What happens to a slip the OCR cannot read?

It is routed as a capture failure, not as a cash difference. The distinction is deliberate: one is fixed by improving the image at source and one needs an investigation into the cash itself, and mixing them is how a reconciliation backlog becomes permanent.

Every bank prints a different slip. How is that handled?

Each bank has an extraction profile covering its layout and denomination set, and a rule set covering derivations, counter behaviour and tolerance. Adding a bank is configuration. There is no assumption anywhere that slips will converge on a common format.

Why match denomination by denomination rather than on the total?

Because two errors that cancel out produce a correct total and an incorrect cassette. A denomination-level comparison catches that; a total-level one certifies it as reconciled.

What does the confidence score change in practice?

It decides what a person looks at. A score per field means review is targeted at the values that were genuinely uncertain, instead of at whole documents flagged because one column was smudged.

Can the tolerances and rules be changed without us raising a ticket?

Yes. Derivation checks, counter behaviour and tolerances live in the rule engine and are configured by the team that owns the reconciliation. Changes are recorded, so a shift in outcomes can always be traced to the rule change that caused it.

Does this only apply to ATM replenishment?

It was built on a national ATM estate, where the volume and the format spread are hardest. The same shape applies wherever a physical slip has to agree with a record someone else holds — counter takings, deposit slips, vault transfers.

← All solutions

Solutions

Slip image to reconciliation status, without the manual read

Cassette and audit slips captured denomination by denomination with a confidence score on each field, absorbed into per-bank profiles and a configurable rule engine, and matched against custodian records — with unreadable slips and disagreeing slips kept apart.

Book a demo

Per-bank extraction profilesConfidence score on every fieldDenomination-level matching

~40,000

ATMs in the estate

A national footprint, replenished by teams working to different bank formats.

~160,000

Slip documents a day

Cassette and audit slips from every replenishment, previously read by hand.

~89%

Reconciled on clean images

Where capture quality is good. The remainder is separated by cause rather than pooled into one backlog.

Cash & bank

Financial servicesLogisticsOCR & extractionN-way matchingERP write-backAudit trail

The challenge

Every replenishment leaves cassette and audit slips behind, and someone has to read them and match them against custodian records and the transaction system. For a national cash management network that manual read is the bottleneck, and it is also where reconciliation errors enter: a denomination transposed on a busy afternoon becomes a cash difference that is investigated weeks later. Slip layouts and denomination sets differ bank to bank, so there is no single format to standardise on and no prospect of persuading every bank to print the same thing. At estate scale the volume is high enough that any backlog compounds daily rather than clearing overnight.

What we built

Slip image to reconciliation status, automatically. OCR captures denomination-wise counts and totals before and after loading, with a confidence score attached to each field rather than to the document as a whole. Per-bank extraction profiles and a configurable rule engine absorb the layout differences, the derivation checks, the counter behaviour and the tolerances each bank works to, so a new format is configuration rather than a code change. Cases are matched denomination by denomination against custodian records — not on the total alone, because two errors that cancel out in the total are still two errors, and only a denomination-level comparison finds them.

The result

Around 89% of cases reconcile automatically where the image is clean, and the rest are separated by cause instead of pooled. Cases that could not be read reliably are kept apart from cases that were read and genuinely disagree, which turns a single reconciliation backlog into two queues with different owners: a capture problem to fix at source, or a cash difference to investigate. Every field carries its confidence score and every case points back to the slip image it came from, so a reviewer resolves the case against the document rather than against a summary of it.

Process

How it works

Four stages between a slip image and a reconciliation status, with the bank-specific behaviour held as configuration rather than built into any of them.

01
Ingest
Cassette slips, audit slips and the custodian records for the same replenishment are brought together as one case. Reconciling a slip in isolation is not possible: the question is always whether what the slip says agrees with what the custodian recorded for that machine on that visit.
02
Read
OCR captures denomination-wise counts and the totals before and after loading. A confidence score is attached to each field rather than to the page, so a single smudged column marks that column for review instead of condemning an otherwise good slip or, worse, passing quietly with a wrong figure.
03
Apply bank rules
Per-bank extraction profiles handle the layout and the denomination set. A configurable rule engine handles the rest: which values are derived rather than printed, how the counters behave, and what tolerance applies. Adding a bank is a configuration exercise, which is what makes the approach survive a changing estate.
04
Reconcile
Cases are matched denomination by denomination against custodian records, within the tolerance configured for that bank. What matches is closed. What does not is split into the two failure types — unreadable, or read and disagreeing — and routed to the queue that can act on it.

