A tier-1 retail bank (>$2B annual revenue) needed to increase mortgage processing throughput without expanding headcount.

The bottleneck wasn’t demand — it was verification. Back-office teams were manually parsing bank statements, rekeying purchase agreements, checking land registries, and validating insurance documents. Inputs were unstructured and inconsistent: low-quality scans, dozens of statement formats, dense legal language. Every action required a traceable audit trail.

We deployed the EigenPal platform inside the bank’s cloud and automated four verification workflows end-to-end — converting a manual, document-heavy process into a structured, auditable pipeline.

Their workflows

1. Salary verification from bank statements

Mortgage specialists were reviewing customer bank statements line by line to verify recurring salary inflows, match employer names, and cross-check against declared income. Statements came from dozens of different banks with different formats.

We extract all transactions, identify probable salary payments, match employer names, and output a structured verification result directly into the bank's internal forms.

2. Purchase agreement extraction and contract pre-fill

Staff were reading house purchase agreements and manually retyping buyer/seller details, property identifiers, purchase price, addresses, and dates into the mortgage contract prep flow. Slow, and prone to rekeying errors.

We extract the relevant fields from the agreement and map them into the bank's required schema, producing a pre-filled contract package ready for review.

Where humans stay in the loop. Both workflows (1) and (2) are maker-checker processes. We automated the maker role. When the system hits edge cases or ambiguities, it flags the exact reason for human review - e.g., "employer name mismatch on page 3" or "low-confidence scan on purchase price." The reviewer doesn't re-read the whole document; they check the flagged issues.

Result: 80% process speedup, ~5 FTEs saved. Human validation preserved where it matters.

3. Land registry transfer verification

Before disbursement, the bank needs to confirm the property is legally registered with the bank as beneficiary. This means navigating a public land registry site and checking that the legal state matches expectations.

The bank already knows exactly what it expects to see - specific property, owner, pledge, bank as beneficiary. This makes it a clean candidate for full automation. We fetch the public record, extract the fields, and run the comparison. High-confidence matches go through with zero human touch.

We only route to manual review when something is missing, inconsistent, or below our confidence thresholds. When we do route, we highlight exactly where the discrepancy is.

Result: 89% straight-through automation rate. Manual review required in 11% of cases.

How it works

You can't just throw a long prompt at an LLM and hope it handles messy real-world documents reliably. We built a multi-stage pipeline designed around predictability, auditability, and continuous improvement.

Scoped LLM blocks. Instead of one monolithic prompt, we constrain each LLM call to a specific, narrow task. This makes the pipeline more robust and every decision traceable. When something goes wrong, we know exactly where.

Deterministic logic where possible. We use scripting and rule-based validation wherever the logic is well-defined. LLMs and OCR handle the unstructured parts. No point using a model where a regex works.

Eval-first development. Every change to prompts or extraction logic runs against evaluation datasets before deployment. We don't ship changes that regress accuracy.

Automated feedback capture. When an operator corrects the system on an edge case, that correction feeds back into our evaluation and training data. The system gets sharper over time as volume increases.

Because the pipeline is modular at the block level, we can swap underlying models as better ones ship - during this project we upgraded the core extraction model mid-deployment and picked up a 2-point accuracy gain with no prompt changes.

Results

Across all four workflows, the pipeline achieved 97% average accuracy on the bank's validation datasets.

80% process speedup on human-assisted workflows
~15 FTEs saved in mortgage back-office operations
89% straight-through automation on deterministic checks
97% accuracy across OCR, LLM, and deterministic pipeline stages