Skip to main content

21. Use Cases and Scenarios

This section presents practical examples that demonstrate how the architecture operates in real-world situations. Each scenario highlights the involved capabilities, execution paths, governance controls, and measurable outcomes.

21.1 Document Generation: User Flow Example​

Scenario: A caseworker generates a templated decision letter for a benefits claim.

Flow:

  1. User selects a case in the DXP work queue and chooses "Generate Letter".
  2. DXP issues a Generate Document capability invocation with structured inputs (case data, template id, locale, compliance flags) and execution context (latency class, cost-sensitivity, user identity).
  3. Capability Gateway validates the contract, applies policy (template permissions, data classification), and selects an executor.
    • If deterministic generator is available and template covers all fields, deterministic executor is chosen.
    • If gaps or freeform content required, AI-assisted executor is permitted under policy and run with strict input filtering.
  4. Executor returns generated document with rendering metadata, confidence and provenance for any AI-generated sections.
  5. DXP renders preview; user may accept, edit (human-in-the-loop), or request regeneration.
  6. Finalised document and audit record (inputs, template version, execution mode, executor id, timestamp) are stored in the document store and linked to the case.

Outcomes & Controls:

  • Deterministic-first reduces cost and preserves predictable latency for common templates.
  • Confidence thresholds and review gates control when human review is required.
  • Full provenance enables audit and post-hoc review of AI contributions.

21.2 Progressive AI Adoption Pilot​

Scenario: Run a pilot to introduce AI-assisted entity resolution for incoming scanned forms.

Pilot plan:

  • MVP: AI-assisted Extract Metadata capability for a limited document set (3 document types) with human-in-the-loop verification.
  • Success criteria: accuracy above baseline deterministic heuristics, acceptable time-to-review, and controlled cost under quota.
  • Scope: routed through MCP with policy limiting AI to non-sensitive fields and logging all invocations.

Pilot steps:

  1. Implement capability contract and initial AI executor with strict schema for outputs and confidence signals.
  2. Route pilot traffic from a single team within the DXP, exposing human review UI to accept/reject extractions.
  3. Monitor accuracy, confidence distribution, cost per invocation, and reviewer throughput.
  4. Iterate prompts, add deterministic post-processing rules for common patterns, and progressively increase document types.
  5. On meeting criteria, promote capability maturity from AI-Assisted to Optimised by adding deterministic fallback and caching for repeated patterns.

Governance:

  • Quotas and stricter rate limits for pilot executors
  • Mandatory provenance and review logging
  • Model substitution plan with rollback capability

21.3 Performance and Cost Impact Scenario​

Scenario: Estimating cost and latency trade-offs when adding AI-assisted summarisation to a high-volume incoming mail pipeline.

Analysis:

  • Baseline: deterministic parsing + rule-based summarisation (cost X, latency Y)
  • AI-assisted: model summarisation reduces reviewer time but increases per-invocation cost (cost X + delta).

Mitigations:

  • Apply sampling: only a percentage of documents use AI summaries initially, based on confidence triggers.
  • Use hybrid path: deterministic extractor first, then AI summariser only for items that fail heuristics or exceed complexity thresholds.
  • Cache summaries for repeat or similar documents to amortise cost.

Metrics to track:

  • Cost per document (deterministic vs AI path)
  • Reviewer time saved per document
  • Latency distribution and tail latencies for synchronous user flows
  • Cache hit rate and cost amortisation

Decision rule example:

  • Promote AI summarisation for a document class when reviewer time saved * reviewer cost > incremental AI cost, adjusted by confidence-weighted error rate.