MODELCATALOG

Recognize the recurring problem shape before selecting architecture or technology.

PATTERN LIBRARY

Reusable Banking AI Patterns

The pattern should stay stable even when the underlying model, orchestration framework, or data platform changes.

01

Knowledge & Research

Users repeatedly search policies, procedures, prior cases, research, or approved knowledge.

Question → Retrieve approved knowledge → Build context → LLM reasoning → Evidence-backed response

Required capabilities

  • Enterprise search / RAG
  • Entitlement-aware retrieval
  • Prompt/context assembly
  • Citation and evidence capture
  • Evaluation for groundedness

Bank examples

  • Policy Q&A
  • Regulatory research
  • Credit research
  • Audit research
  • Negative-news research

Control considerations

  • Approved-source restriction
  • Entitlement enforcement
  • Source citations
  • Unsupported-answer handling
02

Data Investigation

An analyst asks why a metric changed, where a value came from, or what is driving an exception.

Question → Interpret metric → Retrieve metadata → Governed query → Analyze → Explain drivers → Evidence

Required capabilities

  • Semantic / metadata layer
  • Governed SQL or data-query service
  • Data-quality checks
  • Reasoning and variance analysis
  • Evidence package / lineage

Bank examples

  • Regulatory variance analysis
  • Risk reporting
  • Management reporting
  • Data-quality investigation
  • Finance commentary

Control considerations

  • Read-only query execution
  • Approved datasets
  • Query logging
  • Reconciliation to source
  • Human review for material reporting
03

Document Intelligence

The process depends on extracting, validating, and interpreting information from documents.

Document → Classify → Extract → Validate → Policy reasoning → Exception / decision

Required capabilities

  • OCR / IDP
  • Document classification
  • Structured extraction
  • Policy/rules validation
  • Exception workflow

Bank examples

  • Mortgage document review
  • Commercial credit packages
  • KYC documentation
  • Dispute evidence
  • Regulatory submissions

Control considerations

  • Field-level validation
  • Document provenance
  • Confidence thresholds
  • Exception routing
  • Human verification for low-confidence extraction
04

Investigation / Case Management

An alert or case requires gathering evidence from several systems before a recommendation is made.

Alert → Gather case context → Research → Analyze → Recommendation → Human review → Case update

Required capabilities

  • Case-context assembly
  • Multi-system data access
  • Tool calling
  • Research / RAG
  • Evidence synthesis
  • Workflow orchestration
  • Human-in-the-loop

Bank examples

  • AML alert investigation
  • Fraud investigation
  • KYC investigation
  • Operational-risk cases
  • Account takeover investigation

Control considerations

  • Case-level audit trail
  • Evidence-linked recommendations
  • Role-based actions
  • Mandatory review for material decisions
05

Controlled Content Generation

A recurring output must be drafted from approved data and knowledge in a controlled format.

Approved data + approved knowledge + template → LLM → Validate → Evidence → Approval

Required capabilities

  • Template management
  • Grounded generation
  • Structured-data injection
  • Validation rules
  • Approval workflow

Bank examples

  • Regulatory narratives
  • Credit memos
  • Management commentary
  • Customer communications
  • Case summaries

Control considerations

  • Approved templates
  • Source traceability
  • Content validation
  • Human approval
  • Version retention
06

Agentic Workflow

The objective requires multiple steps, dynamic tool choice, iterative reasoning, and controlled execution.

Goal → Plan → Select tool → Execute → Observe → Reason → Continue / stop → Human approval

Required capabilities

  • Agent / tool orchestration
  • State management
  • Enterprise APIs
  • Deterministic workflow execution
  • Policy and guardrail layer
  • Human checkpoints

Bank examples

  • Complex investigations
  • Research workflows
  • Servicing exceptions
  • Operations exception management
  • Multi-system analyst workflows

Control considerations

  • Allowed-tool registry
  • Action permissions
  • Execution limits
  • Full trace logging
  • Escalation / kill switch