Release Plan
The final doc in the set. It operationalizes the PRD’s three-phase scope into concrete entry/exit gates, rollback criteria, and a beta strategy — the release shouldn’t be a launch date, it should be a sequence of trust-earning gates.
1. Release philosophy
Ad Copilot’s biggest adoption risk isn’t feature completeness — it’s a marketer’s trust in AI-driven action on their live budget. Each phase expands scope only after the previous phase has demonstrated the trust and quality signals it needs. No phase is gated purely on a calendar date.
2. Phase 1 — MVP: Creative + Suggest-lane (Meta only)
Scope: US-1, US-2, US-3, US-5, US-6, US-7, US-8 (Suggest/Escalate only), US-9, US-10, US-12, US-13.
Entry criteria
- Meta Ads API integration complete; data-accuracy validated against manual spot-check of 10 beta-partner accounts.
- Brand-safety filter: zero false negatives on a curated known-bad term list.
- Cost-cap enforcement (US-12) load-tested to block reliably at the boundary.
Beta cohort: 5–10 design partners, mix of e-commerce (ROAS) and lead-gen (CPA), actively managing $50K+/mo so drift signals have enough data.
Exit criteria (to Phase 2)
- Suggestion acceptance ≥ 40% across beta, sustained 4+ weeks.
- Zero brand-safety filter failures reaching a live account.
- No unresolved P0/P1 bugs in Action Executor or Audit Log.
- Qualitative: partners report recommendation reasoning is useful (structured interviews, not just analytics).
Rollback / kill-switch: per-account “pause AI” toggle exists from day one. Any partner can disable generation and recommendations instantly.
3. Phase 2 — Multi-platform + budget/bid recs
Scope: Google Ads; extend to budget/bid recommendations (still Suggest-only); consider US-4, US-11, US-14 from Phase 1 learnings.
Entry criteria
- Phase 1 exits met and sustained.
- Google Ads developer token approval secured (Architecture §3.2).
- Confidence calibration reviewed against real Phase 1 outcomes before extending scoring to budget/bid actions.
Rollout: budget/bid recs behind a feature flag — Phase 1 beta cohort first.
Exit criteria (to Phase 3)
- Budget/bid acceptance reaches parity with creative-recommendation acceptance.
- Incremental ROAS north-star shows a positive, statistically-reasonable trend across the expanded cohort.
- No Escalate item unreviewed past its staleness threshold — validating the review flow scales.
4. Phase 3 — Auto-apply lane
Scope: Auto-apply goes live; rollback becomes load-bearing in production.
Entry criteria
- Accounts become Auto-apply-eligible only after meeting the trust-tier bar (data volume + acceptance history) — per-account, not global.
- Platform-level hard ceilings on Auto-apply blast radius are implemented and confirmed non-configurable by admins.
- Rollback tested under failure injection in staging before any real account exposure.
Rollout: opt-in only. Even eligible accounts must explicitly enable Auto-apply per campaign — never on by default.
Success: Auto-apply accounts show equal or better Incremental ROAS vs their own Suggest-lane history; rollback rate low and stable.
5. Feature flags
| Flag | Controls | Default |
|---|---|---|
platform.google_ads | Google Ads availability | Off until Phase 2 entry |
recs.budget_bid | Budget/bid recommendations | Off until Phase 2 |
lane.auto_apply | Whether Auto-apply can ever be reached | Off until Phase 3; then opt-in per account |
account.ai_paused | Per-account master kill-switch | Available from Phase 1; owner-overridable |
6. Post-launch monitoring
- Weekly review of guardrails: LLM cost %, acceptance rate, escalation rate.
- Any single-account cost-cap breach investigated same-week.
- Confidence calibration treated as ongoing release-health work, not a one-time launch task.
What “done” looks like: Case Study → PRD → User Stories → AI Workflow → Metrics Spec → Architecture → Release Plan all trace one thesis — earn autonomy incrementally; make the AI’s boundaries explicit — through every layer.
More in this series