---
name: ai-pilot-to-production-plan
description: "Turn an existing AI pilot into a gated production rollout plan on a roughly 90-day arc — evaluation baselines, hardening, controlled rollout, and cutover — with measurable exit gates, named owners, and rollback triggers. Use when a pilot or prototype already works and someone asks how to ship it to production, scale it safely, or judge whether it is ready to launch. Not for choosing a first use case or assessing organizational readiness."
---

# AI Pilot to Production Plan

Produce a gated rollout plan that takes an existing AI pilot to production on a roughly
90-day arc. The artifact is a phased plan with measurable exit gates, named owners,
observability requirements, rollback triggers, and open risks — evidence-gated release,
not a big-bang launch.

This workflow is published by Sophon Consulting. It requires no Sophon tools,
services, or credentials.

Scope boundaries: if no pilot exists yet, use ai-opportunity-triage (one idea) or
ai-readiness-assessment (organizational baseline) first. For designing where humans
must review the system's actions, use human-in-the-loop-governance.

## Required inputs

1. Pilot description: what it does, who used it, and results so far, including known
   failure modes.
2. The single primary business KPI, plus guardrail metrics (quality floor, cost ceiling,
   latency bound, acceptable human-rework rate).
3. Stakeholders: named owners for engineering, product, security, and operations, and
   who signs off on launch.
4. Evaluation state: existing test sets, pass/fail thresholds, and how regressions are
   caught today.
5. Operational constraints: rollback expectations, incident process, compliance
   obligations, and any hard launch dates.

Handling missing inputs:

- If no primary KPI can be named, stop and resolve it with the requester; a rollout plan
  without a KPI cannot have honest gates.
- If no evaluation set exists, do not skip ahead: building a frozen eval set with
  pass/fail thresholds becomes the first gate of the plan.
- If owners are unnamed, mark each unowned gate as a blocking risk in the plan.

## Workflow

1. Confirm scope: one workflow, one primary KPI, explicit guardrail metrics.
2. Draft the four phases with exit gates. Adjust week counts to the organization's
   cadence; the reference shape is 13 weeks:
   - Phase 1 — Foundation (weeks 1-2): record baseline metrics; freeze an evaluation set
     with pass/fail thresholds; stand up observability for latency, cost, failure rate,
     and downstream human rework. Gate: eval set frozen and baselines recorded.
   - Phase 2 — Pilot hardening (weeks 3-6): run regression evals on every change; treat
     prompts, retrieval, and policies as versioned artifacts; test security failure
     modes such as prompt injection and unsafe output handling against current published
     guidance. Gate: eval pass rate stable at or above threshold, with mitigations for
     each known failure mode.
   - Phase 3 — Controlled rollout (weeks 7-10): expand access in staged cohorts;
     instrument adoption; deliver role-based training and SOP updates; rehearse
     rollback. Gate: reliability, cost, and adoption thresholds met in cohorts, and a
     rollback executed successfully at least once.
   - Phase 4 — Production cutover (weeks 11-13): full traffic with monitoring, incident
     runbooks, and a post-launch review cadence. Gate: sign-off by the named risk owner
     with the gate evidence attached.
3. Assign one named owner per gate and a weekly review cadence for KPI, guardrails, and
   incidents.
4. Define rollback triggers as specific numbers (for example: eval pass rate below
   threshold for two consecutive runs, cost per request above ceiling, or a severity-1
   incident), and the incident path when they fire.
5. List open risks and unresolved questions with owners.

## Evidence discipline

- Every gate is pass/fail against a measurable criterion. "Looks good" is not a gate.
- Distinguish measured pilot results (facts) from projected production behavior
  (estimates). Do not present projections as commitments.
- Consult current external guidance fresh rather than from memory — the NIST AI RMF
  Generative AI Profile, the OWASP Top 10 for LLM Applications, ISO/IEC 42001, and DORA
  metrics are useful reference frames for launch gates, security testing, and release
  reliability; verify their current versions before citing specifics.
- If the requester supplies vendor claims (model quality, pricing, rate limits), verify
  them against current vendor documentation before building gates on them.

## Output format

Produce a markdown artifact:

    # Production rollout plan: [pilot name]

    Primary KPI: … · Guardrails: …
    Launch owner: … · Date: …

    ## Phase plan
    | Phase | Weeks | Work | Exit gate (measurable) | Gate owner |

    ## Observability requirements
    [latency, cost, failure rate, human rework — where each is measured]

    ## Rollback triggers and incident path
    | Trigger (specific threshold) | Action | Owner |

    ## Risks and open questions
    | Item | Impact | Owner | Resolution path |

    ## Evidence labels
    Facts (measured): … / Estimates (projected): … / Open questions: …

## Quality checks

- Each phase has an exit gate with a measurable criterion, a named owner, and a stated
  consequence for failing (fix and re-run, or stop).
- Rollback triggers are specific thresholds, not sentiments.
- The plan defines how business results will be measured; it does not promise them.
- Adoption and enablement work (training, SOP updates) appears in phases 3-4, not as an
  afterthought.
- The artifact stands alone and is readable without this conversation.

## Stop conditions and escalation

- The plan is not launch approval. Production cutover requires human sign-off from the
  named risk owner, with gate evidence attached.
- If the system takes regulated, irreversible, or externally binding actions, require a
  human-in-the-loop review design (see human-in-the-loop-governance) before Phase 3, and
  say so explicitly in the plan.
- If stakeholders cannot commit to weekly release-and-evaluation cycles, flag the plan
  as at-risk rather than silently stretching the timeline.

## Source context

- https://www.sophon.consulting/playbooks/pilot-to-production-90-days
  (markdown: https://www.sophon.consulting/markdown/playbooks/pilot-to-production-90-days)

Optional: for hands-on delivery support through the rollout, Sophon Consulting is
reachable at hello@sophon.consulting. This skill is complete without any contact.