How much weight should price get in a cheque scanner evaluation?

No more than 10% in the first pass. Cheap cheque software is usually expensive after exception handling, integrations, reconciliation defects, and audit remediation. Score recognition accuracy, workflow controls, and fraud handling first.

How many cheques do I need in a blind OCR test?

A minimum of 2,000–5,000 items covering personal, business, cashier’s, and dividend cheques in printed, handwritten, and mixed forms. For bank-grade decisions, use 25,000+ items. Build the ground truth with two independent human keyers, never let the vendor create the truth file.

What is the most dangerous metric in cheque OCR?

False accept rate — the system accepting an incorrect read as correct. It creates financial loss directly. A high-confidence wrong recognition is more dangerous than low accuracy, because it bypasses the exception queue entirely.

What should a vendor demo live?

At minimum: a clean straight-through item, a poor-quality image routed to rescan, a CAR/LAR mismatch through exception and approval, a duplicate deposit detection, a high-value item requiring a second approver, a bank rejection routed to repair, and a simulated integration outage that recovers without duplicate posting.

What is straight-through processing (STP) and why does it matter?

STP is the share of cheques processed end-to-end without human intervention. It is the best single proxy for operating cost, but it must be measured with false-accept monitoring — raising STP by accepting risky reads is how losses happen.

Procurement Guide

Cheque Scanner Evaluation Guide

Cheque processing software should be evaluated as a controlled financial-processing system, not a document-capture tool. OCR accuracy matters, but procurement succeeds or fails on four harder questions: can it create a reliable, auditable item from imperfect input; can it prevent high-cost mistakes; can it route exceptions with enforceable controls; and can it integrate cleanly with your bank, core, ERP, and archive.

1. Start with your use case, not a generic RFP

The same category covers very different buyers. Pick the row that matches your operation and evaluate against it:

Use case	Buyer	Key evaluation dimension
Bank branch capture	Banks, building societies	Clearing compliance, scanner support, teller controls
Remote deposit capture	Banks, corporates	Risk controls, duplicate detection, entitlements
Lockbox / remittance processing	Banks, BPOs, corporates	OCR + remittance matching + ERP integration
Corporate deposit automation	Treasury, finance	Bank connectivity, AR matching, approval controls
Back-office cheque clearing	Financial institutions	File standards, throughput, operational resilience
Mailroom / back-office capture	Back offices, shared services	Station sizing, batch workflow, cut-off management

2. Use a weighted scorecard

A suggested starting weighting — adjust to your own risk profile, but keep price out of the top two slots:

20%Recognition accuracy and validationDetermines straight-through processing and exception cost

15%Workflow controls and operational governancePrevents unauthorized correction, posting, release, and override

15%Fraud and risk controlsProtects against duplicates, alterations, stale items, and mis-posting

15%Integration architectureDetermines implementation cost and long-term maintainability

10%Standards and compliance fitCritical for bank channels, image exchange, audit, and retention

8%Performance and resilienceMatters at peak deposit windows and batch cut-offs

7%User experience and exception productivityDrives real operating cost

4%Reporting, MI, and analyticsNeeded for SLA, audit, fraud, and process improvement

4%Vendor viability and roadmapReduces platform and support risk

2%Commercials and TCOImportant, but should not dominate a control-critical system

3. Accuracy is not one number

Cheque recognition is a set of field-level reads with different risks. Evaluate each one separately:

Component	Reads	Risk if wrong	Metric
MICR	Sort/routing, account, cheque number	Posting to wrong account, failed clearing, returns	Field exact-match rate; false accept rate
OCR	Machine-printed text: payee, memo, amounts	Payee, memo, remittance, printed amount	Field exact-match and confidence calibration
ICR	Handwritten text	Handwritten payee, date, amount	Exact match; review-routing accuracy
CAR	Courtesy amount (numeric box)	Wrong amount posting	Exact match; false positive rate
LAR	Legal amount (amount in words)	Amount validation against CAR	CAR/LAR agreement rate
IQA	Image quality: skew, blur, crop, front/back	Rejected images, unreadable archive	IQA pass/fail, rescans, downstream rejects

The most dangerous failure mode is not low accuracy — it is high-confidence wrong recognition. A false accept bypasses the exception queue and posts money where it should not go. Ask every vendor for their false accept rate, and expect a real answer.

4. Run a blind test on your own cheques

Build a representative set — personal, business, cashier's, dividend, and refund cheques; printed, handwritten, and mixed; branch, desktop, lockbox, and archive image sources; quality edge cases like folds, skew, stamps, and faded ink; operational edge cases like post-dated and stale items; and fraud samples (altered payee, washed cheques, number reuse). Minimum 2,000–5,000 items; bank-grade decisions need 25,000+.

Ground truth: two independent human keyers, then adjudicate mismatches. Never let the vendor create the truth file.
Track character accuracy, but decide on field exact-match and item-level accuracy — they predict straight-through processing.
Measure exception rate and correction time; they are your operating-cost drivers.
Measure duplicate detection recall — the fraud exposure you actually carry.

5. Workflow controls separate a tool from a platform

The audit-friendly controls that keep a capture operation defensible:

Role-based access & segregation of duties

Role-based queues and actions; a user who captures or corrects an item cannot approve or release it above thresholds; maker-checker review for high-value, corrected, and overridden items; field-level permissions; break-glass access that is time-limited and logged.

Exception queue management

Separate queues for capture, image quality, recognition, amount mismatch, duplicates, fraud review, high value, returns, and unbalanced batches — with ageing, cut-off timers, and escalation.

Audit trail that reconstructs everything

Original image hash, enhanced image, raw read, corrected value, user, before/after values, timestamp, device, approval decision, rule triggered, override reason, export status, downstream response. Every decision, automated or manual.

Fraud controls

Exact MICR+amount+date duplicate detection; cheque-number reuse alerts; same-image-different-metadata; cross-channel duplicates; CAR/LAR mismatch; payee and amount alteration indicators; positive pay integration with maker-checker exceptions.

6. The vendor demo script

Require these scenarios live. A vendor that cannot demo them is not ready for production scrutiny:

1Clean straight-through item: capture → MICR → amount → validate → batch → export → audit
2Poor image (skewed, low contrast, missing rear) → quality rejection → rescan → audit
3CAR/LAR mismatch → exception → operator corrects → supervisor approves → audit
4Duplicate deposit (same cheque twice) → duplicate candidates → decision workflow
5High-value item above threshold → requires second approver → same-user rejection
6Bank/core rejection → repair queue → resubmit without duplicate
7Integration failure (simulated ERP outage) → retry → queue → alert → reconciliation

7. Red flags

Claims "99% OCR accuracy" without field-level evidence

Cannot report false accepts

Cannot separate raw OCR values from corrected values

Cannot process your blind test set

Require broad admin permissions for daily operations

Lack maker-checker controls

Lack immutable audit history

Cannot replay failed batches safely

Cannot prove idempotency (resubmitting an item must not double-post)

No duplicate detection beyond exact cheque number matching

Treat signature verification as definitive proof

Cannot name the cheque/image standards they comply with

The sharpest procurement test

Give each vendor the same blind cheque set, the same workflow-control matrix, and the same integration failure scenarios. Score the actual outputs, not the slideware.

Run a Pilot with Chequedb

Scanner hardware catalog Mailroom capture use case Bulk processing Cheque scanning software

API & Integration

Hardware & Deployment