Practical guide

AI launch plan verification checklist

A launch plan is ready for a decision only when a reviewer can trace its claims to current evidence, reproduce its gates, identify who may pause the release, and exercise the fallback path. A polished roadmap is not enough. This checklist verifies an existing AI launch plan. It complements the FLIGHT planning method by testing whether the plan can actually control a release. Copy it into the.

By Two Prune

A launch plan is ready for a decision only when a reviewer can trace its claims to current evidence, reproduce its gates, identify who may pause the release, and exercise the fallback path. A polished roadmap is not enough. This checklist verifies an existing AI launch plan. It complements the FLIGHT planning method by testing whether the plan can actually control a release. Copy it into the launch record, complete every field, and attach evidence links rather than relying on verbal assurance. The checklist uses CLEAR: Context, Limits, Evidence, Authority, and Recovery.

Verification header

Freeze these fields before the review. If the running system changes after verification, record the change and decide which checks must be repeated.

  • [ ] Capability and release version:
  • [ ] Intended audience and maximum exposure:
  • [ ] Environment and planned release window:
  • [ ] Launch decision owner:
  • [ ] Independent reviewer:
  • [ ] AI model, service, prompt, retrieval source, and policy versions:
  • [ ] Plan version and evidence cutoff time:
  • [ ] Review outcome: pass / repair / hold / stop
  • [ ] Next review date:

C: Verify the decision context

Ask the plan owner to demonstrate the journey in a production-like environment. A diagram alone does not show that identity, permissions, data flow, AI behavior, and fallback work together.

The GOV.UK Service Manual recommends that an AI-enabled service assessment include a service brief, recent prototype or development environment, high-level architecture, an explanation of how AI is used, and team members able to answer questions across ownership, technology, research, design, and performance. That is a useful evidence packet for this check, even when the service is outside government. GOV.UK: Book a service assessment

  • [ ] The plan states the exact release decision, not only a target date.
  • [ ] The user need is supported by current research or operational evidence.
  • [ ] The team compared AI with a non-AI or simpler alternative.
  • [ ] The intended user journey is complete from entry to support or exit.
  • [ ] The release unit is small enough to control but large enough to test the real journey.
  • [ ] Dependencies, integrations, data sources, and affected teams are named.
  • [ ] Success measures have definitions, denominators, segments, windows, owners, and linked decisions.
  • [ ] Assumptions and unresolved questions are separated from observed facts.

Sources: [1]

L: Verify limits and user protection

Current GOV.UK guidance for AI in services emphasizes user need, clear communication about effects on data and outcomes, accuracy, security involvement, continuous monitoring, and readiness to stop using the AI technology. Applicability depends on context, but the guidance supports evidence-led review of the boundary. GOV.UK: Using artificial intelligence in services

  • [ ] Intended use and excluded uses are explicit.
  • [ ] The plan identifies affected people, foreseeable misuse, and material failure modes.
  • [ ] Users can tell when AI changes the information or outcome they receive.
  • [ ] A human contact, escalation, challenge, or override route is available where the workflow needs one.
  • [ ] Permitted data classes, storage, access, retention, and deletion paths are documented.
  • [ ] Security, privacy, accessibility, records, legal, and operational reviewers are involved where relevant.
  • [ ] Known limitations appear in the interface or guidance at the point they matter.
  • [ ] The launch team can stop using the AI component without losing the whole service where feasible.

Sources: [2]

E: Verify the evidence gates

For every gate, require a row in this table:

| Gate | Acceptance rule | Test method | Frozen evidence | Result | Reviewer | Exception | | --- | --- | --- | --- | --- | --- | --- | | | | | link and version | pass / fail / incomplete | | |

NIST's AI RMF Core calls for documented test methods and results, evaluation in conditions resembling deployment, independent review where appropriate, and post-deployment mechanisms for feedback, override, incident response, recovery, and change management. Use those outcomes to challenge the evidence packet, not as a universal release verdict. NIST AI RMF Core

  • [ ] Test cases represent ordinary, edge, adversarial, and out-of-scope conditions.
  • [ ] Expected results or scoring rules were frozen before the final run.
  • [ ] Tests use conditions similar to the intended deployment context.
  • [ ] Results include error categories, uncertainty, and examples rather than one aggregate.
  • [ ] Retrieval, citation, calculation, access, refusal, and fallback behavior are tested where relevant.
  • [ ] The reviewer can reproduce a risk-based sample from source inputs to recorded result.
  • [ ] Failed and incomplete checks remain visible.
  • [ ] Exceptions name an owner, exposure, compensating action, expiry, and approval.
  • [ ] The plan defines what triggers continue, repair, hold, stop, or rollback.

