Client Q&A — Finastra
Internal use only. Q&A prepared during Finastra prospect engagement. Contains technical and competitive context for internal reference. For the partner-facing competitive comparison, see the OCR section in Battlecard: Smartflow vs. Manual Process.
Accuracy & Technology
Q: What is the accuracy of the OCR and how long does it take to train?
A: Traditional OCR solutions (not powered by LLMs) have historically ranged from 70–90% accuracy depending on the type of extraction (clean text vs. handwritten images) and document complexity. These solutions come pre-trained and often cannot be further trained with a client's own internal data, as they are centrally managed by the software provider (e.g. Amazon Textract, ABBYY FineReader).
LLM-based solutions provide a significant increase in accuracy and a stronger platform for additional analytical capabilities. In a recent production deployment for a large HK securities house, we processed more than 25,000 data points — including handwritten images and non-English languages — and achieved a 99.5% accuracy rate with our Human-in-the-Loop capability.
Q: What is the % accuracy of document scanning — LLM-based vs traditional OCR?
A: For traditional OCR, banks typically see 80–90% field-level accuracy on semi-structured documents — often sufficient to justify automation of basic document capture.
Modern AI/LLM-powered document understanding solutions can reach 90–99% accuracy on structured and semi-structured financial documents when properly trained and configured, especially for key data fields in loan packages. The improvement comes from better context handling and layout understanding, not from ledger technology. Actual performance still depends on document quality, template variation, and validation rules.
Competitive Positioning
Q: Who are the competitors on the market and how does Smartflow's technology differentiate?
A: Our primary competition is traditional OCR technology — Amazon Textract, ABBYY FineReader, or generic extraction-focused engines that may use LLMs.
"Generic LLM tools extract text. Smartflow produces trusted, evidence-linked, audit-grade loan data that flows directly into your systems of record — without ever leaving your environment."
| Generic LLM Extraction | Smartflow | |
|---|---|---|
| Domain Model | General-purpose schema inference | Pre-trained on APLMA/LMA/LSTA templates; understands loan-specific ontology (facility, tranche, covenant, agent) |
| Output Target | JSON/text blob | LoanIQ-ready field map with bank-specific customisation per deployment |
| Accuracy Floor | 80–85% on complex legal docs | 90%+ on credit agreements; 99%+ with HITL gate |
| Validation Logic | None / user-defined | Business rule engine with LoanIQ format constraints (char limits, date formats, field cardinality) |
| Evidence Chain | Text extraction only | Page/section/snippet provenance per field — regulatory-grade audit trail |
| Covenant Intelligence | Extract text | Extract + calculate ratios + version-aware threshold application + breach prediction |
| Deployment Model | Cloud (data leaves perimeter) | Edge-deployed; data sovereignty by design |
| System Integration | Webhook/API to anything | Native LoanIQ push, with support for custom connectors |
| Workflow Layer | None | HITL review queues, approval gates, SLA tracking, exception management |
| Failure Mode Handling | Hallucination undetected | Confidence gating — low-confidence extractions blocked from production state until reviewed |
Pricing
Q: What is the approximate price of an LLM-based solution vs. traditional OCR?
A: A traditional enterprise OCR solution for banks typically ranges from ~$50K to $300K+ per year, depending on document volume, licensing model, and customisation. LLM-based document intelligence solutions for loans start at $100K and can scale up to $300K per year depending on usage.
The higher starting cost is generally justified by improved accuracy, better handling of unstructured and complex loan documents, reduced manual review effort, and higher end-to-end automation rates. Actual pricing varies based on vendor, deployment model (cloud vs. on-prem), compliance requirements, and transaction scale.
DLT vs. LLM Clarification
Q: What are the benefits of DLT-based OCR vs. traditional OCR in the context of LoanIQ?
A: It is worth clarifying what "DLT-based OCR" means in this context. In document processing, DLT typically refers to Deep Learning Technology (AI-based text recognition). However, references to tamper-proof documents and audit trails align more with Distributed Ledger Technology (blockchain) — which is a separate concept from OCR.
OCR, whether traditional or AI-based, does not inherently prevent document tampering. Immutability and auditability are features of ledger or system architecture, not text recognition technology.
For Smartflow's purposes: the answers above address AI/LLM-based OCR improvements, not blockchain-based document integrity. Our audit trail is achieved through field-level provenance and session logging — not DLT.