Client
Granby Finance
Sector
Financial services
Headline
96% · Documents processed without review
Time to ship
Nine weeks
The problem
Granby Finance is a mid-sized financial services firm in Liverpool running a high-volume document-intensive operation — loan applications, KYC packs, supplier statements, internal compliance documents. The operations team was reading every document by hand, extracting structured fields, validating against business rules and pushing them into the operations system. Volume was growing 15% year-on-year; the team had been at full capacity for three quarters and overtime spend was meaningful. Compliance review was a separate, slower pipeline that ran on top of an already-slow operations pipeline.
What we built
A document intelligence pipeline coupled with a multi-step reconciliation agent. The document side: extraction and validation pipeline that reads the inbound document, extracts the structured fields, validates against Granby's business rules, and routes low-confidence cases to a human reviewer with the model's reasoning visible.
The agent side: an automated reconciliation workflow that takes the extracted fields, pulls the corresponding data from the operations system and the customer master, identifies any exceptions, drafts the recommended action and escalates anything outside its decision boundary to a human compliance reviewer. The agent does the cross-system data pull and the candidate matching; the human makes the decision.
Both systems share an audit log: every model call, every retrieved record, every human decision is timestamped with versioned model identifiers. The compliance team has a single dashboard view of every model-touched case.
What changed
- 96% of documents processed end to end without human review against a confidence threshold tuned by the compliance team. The remaining 4% route to a reviewer with full context — the headline number Granby committed to before we started.
- Reconciliation cycle time down from 3 days to under 6 hours for the most common case types.
- Exception identification up: the agent surfaced 14 cases in the first quarter that the manual process had been missing — small in count, meaningful in value.
- Overtime spend in operations down 31% in the first quarter post-launch.
- Compliance team turnaround on flagged cases down from days to hours because the case file arrives pre-structured with the model's evidence.
What we did not do
We did not have the AI system make any final compliance decisions. The agent prepares the case file; a human compliance officer signs off. The regulatory expectation for human oversight of automated decision-making in financial services is clear, and we built to it.
We also did not deploy any model that materially influences credit or suitability decisions about individual customers. Those projects sit in a different category and need a regulated path to deployment; this engagement was deliberately scoped to operations and document workflows.
Compliance and audit
Models hosted in UK regions; customer and transaction data does not leave UK infrastructure. Full audit log per case — inputs, model versions, retrieved records, reviewer decision, timestamp. The Granby compliance team can reconstruct any case end to end at any point in the retention window.
What is next
Granby's operations lead is now the named owner. We are on retainer for the next two quarters tuning the system as new document types and product lines are added. A larger second phase — extending the agentic workflow into the dispute and complaint handling pipeline — is scoped for the next quarter.
Related reading
For the broader pattern: AI for financial services in Liverpool: practical use cases for 2026.