Sources: [3]

A: Verify authority and operating ownership

Ask each critical owner to explain their action without help from the plan author. A name in a RACI table is weak evidence if the person cannot access the tools or does not know the trigger.

  • [ ] One person owns the release decision.
  • [ ] The people who may pause or roll back are named and reachable.
  • [ ] Risk acceptance is assigned to an authorized person, not the AI system or launch coordinator by default.
  • [ ] Support, incident, communications, data, security, and vendor owners know their responsibilities.
  • [ ] On-call access, credentials, dashboards, contact routes, and escalation paths have been exercised.
  • [ ] The decision owner has reviewed dissenting views and open exceptions.
  • [ ] The final decision, time, evidence version, and rationale will be recorded.
  • [ ] Post-launch review dates and change-approval authority are set.

R: Rehearse recovery and continuing control

Record the rehearsal:

| Scenario | Start state | Trigger | Expected action | Observed action | Time | Gap | Retest | | --- | --- | --- | --- | --- | --- | --- | --- | | | | | | | | | |

  • [ ] Monitoring detects the failures that matter to the boundary.
  • [ ] Alerts reach a responder within the required time.
  • [ ] The team has rehearsed pause, rollback, degraded mode, and recovery.
  • [ ] Rollback restores compatible data, configuration, model, prompt, and retrieval versions.
  • [ ] The service handles an unavailable or degraded third-party AI dependency.
  • [ ] Incident records preserve the evidence needed for investigation.
  • [ ] Support guidance covers known limitations and escalation.
  • [ ] The plan specifies how user feedback and challenges are reviewed.
  • [ ] Model, prompt, data, policy, and interface changes trigger proportionate retesting.
  • [ ] The service can be decommissioned with clear data and communication steps.

Independent sampling protocol

Choose all critical gates and a risk-based sample of the rest. For each sample, an independent reviewer should:

Expand the sample when a critical result cannot be reproduced, a source version is wrong, an exception is hidden, or the deployed configuration differs from the record.

  • locate the frozen source without guidance;
  • reproduce the test or calculation;
  • confirm the acceptance rule was applied as written;
  • inspect any exception and approval;
  • verify the running version matches the reviewed version.

Worked test 1: internal policy assistant

A plan proposes releasing a policy-search assistant to 300 employees after 94 of 100 test questions receive an acceptable answer. CLEAR holds the release.

The context check finds that all test questions came from the policy team, so ordinary employee wording is missing. The limits check finds no user-facing route for a disputed answer. The evidence check shows the aggregate hides four access-control questions where the assistant summarizes a general policy but misses a controlling exception. The recovery rehearsal also shows that disabling the model leaves a broken page rather than a searchable policy index.

The plan receives a repair outcome. The team adds representative user phrasing, freezes exception-focused tests, creates a human escalation route, and implements the non-AI fallback. A second review samples every access-control question and the rollback scenario before limited exposure begins.

Worked test 2: AI-assisted launch communication

A marketing team uses AI to draft release emails and help-center copy. The underlying product has no AI component. CLEAR narrows the verification boundary: the launch plan does not need model-operation monitoring, but the public copy still requires source control, claim review, approval, and correction handling.

The reviewer traces each feature claim to the released build, catches one sentence describing a future integration as available, and records a repair. The corrected copy passes its source sample. This example shows why the checklist should be tailored to how AI affects the launch, rather than treating every use as an AI product release.

Hold conditions

Hold the launch decision when the reviewed version is unclear, critical evidence is missing, a required owner is absent, the user journey cannot be demonstrated, high-consequence failures are hidden by an aggregate, a hold or exception lacks authority, recovery cannot be rehearsed, or the running configuration differs materially from the evidence packet.

Two Prune can use a launch-plan review as a realistic AI-enabled work assessment. It evaluates how a participant frames the release, navigates evidence, uses AI, verifies claims, handles risk, and communicates a recommendation. It does not approve a product launch, operate the service, or make employment decisions.

Sources

GOV.UK Service Manual: Book a service assessment

  • GOV.UK Service Manual: Using artificial intelligence in services
  • GOV.UK Service Manual: What happens at a service assessment
  • NIST AI Risk Management Framework Core

Sources: [1] [2] [4] [3]

Sources

  1. 1.GOV.UK: Book a service assessment · UK Government
  2. 2.GOV.UK: Using artificial intelligence in services · UK Government
  3. 3.NIST AI RMF Core · National Institute of Standards and Technology
  4. 4.GOV.UK Service Manual: What happens at a service assessment · UK Government