Real-time decisioning architecture for the bank

What is real-time decisioning in a bank

Every customer interaction has a potential decision: approve transaction or block, show offer, escalate, log audit. Without a centralised layer — decisions scattered, inconsistent.

Banking decisioning consolidates: trigger event → context fetch → policy + model evaluation → decision + audit.

Structural components

Trigger ingress. Events from event bus, channels, transactions.

Context fetch. Profile from CDP, current state (balance, recent activity, active cases).

Policy engine. Declarative rules (regulatory, risk appetite, customer preferences). Change without deploy.

Model serving. ML models (PD, fraud score, propensity). Versioned, A/B-able.

Decision logic. Combination of policies + model scores + business rules.

Audit trail. Every decision with full input, models, policies, output, owner.

A/B framework. Decision compared to alternative for learning.

Latency budget

Banking specific:

• Card authorisation: <500ms end-to-end (network limit).
• Online banking transaction: <1s acceptable.
• Retail credit decision: <5s.
• Mobile app NBA: <500ms.

Budget allocation:

• Trigger ingress 20-50ms.
• Context fetch 30-80ms.
• Model serving 20-100ms.
• Policy evaluation 10-30ms.
• Decision logic 10-30ms.
• Audit (async) 10-30ms.

Where banking-specific is harder

Regulatory traceability. Every automated decision must be explainable per regulator (especially for credit and AML).

Model bias. Credit models in particular — disparate impact monitoring mandatory.

Override audit. Manual overrides — full chain of approval.

False positive cost. Blocking a legitimate transaction — customer churn.

Where it usually breaks

Latency not measured — channels timeout.

Policies in code — changes require deploy.

Models not retrained — accuracy decays unnoticed.

Audit incomplete — investigation cases impossible.

A/B testing impossible — product does not learn.

Operating model

Owner — Head of Decisioning, between risk and technology.

Teams: platform engineering, decision science, policy management, channel integration.

Routine — weekly decision review, monthly model review.

Related

• /en/architecture/banking-cdp-architecture/ — context source
• /en/architecture/banking-event-bus-architecture/ — event bus
• /en/architecture/banking-mlops-architecture/ — MLOps
• /en/solutions/banking-credit-decisioning-platform/ — credit decisioning