Multi-Agent Orchestration in Toronto, Canada

Banking / Financial Services

A Big Five Canadian bank's (top-5 by assets) Financial Intelligence Unit (FIU) processed an average of 12,000 transaction monitoring alerts monthly across retail, commercial, and wealth management business lines. The unit: 45 Level 1 analysts (initial alert review and triage), 18 Level 2 investigators (detailed investigation and STR preparation), 5 senior investigators (complex case review and regulatory liaison), and 2 FINTRAC reporting officers. The problems: (1) Alert volume overwhelming capacity. The bank's rule-based transaction monitoring system (Actimize) generated 12,000+ alerts monthly with a 96.5% false positive rate. Level 1 analysts reviewed each alert to determine if investigation was warranted — averaging 18 minutes per alert review. At 12,000 alerts: 3,600 person-hours monthly for Level 1 triage alone. The team consistently fell behind, carrying a backlog of 2,000-3,000 un-reviewed alerts — creating regulatory risk. (2) Investigation quality was inconsistent. Level 2 investigation quality varied significantly by analyst experience and tenure. Junior investigators frequently missed relevant transaction patterns, producing incomplete STRs that required multiple revision cycles. Average investigation time: 4.5 hours for standard cases, 12-18 hours for complex cases. STR revision rate: 34% (STRs returned by senior reviewers for additional work). (3) FINTRAC filing deadlines at risk. FINTRAC requires STR filing within 30 days of the transaction monitoring alert. The backlog and revision cycles meant the bank was filing 15-20% of STRs within 5 days of the deadline — creating regulatory risk if any additional delays occurred. (4) Bilingual requirements. The bank operates in both English and French. Quebec-based customers communicated in French, and FINTRAC accepts filings in both languages. The FIU's French-language capacity was limited — only 8 of 45 Level 1 analysts and 4 of 18 Level 2 investigators were bilingual, creating bottlenecks for Quebec-originated alerts.

We built a 5-agent AML investigation orchestration system over 14 weeks. (1) Alert Triage Agent: processing each transaction monitoring alert through an AI assessment pipeline. The agent analysing: transaction details (amount, type, counterparty, geography, timing), customer profile (account age, typical activity patterns, KYC risk rating, occupation, source of wealth), alert context (which monitoring rule triggered, alert history for this customer, related alerts for connected accounts), and open-source intelligence (adverse media screening, corporate registry checks for commercial accounts). The agent producing a structured triage assessment: recommended disposition (close — false positive with documented rationale, investigate — with preliminary typology classification, escalate — indicators of serious criminal activity requiring immediate senior review), confidence score (0-100), and supporting evidence summary. Using Claude Sonnet 4.6 — strong on structured analysis and evidence-based reasoning with lower hallucination risk (critical for regulatory applications). Alert processing: 3-4 minutes per alert (vs. 18 minutes manual). False positive identification accuracy: 94.2% (the agent correctly identifying false positives that analysts would also close, while flagging genuine suspicious activity for investigation). (2) Investigation Agent: for alerts recommended for investigation, conducting automated evidence gathering and analysis. The agent: pulling 12-month transaction history for the subject and related accounts, identifying transaction patterns matching FINTRAC typologies (structuring — transactions just below $10K CAD reporting threshold, rapid movement of funds, layering through multiple accounts, third-party payments inconsistent with business profile), mapping account relationships (identifying connected accounts through shared addresses, phone numbers, beneficial ownership, or transaction patterns), checking against Canadian sanctions lists (SEMA, Criminal Code listed entities, UN Security Council lists), and generating an investigation narrative with evidence citations. Processing: 25-40 minutes per investigation (vs. 4.5 hours manual for standard cases). (3) STR Generation Agent: for cases warranting FINTRAC reporting, drafting the Suspicious Transaction Report in compliant format. The agent populating: Part A (reporting entity information), Part B (transaction information — structured data from the investigation), Part C (subject information — KYC data), and the narrative section (describing the suspicious activity, the indicators identified, and the rationale for reporting — in either English or French based on the customer's language of record). Generating reports in both English and French with equal quality — addressing the bilingual capacity constraint. (4) Quality Review Agent: reviewing every triage decision and STR before human review. The agent checking: completeness (all required investigation steps performed, all FINTRAC-required fields populated), consistency (triage decision consistent with the evidence, typology classification matching the observed patterns), regulatory compliance (STR format meeting FINTRAC specifications, filing within the 30-day window), and risk flags (identifying cases that may require law enforcement referral or OSFI notification beyond standard STR filing). (5) Bilingual Processing Agent: handling French-language communications, documents, and customer interactions throughout the pipeline. The agent: translating French-language customer documents and communications for anglophone investigators, generating French-language investigation notes and STR narratives for Quebec-originated cases, and ensuring consistent regulatory terminology in both languages (FINTRAC's bilingual reporting vocabulary). The orchestrator managing the pipeline with human gates at two critical points: after triage (Level 1 analyst confirming the AI disposition for a sample of cases — initially 100%, reducing to 20% as confidence builds), and after STR generation (Level 2 investigator reviewing and approving every STR before FINTRAC submission — this gate remains at 100% regardless of AI confidence, per regulatory expectations).

Alert processing: Level 1 triage time reduced from 18 minutes to 4 minutes per alert (AI triage + analyst review of AI assessment). The backlog of un-reviewed alerts eliminated within 3 weeks of deployment — from 2,500 pending alerts to zero. Ongoing: alerts triaged within 24 hours of generation (vs. 3-7 day delays previously). False positive reduction: the AI triage agent's false positive identification (94.2% accuracy) enabled the team to focus investigation resources on the 420 genuinely suspicious cases monthly (vs. reviewing 12,000 alerts to find them). Investigation quality: STR revision rate dropped from 34% to 8% — the AI investigation agent producing more complete, consistent evidence packages that required fewer revision cycles. Average investigation time: reduced from 4.5 hours to 1.2 hours (AI investigation + investigator review and approval). Bilingual capacity: French-language cases processed at the same speed and quality as English-language cases — eliminating the bilingual bottleneck. Quebec-originated alert backlog (previously 40% larger than English-language backlog due to capacity constraints) eliminated. FINTRAC compliance: zero STRs filed within 5 days of the 30-day deadline (vs. 15-20% previously). Average filing: 12 days after alert generation (vs. 24 days previously). Regulatory examination: OSFI's subsequent AML examination noting significant improvement in the bank's transaction monitoring effectiveness and investigation quality. Staff redeployment: 15 Level 1 analysts redeployed to enhanced due diligence, customer outreach, and proactive monitoring programmes — the FIU shifting from reactive alert processing to proactive financial crime prevention. Annual value: CAD $2.8M in analyst productivity, CAD $1.2M in reduced regulatory risk (fine avoidance, examination remediation), and immeasurable reputational risk reduction. Development cost: CAD $340,000. Payback: under 8 weeks.