行业知识 · v2.3.0 · 资料核对 2026-10-03

Deliver AI Project

Manage uncertainty explicitly. Plan AI projects around evidence and risk reduction, not only feature completion.

Operating rules

Separate product ownership from project control, even when one person performs both roles.
Treat data, evaluation, safety, operations, and user adoption as first-class work packages.
Use ranges and assumptions for uncertain work; time-box technical discovery.
Never hide a scope increase inside an unchanged schedule or budget.
Escalate with facts, impact, attempted actions, options, and a requested decision.
Do not mark work complete until acceptance evidence and operating ownership exist.

Workflow

1. Establish the charter

Define purpose, outcome, scope/non-scope, sponsor, product owner, project manager, decision rights, constraints, assumptions, budget, milestones, risk level, acceptance, operations owner, and closure criteria.

2. Build phase gates

Read references/delivery-playbook.md. Use gates such as problem validated, data available, baseline established, feasibility proven, offline quality passed, operational readiness passed, canary passed, and scaled release accepted.

3. Plan the work

Create WBS packages for product, UX, data, model, retrieval/tools, engineering, evaluation, security/compliance, operations, training/adoption, procurement, and project governance. Map dependencies and identify the critical path.

4. Assign ownership and cadence

Create RACI for major deliverables and decisions. Set standup, risk, decision, steering, evaluation, launch, and vendor cadences. Give every action one accountable owner and due date.

5. Control uncertainty

Maintain RAID and decision logs. Define triggers, preventive actions, contingencies, reserves, and escalation thresholds. Convert unknowns into time-boxed experiments with a decision at the end.

6. Control change and recovery

For each change, show benefit, cost, risk, dependency, baseline impact, options, approver, and new commitment. For delay, choose explicitly among scope, time, resources, quality above the minimum, or solution approach.

7. Accept and hand over

Verify functional, evaluation, safety, performance, cost, observability, rollback, documentation, training, vendor, and operations criteria. Transfer ownership and confirm support, incident, maintenance, and re-evaluation processes.

8. Produce the artifact

Copy assets/ai-project-charter-and-plan.md. Use assets/raid-log.csv for a local RAID register.

Quality gate

Confirm that the delivery system includes:

Outcome and acceptance, not just feature lists.
A baseline schedule with assumptions and ranges.
Data and evaluation work before final development commitment.
Named decision makers and clear escalation paths.
Critical dependencies, external vendors, and fallback options.
Risks for quality, privacy, security, cost, adoption, and operations.
Controlled change with updated commitments.
Go/No-Go, rollback, handoff, retrospective, and closure.

If the date, scope, resources, and quality are all declared fixed while evidence shows the plan is infeasible, surface the conflict and request a tradeoff decision.