Every step is clickable — see responsibilities for each discipline: PM · Designer · Development · AI
What gets produced and handed between phases — made explicit so nobody has to guess.
Who is in the room at each step, and how distributed teams coordinate across time zones.
| Step | Product Manager | Program Manager | Designer | Development | Technical Writer | AI |
|---|---|---|---|---|---|---|
| Pre-Discovery | Primary | None | Collaborate | Aware | None | Collaborate |
| Discovery | Primary | None | Collaborate | Collaborate | Collaborate | Collaborate |
| Prototype | Collaborate | None | Primary | Collaborate | None | Collaborate |
| Non-Functional Requirements | Primary | Collaborate | Collaborate | Collaborate | None | Collaborate |
| Client/Stakeholder Validation | Primary | Collaborate | Collaborate | Collaborate | None | Collaborate |
| Functional Spec Draft + Confirm | Collaborate | None | Collaborate | Aware | None | Primary |
| Technical Spec Draft + Confirm | Aware | None | Aware | Collaborate | None | Primary |
| Epics/Stories | Primary | None | Collaborate | Primary | None | Primary |
| Build | Collaborate | None | Collaborate | Collaborate | None | Primary |
| UAT | Primary | Collaborate | Collaborate | Collaborate | Collaborate | Collaborate |
| Company Readiness | Primary | Collaborate | Collaborate | Collaborate | Collaborate | Collaborate |
| Release | Collaborate | Collaborate | Collaborate | Primary | Collaborate | Collaborate |
| Adoption | Collaborate | None | Collaborate | Aware | Collaborate | Primary |
| Maintenance loop | Aware | None | None | Collaborate | Aware | Primary |
| Quarterly Planning | Collaborate | Primary | Collaborate | Collaborate | Collaborate | Collaborate |
| Ceremonies | Collaborate | Collaborate | Collaborate | Primary | Collaborate | Collaborate |
| Metrics and Dashboards | Collaborate | Primary | Collaborate | Collaborate | None | Collaborate |
Cross-functional collaboration does not mean pulling every engineer into every product call — that is low value and slows everyone down. The intent is to involve the right discipline at the right moment: Dev shows up early in Discovery to flag feasibility, not to sit through every requirements discussion. The Collaborate and Aware tiers above exist to make these calls intentional rather than habitual.
Teams can use Microsoft Teams, Slack or some other mechanism they choose to facilitate async communications. Rather than the everyone-in-the-room model used by fully co-located teams, cross-time-zone handoffs follow this pattern:
| Step | What happens | Time cost |
|---|---|---|
| 1. Written handoff | Originating team produces a structured artifact (spec, PR, design note) with explicit open questions flagged inline — not left implicit. | Same day |
| 2. Recorded context | For anything nuanced, a short recorded walkthrough accompanies the artifact so the receiving team doesn't need a live call to understand intent. | Same day |
| 3. Agent-assisted triage | Receiving team's agent reviews the handoff against context layer, flags genuine ambiguities versus questions it can answer itself. | Minutes, async |
| 4. Targeted async question | Only genuine ambiguities are escalated to a human, asked specifically rather than open-ended, batched where possible. | One overlap window |
| 5. Scheduled sync (exception, not default) | Reserved for decisions that are genuinely contested or carry high ambiguity. Not the default path for routine handoffs. | As needed |
Where AI does the work, where it assists, and where humans retain full control — across the whole workflow.
The percentage split is directional, not a literal time measurement — it communicates which party is doing the bulk of the generative work at each step. A step that is mostly AI-colored still has a human confirmation point; a step that is mostly human-colored may still use AI for research, drafting, or summarization in the background. No step in the AI-SDLC removes human accountability entirely — every phase that produces a durable artifact has a named human owner.
Terms used throughout the AI-SDLC, defined for readers across all disciplines — not assuming prior familiarity.
A short written record capturing a significant technical decision: what was decided, why (the reasoning), and what alternatives were considered and rejected. ADRs are produced during Technical Spec and carried forward as a required input to engineering — they prevent the same trade-off conversation from happening repeatedly, and they give future contributors context for why the system is built the way it is.
The explicit go decision, made by the Dev lead (HI-CTRL), that releases a scoped work unit into the AI orchestration loop. It is the gate between human direction and agent execution: task intent is set and TDD test cases are confirmed, then dispatch is approved. No agent begins implementation until dispatch approval is given, and execution authority remains with the human throughout.
The agreed checklist a piece of work must satisfy before it counts as complete. In the AI-SDLC, DoD confirmation means: test coverage threshold met, all CI checks passed, human review confirmed, and (where applicable) design fidelity confirmed. DoD confirmation is a quality check that flows directly into automated deployment — it is not a separate stop-and-wait approval step.
Validation that a built feature actually does what was intended — checked against the original functional spec and acceptance criteria, not just against technical requirements. A UAT walkthrough is the structured session (live or recorded) where PM, Designer, and relevant stakeholders step through the feature to confirm it behaves as expected before Release.
The persistent, continuously-updated source of project knowledge that AI agents draw from across every phase — including the codebase, prior ADRs, design artifacts, specs, and decisions. It is what allows an agent drafting a spec or a fix to be "grounded" rather than guessing, and it is the mechanism that replaces manual handoffs between phases.
A confirmation point that work flows through once criteria are met — replacing the term "gate" used in earlier versions of this workflow. The distinction matters: a quality check is part of the continuous flow, not a stop-and-wait approval queue. Work that meets the criteria proceeds immediately; work that doesn't loops back to the relevant step rather than waiting in a queue.
Writing test cases before writing the implementation. In the AI-SDLC, the agent generates test cases from acceptance criteria during Technical Spec — before any code is produced — so the implementation has a concrete target to satisfy from the start.
Shorthand for "human in control" — the principle that a named human holds execution authority (Producer Mode) and merge authority (Reviewer Mode) throughout the AI-orchestrated development loop. No agent output reaches production without HI-CTRL confirmation.
A framework for clarifying who is Responsible, Accountable, Consulted, and Informed (RACI) — or an adapted variant — for a given decision or deliverable. Several open questions in this workflow (acceptance criteria ownership where TPMs remain active, ceremony ownership without a TPM) are pending a team-specific RACI, currently in draft. See the note on the Standup Pre-work and Retro + Learning cards in the Continuous Loop.
A quarterly planning cadence used across all products and teams, where cross-team and cross-product dependencies are surfaced and sequenced before the quarter begins.