Phase-Gated Rollout Roadmap

Three phases, each with gated entry, defined exit criteria, retro checkpoints, and stakeholder communication cadence. Gate decisions are made by the Automation CoE lead + executive sponsor; no phase advances without both sign-offs.

Phase 1

Foundation & Quick Wins

Weeks 1–8

Deterministic-only

Phase 2

AI-Assisted Operations

Weeks 9–20

Human-in-the-loop AI

Phase 3

Scale & Autonomy

Week 21+

Conditional autonomy

1 Foundation & Quick Wins

Weeks 1–8 2–3 deliveries Low risk

Objective: Prove the pipeline works end-to-end — intake → evaluation → build → deploy → measure. Ship 2–3 deterministic automations that are unambiguously low-risk and high-confidence. No AI touches production. Build delivery muscle, tooling, and stakeholder trust before introducing probabilistic systems.

Candidate Slots (from Step 4 Queue)

Slot	Process	Quadrant	AI Gate	Why This Phase	Target Week
1A	Expense Report Approval	Q1 — Quick Win	PASS	Highest priority score (95), strategic deadline, low effort, well-defined rules — classic workflow automation. No AI needed; deterministic rule engine suffices.	Ship Wk 4
1B	New Hire Onboarding Tickets	Q1 — Quick Win	PASS	Second-highest Q1 priority (77), clean gate, touches HR — good cross-functional visibility. Builds stakeholder confidence across a second department.	Ship Wk 7
1C (stretch)	Email Triage & Routing	Q3 — Filler	PASS	If capacity allows after 1A and 1B are on track. Lower impact but simple rules-based classification — good for onboarding a new team member.	Ship Wk 8

Phase 1 Definition of Done

2 automations live in production, processing real transactions for ≥2 weeks with zero incidents requiring rollback
Each automation has: a runbook, an ownership assignment, and a dashboard/metrics tile showing volume, error rate, and time saved
Intake pipeline processed ≥10 submissions; evaluation rubric applied to ≥8; prioritized queue maintained and visible
Delivery metrics baselined: actual vs. estimated build time, actual vs. estimated effort savings
Playbook v1 drafted: the "how we built it" document covering toolchain, deployment pattern, and lessons learned

Phase 1 → Phase 2 Gate Criteria

All Phase 1 DoD items satisfied.CoE Lead

Retro complete with documented action items. Top 3 lessons distributed to stakeholders.CoE Lead

AI governance body engaged (if one exists). Phase 2 AI candidates presented with their Step 3 gate results. No surprises.CoE Lead + Sponsor

Human-in-the-loop mechanism designed and tested — review UI, confidence thresholding, escalation path. Must exist before any AI output reaches a human reviewer.Tech Lead

AI vendor/self-hosting decision finalized — contracts signed, data-processing agreements executed, security review completed for the model provider(s) Phase 2 will use.Legal + Infosec

1C stretch goal delivered. Not gating, but demonstrates team can parallelize.CoE Lead

Queue refreshed — new submissions scored, sequence re-evaluated against Phase 1 learnings.CoE Lead

🔄 Retro checkpoint — End of Week 8
Attendees: CoE team, executive sponsor, 1–2 process owners from Phase 1 deliveries. Agenda: (1) What worked / what didn't in intake→delivery pipeline, (2) Actual vs. estimated build time and savings — recalibrate Step 2 Ease anchors if off by >50%, (3) Stakeholder feedback on visibility and communication, (4) Go/no-go recommendation for Phase 2. Output: retro doc + updated playbook + signed gate decision.

2 AI-Assisted Operations

Weeks 9–20 3–5 deliveries Moderate risk

Objective: Introduce AI into production — but always with a human between the model and the real world. Build the AI operations playbook: prompt management, confidence scoring, review UX, output quality monitoring, and model drift detection. Start with Q1 CONDITIONAL processes whose guardrails can be resolved quickly, then progress to Q2 Strategic Bets.

⚠ Hard rule for Phase 2: No AI output may take a consequential action (payment, filing, customer-facing communication, system-of-record write) without human approval. This rule relaxes only in Phase 3, and only for processes that clear a heightened autonomy gate.

