Fraud Detection¶
Documentation status
Last verified: 26 Sep 2026. Readiness uses the single rubric and status vocabulary.
Scores customer transaction behaviour and delivers investigator-facing AI reports to the banking dashboard.
Purpose¶
Give risk and compliance teams continuous visibility into unusual customer behaviour — unusual times, locations/devices/merchants, and amount or velocity patterns — with a structured report they can act on.
Users¶
| Audience | How they use it |
|---|---|
| Risk / fraud operations | Review investigator reports and decide next steps in the banking application |
| Partner / client engineers | Call scoring APIs and accept dashboard report deliveries |
| Compliance / product | Set expectations for decision policy (currently provisional) |
Business problem¶
Manual review cannot keep pace with every transaction. Banks need behavioural context (what is normal for this customer), automated scoring when activity looks unusual, and a consistent report handoff into the operations workspace — without letting a language model invent approve / pend / block decisions.
Capabilities¶
| Capability | What it does |
|---|---|
| Behavioural profiling (baselines) | Builds time, location/device/merchant, and amount/velocity profiles from historical activity |
| Rule evaluation | Scores a transaction against configured rules (including a dedicated case-oriented rule path) |
| ML scoring | Scores a transaction with the ML prediction service |
| Report generation | Validates the prediction contract, applies decision policy, attaches an AI narrative, and delivers a structured report |
| Operational handoff | Posts the report to the banking application’s report endpoint for investigator review |
Invariant: the AI narrative may explain a decision; policy code produces approve / pend / block. See Fraud output contract.
Provisional decision policy
Automated primary decisions and escalations use fraud-decision-v1, which is provisional / non-production until compliance signs off. Do not treat policy output as a filed regulatory decision.
End-to-end workflow¶
flowchart LR
batch[Start_customer_baselines]
profile[Build_or_skip_profiles]
score[Score_transaction]
report[Compose_report]
dash[Dashboard_report]
batch --> profile --> score --> report --> dash
- Baselines — schedule behavioural checks for scoped customers; persist profiles or skip on cold start (too little history).
- Scoring — call ML or rule scoring for a transaction (authenticated).
- Report compose — validate prediction fields, apply policy, attach narrative.
- Handoff — deliver the structured report to the banking application.
- Post-report work (assignment, notes, closure) — owned by the banking application where implemented; not claimed as complete in the Fraud AI products. See Case Management and Transaction monitoring lifecycle.
Inputs¶
| Input | Typical source |
|---|---|
| Customer identifiers for baseline batches | Banking application / orchestrator |
| Transaction and context for scoring | Banking application backend |
| Auth material | Bidirectional JWT + tenant header — Auth and tenancy |
Outputs¶
| Output | Consumer |
|---|---|
| Prediction (score / flags / conditions) | Immediate HTTP response to scoring caller |
| Structured fraud report (decision, escalations, evidence, narrative) | Banking application report endpoint |
| Escalation labels (for example investigator or SAR review) | Labels on the report only — not a filed regulatory submission |
Integrations¶
- Banking application / dashboard (report delivery)
- Model provider for narrative generation (server-side)
- Event infrastructure for baseline and report pipelines (platform-operated)
Partner route details: Fraud API.
Readiness¶
Rated with the single readiness rubric. Reasons for each rating are on that page.
| Gate | Rating |
|---|---|
| Code | PARTIAL |
| Integration Verified | NOT MET (ML scoring output and the report step's accepted tiers are incompatible) |
| Ops Ready | PARTIAL |
| Security Approved | NOT MET |
| GA | NOT MET |
The automated decision policy fraud-decision-v1 is provisional.
See Product readiness summary.
Limitations¶
- Cold-start customers may skip profile persistence; rule scoring is the expected path for those customers.
- A successful score does not guarantee a dashboard report (contract rejection, delivery failure, or score/tier mismatch can block compose).
- Full analyst case lifecycle is not part of the Fraud AI product boundary — see Case Management.
- Decision policy remains provisional until compliance GA.
Related documentation¶
- Transaction monitoring lifecycle
- Case Management (rule scoring vs durable cases)
- Fraud output contract
- Fraud API · Auth and tenancy
- AML and Regulatory Reporting (escalation labels vs filing)