Stage 1 Build and Acceptance
Updated 9 September 2026
The Internal Dogfood MVP build from the original vision, pages 10–11, is complete and deployed on 8 September 2026. AI inference remains disabled at the user's request. The human-review workflow is built; model output has only been represented by clearly labeled fictional local fixtures. This is a technical pilot release, not production approval or evidence that the business targets have been met.
Build Checklist
- [x] Data use: purpose-specific decisions, evidence, reuse withdrawal, capture exclusions and reviewed retention/deletion.
- [x] Administration: platform policy, usage limits, staff access/offboarding, workspace-administrator roles and an isolated-deployment directory.
- [x] AI workflow: bounded suggestions, source evidence, permission/quota checks and explicit human review. Live inference is disabled.
- [x] Measurement: queue/workload, ownership, response and backlog age, knowledge yield/reuse, and recorded pilot observations.
- [x] Delivery: 38 application tests, scoped desktop/narrow UI review, live SSO/disabled-state checks, refreshed guides/screenshots, shareable package and committed slices.
Original Scope Coverage
| Stage 1 Requirement | Delivered Behavior | Evidence and Limits |
|---|---|---|
| One WhatsApp identity per subsidiary | Separate runtime, database, credentials and Teams binding; startup rejects conflicting database bindings. | One real workspace; two-database isolation tests. Additional subsidiaries need their own provisioned accounts. |
| Teams workspace and companion web admin | Shared Inbox plus separately authenticated Platform Administration. | Live Teams and administrator sign-in passed. |
| Ownership lifecycle | Unassigned, In Progress, Waiting for Customer, Escalated and Resolved queues. | Collision/stale-action tests; real single-user claim and follow-up evidence. |
| Assignment, notes and escalation | Claim, handover, internal notes, structured escalation, supervisor decisions and recovery. | Live same-user escalation/recovery; distinct-person workflows tested locally. |
| Verified/unverified customer profile | Evidence-backed facts with explicit verification and correction history. | Live fictional profile rehearsal and local authorization tests. |
| Relationship timeline | Human-confirmed profile links across separate conversation histories. | Cross-channel links and reversals tested locally; no automatic identity merging. |
| Searchable archive | Search captured history and open the matching case. | Archive and 200-case queue tests. |
| Intent, outcome, process and knowledge tags | Case Context and source-linked knowledge nomination. | Internal context and guarded revision tests. |
| Service reporting | Backlog, workload, response, reassignment, resolution, ownership, unresolved age and knowledge metrics. | Live queue read; local metric tests. Operating evidence starts empty. |
| Audit history and permissions | Role/tenant/workspace checks, versioned actions, audit history and dynamic offboarding. | Authorization, retry, restart and isolation tests; one live administrator. |
| Structured employee feedback | Case-linked suggestions, supervisor decisions and period reporting. | Live fictional feedback and local distinct-role checks. |
| Limited AI with human approval | Summary/category suggestions and unpublished knowledge drafts with source review. | Local fixed fixtures and injected transport responses only. No live model calls or autonomous replies. |
The original service-enquiry and escalation journeys are covered by these controls. Cross-subsidiary learning patterns and advanced process intelligence remain later stages. LINE and WeChat are additional verified technical demonstrations.
Administration Boundary
99 Venture's separate web console owns platform capabilities, policy, usage limits, workspace registration and administrator grants. Workspace administrators manage ordinary staff and pilot evidence for their own runtime; they can read policy and health. Microsoft administration supplies sign-in and Teams installation only. Support supervisor access does not grant platform administration.
The directory records isolated deployments and opens their independently authenticated consoles. It does not provision resources, combine transcripts or provide automatic remote health checks. This keeps the first multi-workspace setup explicit and avoids a shared cross-subsidiary database.
Acceptance That Requires Real Operations
- 99 Venture names the subsidiary, policy owner, knowledge owner and support manager; approves purposes, capture and retention rules; and supplies client-owned accounts.
- Separate real employees rehearse assignment, escalation, recovery, review and offboarding in isolated client workspaces.
- Establish a baseline, gather staff feedback and run enough real pilot work to assess the proposed gates below.
- Complete credential rotation, alerts, encrypted off-host backup/recovery and an Azure cutover rehearsal before production.
- Approve and validate a provider, exact model, data handling and cost ceiling before deliberately enabling AI. Keep it disabled meanwhile.
Proposed Business Gates
| Measure | Original Proposed Gate |
|---|---|
| Eligible conversations completed in Teams | At least 70% |
| Conversations with an accountable owner | At least 95% |
| Reassigned cases completed without repeating key facts | At least 80% |
| Approved knowledge per 100 resolved cases | At least 5 |
| Staff who prefer OCDIT in the survey | At least 70% |
| Median first-response and unresolved-age improvement | At least 15% |
| Material permission or audit incidents | Zero |
| Workflows reusing accepted knowledge | At least 3 |
These are proposed decision gates, not measured achievements. The one-user pilot and fictional local tests cannot establish them. See Verification, Staff Guide and Platform Administration.