Delivery Wave Plan

Wave	Process	Quadrant	AI Gate	Guardrails to Clear	Target
2A	Customer Sentiment Classification	Q1 — Quick Win	CONDITIONAL	Human-review loop confirmed; bias monitoring plan drafted; AI governance sign-off obtained	Ship Wk 12
2B	Invoice Exception Handling	Q2 — Strategic Bet	PASS (+ enabler)	None — PASS gate. But as first Q2 item, requires extra design review. Enabler bonus: unlocks downstream AP automation candidates.	Ship Wk 16
2C	Contract Clause Extraction	Q2 — Strategic Bet	CONDITIONAL	Data safeguards (PII in contracts); explainability for extracted clauses; vendor diligence complete	Ship Wk 19
2D (stretch)	Meeting Notes Summarization	Q3 — Filler	PASS	None. Low-risk AI warm-up for new team member.	Ship Wk 20

Phase 2 Internal Milestones

Week	Milestone	Owner
Wk 9	HITL review UX deployed to staging; AI model endpoint configured; first prompt library created	Tech Lead
Wk 10	First AI output reviewed by a human (internal test); confidence scoring calibrated on historical data	Tech Lead + QA
Wk 12	Ship 2A: Sentiment Classification live — AI proposes, human confirms	Delivery Lead
Wk 13	Mid-phase retro: HITL UX feedback, prompt drift check, guardrail status for 2C	CoE Lead
Wk 16	Ship 2B: Invoice Exception Handling live with human approval on exceptions	Delivery Lead
Wk 19	Ship 2C: Contract Clause Extraction live	Delivery Lead
Wk 20	Ship 2D: Meeting Notes Summarization (stretch); AI ops playbook v1 published	Delivery Lead

Phase 2 Definition of Done

≥3 AI-assisted processes live with human-in-the-loop, processing real transactions for ≥2 weeks each
AI ops playbook published: prompt versioning, confidence threshold tuning, output quality monitoring, model drift detection, and incident response for AI-specific failures
HITL metrics baselined: review acceptance rate, time-per-review, AI-suggested vs. human-corrected rate
Bias monitoring running on at least one process; first fairness report produced
All CONDITIONAL guardrails from Phase 2 processes resolved and signed off
Team velocity calibrated: actual build time for AI-assisted processes compared to Phase 1 deterministic baseline

Phase 2 → Phase 3 Gate Criteria

All Phase 2 DoD items satisfied.CoE Lead

Zero AI-related incidents that reached a customer, vendor, or regulator. Internal AI errors caught by HITL are expected and acceptable — they prove the loop works.CoE Lead

Autonomy rubric defined and approved. Which processes from the Phase 3 candidate list qualify for reduced or removed human review? Criteria: blast radius, confidence history (≥95% accuracy over ≥4 weeks), verifiability, reversibility.CoE Lead + Sponsor + Legal

Cost model validated. Actual AI inference costs, HITL labor, and infrastructure spend compared to Phase 1 deterministic baseline. ROI model updated for Phase 3 forecasting.CoE Lead + Finance

Phase 3 candidate list re-scored through Step 2 evaluation and Step 3 AI gate, incorporating what we learned about real AI risk in Phase 2.CoE Lead

Third department engaged. Phase 1 and 2 covered 2 departments. Phase 3 should span ≥3.Sponsor

🔄 Retro checkpoint — End of Week 20
Attendees: CoE team, executive sponsor, legal/compliance rep (first time — Phase 3 autonomy decisions require them), 2 process owners from Phase 2. Agenda: (1) HITL effectiveness — did the loop catch what it needed to? (2) Prompt management maturity — are we versioning, testing, and rolling back prompts cleanly? (3) AI incident near-misses — what did we almost ship? (4) Autonomy rubric review — which processes, if any, are candidates for reduced review? (5) Go/no-go recommendation for Phase 3. Output: retro doc + AI ops playbook v2 + signed autonomy rubric + gate decision.

3 Scale & Autonomy

Week 21+ Ongoing cadence Managed risk

Objective: Scale the pipeline to a steady delivery cadence. For processes that meet the autonomy rubric, reduce or remove human-in-the-loop — the AI acts directly on systems of record, with monitoring and automatic rollback. Introduce more complex AI patterns (multi-step reasoning, agentic workflows) where the AI gate allows. Expand to new departments and use cases.

