A fast-growing payments company was drowning in compliance paperwork. We built an intelligent NLP pipeline that reduced manual processing by 40% and cut turnaround from 48 hours to 6.
Our client, an FCA-authorised payments company processing over 200,000 transactions per month, had reached an inflection point. Their compliance operations team—responsible for onboarding verification, ongoing monitoring, and regulatory reporting—was spending roughly 60% of their time manually reviewing KYC documents, sanctions screening results, and transaction monitoring alerts.
The average turnaround time for a new customer compliance check had ballooned to 48 hours. With onboarding volumes growing 30% quarter-over-quarter, the team simply could not keep pace. Manual classification errors were creeping in: misfiled Proof of Address documents, incorrectly categorised risk levels, and delayed Suspicious Activity Reports.
Compounding the issue was increasing regulatory scrutiny. The FCA had recently issued updated guidance on anti-money laundering controls, and the firm needed to demonstrate not only compliance but continuous improvement in their processes. Hiring additional analysts was a short-term fix at best—what they needed was a fundamentally different approach to document processing.
The operations leadership came to Insightrix with a clear brief: reduce manual processing time without compromising accuracy, maintain full auditability for regulatory inspections, and deliver a solution that could scale with their growth trajectory.
“We were hiring analysts faster than we could train them, and still falling behind. Every day that a compliance check sat in the queue was a day we risked losing a customer—or worse, missing something we shouldn’t have.”
Head of Operations FCA-Authorised Payments Company, London
We embedded with the compliance team for two weeks, mapping every document workflow end-to-end. This included cataloguing 14 distinct document types across KYC, KYB, and transaction monitoring processes. We audited data sources, identified classification bottlenecks, and documented every compliance touchpoint from initial customer submission through to final sign-off. The audit revealed that 72% of documents followed predictable patterns suitable for automated classification.
Using the audit findings, we designed a multi-stage NLP pipeline built on fine-tuned transformer models. The first stage handled document classification—determining whether an uploaded file was a passport, utility bill, bank statement, corporate filing, or one of ten other document types. The second stage performed entity extraction: pulling names, addresses, dates, company registration numbers, and risk-relevant data points with structured output. All models were trained on anonymised historical data provided by the client, ensuring domain-specific accuracy from day one.
The routing engine was the critical layer that made the system practical for a regulated environment. Every document processed by the NLP pipeline received a confidence score. Documents scoring above the 95% threshold were auto-categorised, validated against existing customer records, and moved to the approved queue. Documents falling below this threshold were flagged for human review with pre-populated extraction results, reducing analyst review time even on edge cases. The system maintained a complete audit trail for every decision, satisfying FCA record-keeping requirements.
We connected the pipeline to the client’s existing compliance management platform via RESTful APIs, ensuring zero disruption to current analyst workflows. The entire system was deployed on AWS infrastructure within the UK region (eu-west-2) to satisfy data residency requirements. We implemented rolling deployments with automatic rollback, comprehensive logging, and real-time monitoring dashboards so the operations team had full visibility into pipeline performance from day one.
Compliance analysts reclaimed nearly half their working hours, redirecting effort toward complex investigations and strategic regulatory projects instead of routine document classification.
End-to-end turnaround dropped from 48 hours to just 6 hours on average, with straightforward cases completing in under 90 minutes. Customer onboarding velocity increased accordingly.
The fine-tuned transformer model achieved 99.2% accuracy on document classification across all 14 document types, exceeding the manual accuracy baseline of 96.8% set by experienced analysts.
From initial discovery session to live production deployment, the entire engagement was completed in four weeks, with zero downtime to the existing compliance platform during integration.
Embedded with the compliance team. Mapped 14 document types, catalogued data sources, identified automation candidates, and benchmarked existing accuracy and throughput metrics.
Fine-tuned transformer models on anonymised historical documents. Built the NLP pipeline for classification and entity extraction. Iteratively validated against compliance team review standards.
Connected the pipeline to the existing compliance platform via APIs. Ran parallel processing alongside the manual workflow for validation. Tuned confidence thresholds based on analyst feedback.
Deployed to AWS UK infrastructure with monitoring and alerting. Delivered documentation, runbooks, and conducted team training sessions. Established a four-week post-launch support window.
Book a free AI readiness call. We'll discuss your challenges and outline a clear path forward.