行业知识 · v2.3.0 · 资料核对 2026-10-03
AI Project Delivery Playbook
Phase gates
| Gate | Evidence | Exit decision |
|---|---|---|
| Problem validated | user workflow, baseline, value, owner | fund discovery / stop |
| Data feasible | source, authority, quality, access, sample | proceed / narrow / stop |
| Technical feasible | prototype, baseline comparison, risks | build / change / stop |
| Product defined | PRD, scope, UX, metrics, acceptance | commit delivery baseline |
| Offline ready | evaluation and hard gates passed | operational validation |
| Launch ready | security, capacity, monitoring, rollback, support | canary / hold |
| Canary accepted | online quality, safety, cost, user outcome | expand / rollback |
| Handover accepted | ownership, runbook, SLA/SLO, maintenance | close project |
Do not commit full-scale scope and dates before critical feasibility evidence exists.
Work breakdown
Include work packages for:
- Product discovery, PRD, UX, adoption, and training.
- Data sourcing, contracts, permissions, cleaning, labeling, and governance.
- Model, Prompt, retrieval, tools, workflow, integration, and infrastructure.
- Evaluation set, rubric, baselines, regression, red team, and acceptance.
- Security, privacy, legal/compliance, threat model, and audit.
- Observability, capacity, cost, rollout, support, incident, and operations.
- Vendor selection, procurement, SLA, change notice, and exit.
- Project governance, communication, change, decision, and documentation.
Each package needs deliverable, owner, estimate/range, dependency, acceptance, and risk.
Estimation under uncertainty
Use three-point ranges for uncertain work. State assumptions. Isolate research/Spike work with a time box, question, method, budget, evidence, and decision. Add contingency based on identified risk, not arbitrary percentage alone.
Avoid treating “model quality” as a single task. Break it into dataset, baseline, experiments, error analysis, remediation, and regression.
RACI and decisions
Use one Accountable owner per outcome or decision. Clarify decision rights for scope, architecture, data use, residual risk, budget, model/vendor, launch, incident shutdown, and acceptance.
Do not make everyone Responsible or Accountable. Consult only roles whose expertise is needed; inform others asynchronously.
RAID
- Risk: uncertain event that may occur.
- Assumption: believed true for planning and must be tested.
- Issue: problem already happening and requiring action.
- Dependency: external deliverable or decision needed.
Write risks as “If [cause/event], then [impact].” Add probability, impact, trigger, prevention, contingency, owner, due/review, and status.
Change control
Require a change record when scope, milestone, budget, quality floor, model/vendor, data use, deployment region, autonomy, risk, or external dependency materially changes.
Show options:
- Accept change and update time/resources.
- Swap scope of equivalent cost.
- Defer to later phase.
- Run discovery before commitment.
- Reject based on value/risk.
Status reporting
Lead with outcome and evidence, then baseline variance, top risks/issues, next milestone, and decisions required. Use red/amber/green only with explicit definitions and recovery actions.
Acceptance and closure
Acceptance covers function, evaluation, safety/privacy, performance, cost, documentation, observability, rollback, support, training, vendor obligations, and ownership. Close only after open items have owners outside the project, financial/vendor closure is complete, assets are archived locally, and lessons/actions are recorded.
Recovery sequence
- Reconstruct facts and the current critical path.
- Identify root cause and remaining uncertainty.
- Quantify outcome, schedule, cost, quality, and risk impact.
- Generate feasible tradeoff options.
- Obtain a decision from the authorized owner.
- Rebaseline, communicate, and monitor recovery evidence.