UK-Based B2B Payments Platform
$2.1M/year in compliance savings by automating KYC/AML document processing.
The client runs a B2B payments platform used by about 3,200 small and mid-sized businesses across the UK and EU for cross-border transactions. As an FCA-regulated company, they have to perform KYC and AML checks on every business that onboards, plus periodic re-verification on existing accounts.
At engagement start, the compliance team was 14 analysts processing about 1,800 KYC applications per month. Each application meant manual review of identity documents, proof of address, corporate registration filings, beneficial ownership declarations, and screening against sanctions lists and PEP databases. Average processing time: 47 minutes per application. About 30% needed at least one round of follow-up because documents were incomplete, expired, or didn't match the onboarding information.
The backlog was persistent. During peak onboarding periods — Q1 and Q4 especially — the queue regularly hit 600+ pending applications, pushing average onboarding time past 11 business days. The sales team estimated losing 15-20 prospects per quarter who abandoned the process before approval because of delays. One enterprise deal worth roughly $340,000 in annual transaction volume was lost because the compliance review took 16 days.
Two prior automation attempts had stalled. A rules-based document classifier built by the internal engineering team hit 61% accuracy on document type identification and was abandoned after four months. An evaluation of a third-party KYC-as-a-service provider revealed integration problems with the existing transaction monitoring system and per-check pricing that would have raised costs at their volume.
14 weeks, starting with the document processing bottleneck and expanding to risk scoring and workflow automation. Each phase went into production against live applications before the next one started.
Diagnose: compliance process audit (weeks 1-2)
We spent two weeks embedded with the compliance team, watching 120 applications move from submission through approval or rejection. The goal was to find where analyst time was going. The answer wasn't surprising, but it was specific: 58% of total analyst time went to three activities that didn't require judgment — checking document type and completeness, pulling entity names and addresses from corporate filings, and cross-referencing those details against the application form. The remaining 42% was risk assessment decisions that needed human judgment. That split defined the automation boundary.
Design and deploy: document AI pipeline (weeks 3-7)
We built a document processing pipeline using OCR, document classification, and entity extraction models, trained on 4,200 historical applications across 23 document types from UK, EU, and offshore jurisdictions. The pipeline classified incoming documents by type, extracted key fields (entity name, registration number, address, director names, dates), and ran automated completeness checks against requirements for each business category. Documents that passed went straight to the risk assessment queue with pre-populated data. Documents that failed were flagged with specific reasons and sent back to the applicant with an automated list of exactly what was missing.
Deploy: risk scoring model (weeks 7-10)
We built a risk scoring model on the client's historical decision data — 14,000 completed applications across three years. The model scored based on jurisdiction of incorporation, business type, transaction volume profile, beneficial ownership structure, and screening results. Applications below a calibrated threshold went to expedited review with a recommended approval. Applications above it were flagged for enhanced due diligence with specific risk factors highlighted. The model wasn't designed to replace analyst judgment on risk decisions. It was designed to prioritize the queue and put the relevant information in front of an analyst so they could decide faster.
Scale: workflow integration and analyst tooling (weeks 10-14)
The document pipeline and risk scoring model were wired into the existing compliance workflow through a custom review interface. Analysts saw each application with documents already classified, key fields extracted and validated, screening results summarized, and a risk score with the factors behind it. The interface was built with the compliance lead and three senior analysts, iterated across four feedback rounds before production deployment. The interface did the mechanical work. The analyst made the decision. Average review time for a standard-risk application dropped from 47 minutes to 8 minutes.
$2.1M
Annual compliance cost savings
Reduced processing time and avoided headcount growth
92%
Manual review time reduction
47 minutes to 8 minutes per application
2.4 days
Average onboarding time
Down from 11 business days
94.3%
Document classification accuracy
Across 23 document types and multiple jurisdictions
73%
Fewer incomplete submissions
Pre-validation catches issues before analyst review
0
Compliance findings in subsequent FCA audit
Full audit trail maintained for every automated and manual decision
The compliance team went from 14 analysts doing mostly document processing to 9 focused on actual risk assessment and complex cases. Five were moved to a new transaction monitoring function the company hadn't been able to staff. The compliance lead said analyst retention improved once the repetitive document-checking work — the biggest turnover driver — was handled by the system. The sales team reported that faster onboarding was winning them enterprise deals, and the company closed two accounts the next quarter that cited onboarding speed in their vendor evaluation.
“We were hiring analysts just to keep the queue from growing. Now the queue barely exists, and the analysts we have are doing work that requires their expertise. The FCA audit was the real proof point — every decision had a clean trail, automated or not.”
Head of Compliance, B2B Payments Platform
Compliance team buried in document review?
If your KYC or AML process is bottlenecked by manual work that doesn't require human judgment, we can fix that.