0
- seller-reported payout discrepancies since cutover
- 4
- countries on one reconciliation model
Where they started
Seller payouts left the account as bulk SEPA files, so one bank debit corresponded to anywhere between forty and three hundred ledger instructions. No one-to-one matching tool could see the relationship, and the operations team resolved it by trusting the file total and investigating only when a seller complained.
What we did
The aggregation pass was configured to reconstruct batches by bounded subset-sum within a one-day window, with bank charges allowed as a residual posted to its own account. Signals was switched on at the same time to baseline payout timing per country.
What changed
Batches now reconstruct automatically, including the ones where a single instruction was rejected and the file total no longer equals the ledger total. That failure mode — previously invisible until a seller complained — is now a break on the day it happens.
We kept both processes for a month, compared them line by line, and found four breaks the old one had been carrying since March.
Idoia Larrañaga
Finance Systems Lead, Rueda Marketplace
Case study 03
Halcyon Lending
Consumer lender · direct debit collections
United Kingdom
- 6 days → same day
- return applied to customer balance
- 100%
- of restorations under four-eyes approval
- 1
- internal audit finding closed
- 34
- reason codes trended monthly
Where they started
Collections were reconciled weekly, which meant an unapplied return could sit for six days while the customer balance showed as settled. With a regulated book, the control gap mattered more than the operational cost, and the internal audit function had raised it twice.
What we did
Daily ingestion of the collection file and the return file, with the return reason code carried through to the exception rather than flattened into a generic mismatch. Assure was configured to require four-eyes approval on any posting that restores a customer receivable.
What changed
Returns are now applied the day they arrive, with the scheme reason code attached. The audit finding was closed at the next review, and the retry decision for a returned collection is made against accurate data rather than a six-day-old balance.
The model does not decide anything. It proposes and it explains, and a deterministic layer checks it before a human sees it.
Nuala Brennan
Chief Risk Officer, Halcyon Lending
In their words
Shorter answers, from more people
We were closing on the eleventh working day and everybody had quietly accepted that. The thing that actually changed was not the match rate — it was that every break arrived with a reason code and a drafted entry, so the queue stopped being an investigation and started being a review.
Aitor Zubeldia
Head of Finance Operations, Kestrel Pay · Ireland
Our acquirer settlements never matched because our ledger books gross and they pay net. Four years of a spreadsheet that one person understood. Finlecy handled it in the first run, and more importantly it posted the merchant discount to its own account instead of quietly eating it out of revenue.
Marta Sequeira
Financial Controller, Novabanc · Portugal
The bulk payout problem is the one nobody solves. One debit on the statement, two hundred rows in our ledger, and every tool we trialled asked an operator to match them by hand. Watching the aggregation pass reconstruct a file from amounts alone was the moment we stopped evaluating vendors.
Joris van Weel
Director of Payment Operations, Volta Mobility · Netherlands
I care about one thing: can I show an auditor what was known on a given day and who decided what to do about it. Replay at the original engine version answered that in a way no other tool we looked at could.
Chiara Bettoni
Group Financial Controller, Aurelio Capital · Italy
Our reconciliation used to be a rota nobody wanted. Two analysts, every morning, comparing exports. They now spend that time on merchant disputes, which is work that actually pays for itself.
Séverine Rochat
VP Operations, Comercia Digital · Spain
Signals flagged that one acquirer had drifted from T+2 to T+4 on the second late file. We had a conversation with them that week instead of finding a hole in the float forecast three weeks later.
Peter Halvorsen
Treasurer, Meridian Treasury · Germany
What sold the team was that the model does not decide anything. It proposes and it explains, and a deterministic layer checks the arithmetic before a human sees it. That distinction matters enormously when you are regulated.
Nuala Brennan
Chief Risk Officer, Halcyon Lending · United Kingdom
We run six currencies and the FX legs were always the argument. Now the conversion drift either sits inside a stated tolerance and gets revalued, or it becomes a break with the effective rate printed on it. The argument ended.
Andrei Lupu
Head of Treasury, Tigris FX · Malta
Implementation took nine days from first file to running in parallel with our old process. We kept both for a month, compared them line by line, and found four breaks the old process had been carrying since March.
Idoia Larrañaga
Finance Systems Lead, Rueda Marketplace · Spain
The tolerance sliders sound like a small feature. They are not. Being able to show the audit committee exactly what a wider fee band would have let through, before approving it, changed how we make those decisions.
Tomasz Wierzbicki
Internal Audit Manager, Novabanc · Portugal
Every one of these started with one day of real data
Not a questionnaire, not a discovery workshop. A statement, a settlement file and a ledger export, run through the engine while you watch.