Casebook: Applying the Architecture Beyond KYC
The capstone uses an auditable case system because regulated review makes the control problems obvious. The architecture is broader than KYC. This casebook shows how the same seven pillars transfer to other expensive workflows.
Use each case as a design exercise:
- what behavior must be evaluated?
- what state must be typed?
- where does human control sit?
- what must be observed?
- what can go wrong securely?
- what does the workflow cost?
- what trust artifact would help adoption?
Case 1: Agentic Revenue Operations
Production Pressure
A revenue team wants AI to research accounts, draft outreach, update CRM fields, and suggest next actions. The expensive failure is not only a bad email. It is silent CRM corruption, embarrassing external communication, duplicate outreach, or an agent spending money on low-value leads.
System Boundary
account signal intake
-> enrichment
-> lead scoring
-> draft recommendation
-> human approval
-> CRM update
-> outreach send
-> outcome tracking
Seven-Pillar Design
| Pillar | Design move |
|---|---|
| Evaluation | golden accounts with expected qualification, disqualification, and escalation outcomes |
| Typed workflow | ProspectStatus, OutreachDraft, ApprovedMessage, CrmMutationRequest |
| Human control | AI drafts and recommends; human approves external sends and high-impact CRM changes |
| Observability | trace account source, model route, draft version, approval, send result, reply outcome |
| Security | CRM write tools are scoped by account, field, and approval state |
| Economics | cheap enrichment first, frontier model only for high-value accounts or ambiguous strategy |
| Distribution | trust artifact: “how the system prevents spam and CRM corruption” |
Hard Rule
The model may draft an email. It may not send a first-touch enterprise email without approval.
Case 2: Civic Evidence Engine
Production Pressure
A civic organization wants to collect public evidence, summarize claims, identify contradictions, and publish explainers. The expensive failure is publishing unsupported claims, mixing opinion with evidence, or losing source provenance.
System Boundary
source intake
-> provenance capture
-> claim extraction
-> evidence clustering
-> contradiction review
-> editor approval
-> public publication
-> correction loop
Seven-Pillar Design
| Pillar | Design move |
|---|---|
| Evaluation | fixtures for unsupported claims, quote fidelity, source-date handling, and contradiction detection |
| Typed workflow | SourceId, ClaimId, EvidenceCluster, EditorDecision, CorrectionRequest |
| Human control | AI prepares claim maps; editors approve public language |
| Observability | record source URL, fetch time, extraction prompt, claim cluster, editor decision |
| Security | untrusted web content cannot become system instruction or publication authority |
| Economics | batch low-priority source clustering; reserve frontier models for contested summaries |
| Distribution | trust artifact: public methodology page with source and correction policy |
Hard Rule
The model may suggest a claim summary. It may not publish a public accusation without editor approval and source traceability.
Case 3: Realtime Translation Quality System
Production Pressure
A conference or live event needs realtime translation. The expensive failure is not only mistranslation. It is latency that makes the stream useless, repeated segments, missing numbers, political phrase distortion, or no way to evaluate style changes.
System Boundary
audio stream
-> transcription
-> segment stabilization
-> translation draft
-> optional refinement
-> listener delivery
-> post-session evaluation
Seven-Pillar Design
| Pillar | Design move |
|---|---|
| Evaluation | corpus with source transcript, reference translation, required terms, number preservation, and latency targets |
| Typed workflow | SessionId, SegmentId, DraftTranslation, CommittedTranslation, Revision |
| Human control | speaker/admin controls session style and glossary; post-session reviewers correct gold data |
| Observability | trace first-token latency, segment commit latency, duplicate segments, provider reconnects |
| Security | listener access is public only when intended; provider credentials stay server-side |
| Economics | realtime path uses bounded models; expensive refinement can be async after the live event |
| Distribution | trust artifact: benchmark report by language pair and event style |
Hard Rule
The system may revise a draft segment. It must not silently rewrite a committed transcript without preserving revision history.
Case 4: Developer Agent for Repository Maintenance
Production Pressure
A developer agent can inspect code, edit files, run tests, and propose fixes. The expensive failure is a destructive command, secret exposure, unreviewed production change, or a patch that passes tests while violating architecture.
System Boundary
issue or task
-> repository inspection
-> plan
-> bounded file edits
-> tests
-> review summary
-> human merge
Seven-Pillar Design
| Pillar | Design move |
|---|---|
| Evaluation | regression tasks with expected diffs, tests, and forbidden destructive behavior |
| Typed workflow | TaskId, ReadOnlyInspection, PatchProposal, ValidatedPatch, HumanMerge |
| Human control | agent may propose and validate; human owns merge and production deploy unless policy says otherwise |
| Observability | record commands, files touched, tests run, failures, and rationale |
| Security | shell tools are permissioned; secrets and destructive commands are blocked or approval-gated |
| Economics | local static checks before expensive model passes; use smaller models for search and summarization |
| Distribution | trust artifact: transparent run log and patch rationale |
Hard Rule
The agent may edit a working tree under policy. It must not silently destroy user changes or deploy production without an explicit release gate.
Transfer Pattern
Across domains, the same architecture repeats:
untrusted input
-> scoped evidence
-> typed workflow
-> model as assistant
-> validation and eval
-> human-owned sensitive transition
-> semantic observability
-> audit or trust artifact
When a new AI product idea appears, do not start with the prompt. Start by filling this pattern.