Banking & Financial Services
Detecting financial crime in real time without drowning analysts in false positives.
Rules-based transaction monitoring had grown for a decade. Each new typology added rules, each rule added alerts, and the investigation queue was permanently beyond capacity.
- Client
- Tier-1 retail and commercial bank
- Region
- Europe
- Scale
- Millions of daily transactions, 24/7 analyst desk
The challenge
What was actually going wrong.
- Alert volumes exceeded what the financial-crime desk could investigate, so genuine risk sat behind noise.
- Rules could not represent relationships between accounts, mules and counterparties.
- Any model change had to be defensible to regulators before it could touch a decision.
The AI solution
How it was engineered.
Graph features and streaming models with analyst feedback captured into retraining.
Graph feature layer
Counterparty relationships modelled as a graph, producing ring, velocity and structuring features that a flat rule cannot express.
Streaming scoring
Models score in-flight on the streaming platform, with rules retained as a hard floor rather than replaced.
Explainability by design
Every score ships with contributing features and a plain-language rationale in the case record.
Analyst feedback loop
Investigation outcomes flow back as labels, so precision improves from the desk's own work.
Delivery
The path to production.
Shadow mode for one quarter before decisioning authority.
Weeks 1–3
Discover
Typology review, label archaeology, model risk classification with second line.
One quarter
Prove
Full shadow mode against live traffic, side-by-side with the rules engine.
Months 5–8
Scale
Decisioning authority granted per typology, staged by risk tier.
Ongoing
Operate
Quarterly model validation, drift and fairness monitoring, audit evidence generated automatically.
The outcome
What changed, and how we know.
Material
Reduction in false positives
At equal detection, client-reported
1 qtr
Shadow mode before go-live
Regulator-visible evidence period
100%
Decisions with recorded rationale
Every score explainable
- Analyst capacity moved from clearing noise to investigating genuine networks.
- Detection held while the alert queue became workable within shift capacity.
- Model governance artefacts are produced by the pipeline, not assembled before each audit.
“Shadow mode was non-negotiable. It is why the model went live at all.”
What is next Extend graph features to trade finance and onboarding risk scoring.
More case studies
Other programmes in production.
Manufacturing
From reactive maintenance to predictive operations: reducing critical equipment downtime.
Read case studyHealthcare & Life Sciences
Compressing regulated document drafting while keeping clinicians in control.
Read case studyGovernment & Public Sector
A sovereign caseworker copilot that shortens handling time and logs every step.
Read case studyStart here
What should your enterprise be able to do next?
Tell us the outcome you want to achieve. We will help you identify, engineer and scale the right AI solution.