What it does

Inside the solution

Denomination-wise capture

Counts by denomination and totals before and after loading, read from the slip as printed. Reconciling on the total alone hides two offsetting errors inside a figure that looks correct, so the comparison is made at the level the cash actually moves.

Confidence on each field

A score travels with every extracted value rather than with the document. One unreadable column marks that column for review while the rest of the slip proceeds, which is the difference between a targeted check and a re-read of the whole page.

Per-bank extraction profiles

Slip layouts and denomination sets differ bank to bank and no standard format is coming. Each bank gets a profile describing what its slip looks like, so a new bank in the estate is onboarded rather than developed for.

A configurable rule engine

Derivation checks, counter behaviour and tolerances are configuration, not code. The rules that decide whether a case reconciles can be read and changed by the team who owns the reconciliation, and every change is recorded against the cases it affected.

Two queues, not one backlog

A slip that could not be read and a slip that was read and disagrees are separated at the point of failure. The first is a capture problem with an owner at source; the second is a cash difference to investigate.

Case-level traceability

Every reconciled or exception case points back to the slip images and the custodian record it was built from, so a reviewer works against the documents rather than against a summary that has already interpreted them.

Estate-scale throughput

The flow runs against a national footprint of around 40,000 machines and roughly 160,000 slip documents a day. Volume shapes the design: anything requiring a person per document does not survive contact with that number.

Write-back and audit trail

Reconciled outcomes are written back to the systems that hold the cash position, and the status of each case, its confidence scores and the rules applied are retained for audit long after the slip itself is archived.

DESIGN PRINCIPLE

Two kinds of failure

A slip that could not be read and a slip that was read and disagrees are different problems with different owners. Merging them is why manual reconciliation queues never shrink: the queue mixes work that a better photograph would remove entirely with work that genuinely needs an investigation, and neither gets the attention it needs.

Could not be read is a capture problem: Low confidence on a field points at the image, the printer or the way the slip was photographed. The fix belongs at source, and counting these cases by machine and by team is what makes that fix findable.
Read and disagreeing is a cash question: The slip was captured correctly and it does not agree with the custodian record. That is a difference to investigate, and it deserves a person — which it will not get if it is sitting behind a thousand blurry images.
Tolerance is configuration, not judgement: What counts as agreement differs by bank. Holding the tolerance in the rule engine means the same case is decided the same way every day, and that a change in policy is a change in one place with a record of when it took effect.

Foundations

The VapusFin models behind it

Cash Position & Forecasting

Cash Ledger

Live balances across every account and entity, continuously reconciled to bank feeds and the general ledger, with a forward view built from what is actually scheduled.

Details

Continuous Reconciliation

Reconciliation

Match bank, ledger, intercompany and tax accounts continuously rather than at period end, and investigate breaks with their likely cause already attached.

Details

Questions

Frequently asked

What happens to a slip the OCR cannot read?: It is routed as a capture failure, not as a cash difference. The distinction is deliberate: one is fixed by improving the image at source and one needs an investigation into the cash itself, and mixing them is how a reconciliation backlog becomes permanent.
Every bank prints a different slip. How is that handled?: Each bank has an extraction profile covering its layout and denomination set, and a rule set covering derivations, counter behaviour and tolerance. Adding a bank is configuration. There is no assumption anywhere that slips will converge on a common format.
Why match denomination by denomination rather than on the total?: Because two errors that cancel out produce a correct total and an incorrect cassette. A denomination-level comparison catches that; a total-level one certifies it as reconciled.
What does the confidence score change in practice?: It decides what a person looks at. A score per field means review is targeted at the values that were genuinely uncertain, instead of at whole documents flagged because one column was smudged.
Can the tolerances and rules be changed without us raising a ticket?: Yes. Derivation checks, counter behaviour and tolerances live in the rule engine and are configured by the team that owns the reconciliation. Changes are recorded, so a shift in outcomes can always be traced to the rule change that caused it.
Does this only apply to ATM replenishment?: It was built on a national ATM estate, where the volume and the format spread are hardest. The same shape applies wherever a physical slip has to agree with a record someone else holds — counter takings, deposit slips, vault transfers.

Want the same outcome?

Send us a day of slips across two or three banks and we will show you what read cleanly, what did not, and which cases genuinely disagree with your custodian records.

Book a demo

Last reviewed 4 September 2026

Slip image to reconciliation status, without the manual read

The challenge

What we built

The result

How it works

Ingest

Read

Apply bank rules

Reconcile