Trust boundary
The model proposes. The application checks access and validates tool parameters. Source permissions must survive indexing, retrieval and cache reuse.
02 / REFERENCE ARCHITECTURES
Design references with explicit boundaries and failure paths. These are not TechUnity client engagements or measured results.
DESIGN REFERENCE / 01
Retrieve within the caller’s permissions, produce a sourced proposal, and require approval before an external action.
The model proposes. The application checks access and validates tool parameters. Source permissions must survive indexing, retrieval and cache reuse.
Missing evidence or stale permissions → refuse or escalate. Approval binds to a specific proposal, parameters and expiry; changed input requires another review.
Deliverable example: data-flow map, permission matrix, test set for denied access and unsupported answers, and approval audit records.
DESIGN REFERENCE / 02
Place a narrow, observable adapter between the workflow and the existing system of record.
The target system owns the authoritative state. Restrict credentials to required operations and define timeouts, concurrency limits and retryable errors.
Unknown write result → reconcile by operation ID before retry. A software rollback does not undo an external write; use a documented compensation or manual recovery path.
Deliverable example: API contract, idempotency test, retry policy, reconciliation procedure and credential revocation checklist.
DESIGN REFERENCE / 03
Treat evaluation, operational visibility and human intervention as explicit release conditions.
Track model, prompt, tool and data versions. Define success, refusal and escalation separately, alongside latency and cost budgets.
Quality or cost threshold breached → disable affected actions and notify the owner. Resume only after reviewed evidence; rollback and compensation are separate procedures.
Deliverable example: evaluation report, alert routing, stop-switch drill, recovery runbook and signed acceptance record.
PUBLIC CASE STUDY · MORGAN STANLEY
Morgan Stanley’s June 26, 2024 announcement describes AI @ Morgan Stanley Debrief: with client consent, it generates meeting notes and action items, drafts an email for an advisor to edit and send at their discretion, and saves a note into Salesforce.
A useful workflow needs an input permission boundary, a reviewable output and a clear system destination. This is our reading of the public description, not a claim about its internal architecture.
The announcement does not establish the detailed access-control design, recovery guarantees or transferable ROI for another business. These require separate evaluation.
Third-party public case; not a TechUnity client, partnership or delivery claim.
Read Morgan Stanley’s original announcement ↗