⚠ Autonomy is earned per process, not granted to the phase. Each process in Phase 3 must individually qualify for reduced review against the autonomy rubric. Processes that don't qualify remain human-in-the-loop indefinitely, regardless of phase.

Autonomy Rubric (Process-Level Qualification)

Criterion	Threshold for Reduced Review	Threshold for Full Autonomy
Accuracy History	≥95% correct over ≥4 weeks of production HITL with ≥500 reviews	≥99% correct over ≥8 weeks with ≥2,000 reviews; zero critical errors
Blast Radius	Internal-only impact; error causes rework, not external exposure	Reversible actions only — every automated action has a rollback path or compensating transaction
Confidence Correlation	Model confidence score predicts correctness — low-confidence outputs are measurably less accurate	Confidence threshold identified where precision is ≥99.5% above that threshold
Drift Stability	No accuracy degradation >2% in any rolling 2-week window	No accuracy degradation >1% in any rolling 4-week window; drift alerting configured
Fallback	Deterministic fallback path exists and has been tested	Fallback runs automatically on confidence below threshold; no human needed to trigger it

Initial Phase 3 Candidate Slots (subject to re-scoring at Phase 2 gate)

Slot	Process	Quadrant	AI Gate	Phase 3 Approach
3A	Vendor Risk Assessment	Q2 — Strategic Bet	CONDITIONAL	Start with HITL (dependency must resolve first). If accuracy meets rubric, candidate for reduced review. High value but currently blocked — Phase 2 enabler (Invoice Exception) may unblock it.
3B	Regulatory Report Generation	Q2 — Strategic Bet	FAIL → Deterministic track	RPA/deterministic build. NOT an AI candidate — the AI gate failure was definitive. Build as rules-based report assembly with human sign-off. Deadline pressure makes this the top deterministic-track item.
3C	New candidates from intake	—	—	Continuous intake pipeline produces new candidates. Phase 3 should have 5–8 active items in the prioritized queue at all times. Target: 1 delivery every 2–3 weeks.
3D	Cross-department expansion	—	—	Proactive outreach to departments not yet represented in the pipeline. Run a 2-hour "automation workshop" per department to generate 3–5 qualified submissions.

Phase 3 Definition of Done (Rolling)

≥1 process operating with reduced HITL under the autonomy rubric, with ≥4 weeks of stability
Steady delivery cadence established: 1 go-live every 2–3 weeks, mix of AI-assisted and deterministic
Pipeline spans ≥3 departments; intake submissions arriving at ≥2/month without active solicitation
Governance loop (Step 6) running: quarterly pipeline reviews, score recalibration, stale candidate pruning
AI ops runbook matured: prompt regression testing, automated drift detection, incident playbook tested with a tabletop exercise
Cost model reviewed quarterly — actual savings tracked against projections from evaluation sheet

Stakeholder Communication Cadence

Applies across all phases. Escalation path: CoE Lead → Executive Sponsor → Steering Committee.

CoE Team

Daily standup (15 min). Blockers, today's focus, WIP limit check. Async update in Slack/Teams if nothing to flag.

Process Owners (Active)

Weekly status email or Slack post. Build progress, UAT schedule, go-live date confidence. Flag scope changes immediately.

Executive Sponsor

Biweekly 30-min sync. Pipeline health, phase gate status, blocker escalation, resource needs. Phase gate decisions happen here.

Leadership / Steering Committee

Monthly dashboard + 1-pager. Deliveries completed, savings realized, pipeline depth, upcoming 4-week plan. Phase gates presented as go/no-go decisions.

All-Hands / Company-Wide

Quarterly 2-slide update. What shipped, what's next, how to submit a process. Purpose: drive intake submissions and visibility.

Legal / Compliance / Infosec

Per AI process — review at Step 3 gate (before build) and again at pre-go-live. Plus quarterly AI risk review covering all live AI processes.

Phase advancement authority: Gate decisions require joint sign-off from the Automation CoE Lead and the Executive Sponsor. Either can veto. A veto does not kill the initiative — it triggers a 2-week remediation sprint to address the blocking concern, after which the gate is re-evaluated.

Integration: This roadmap consumes the sequenced queue from Step 4 and the AI gate results from Step 3. It feeds directly into Step 6 — the governance loop — which monitors actual delivery against this plan and triggers the quarterly recalibration cadence referenced throughout.