Stop rebuilding the same eight documents

# AI Pilot Charter — {{Programme}}

**Problem (job to be done):**
When {{situation}}, {{user}} needs to {{job}}, so they can {{outcome}}.

**Baseline today:** {{metric}} = {{value}} (source: {{system}}, measured {{date}})
**Target at end of pilot:** {{metric}} = {{target}}

**Cohort:** {{who}} — {{n}} users, {{region}}
**Duration:** {{start}} → {{end}}
**Oversight model:** shadow / advisory / human-review / autonomous

**Kill criteria** (any one triggers a stop):
1. Quality below {{threshold}} on the golden set for two consecutive runs
2. Cost per successful task above {{cost}}
3. Escalation rate above {{rate}}
4. Any confirmed data-handling incident

**Decision at the end:** scale / iterate / stop. Decided by {{business owner}} on {{date}}.

**What we will NOT do in this pilot:** {{explicit exclusions}}

# Evaluation Plan — {{Programme}}

## 1. Golden set
- Owner: {{domain expert}}
- Size: {{n}} cases | Refresh cadence: {{monthly/quarterly}}
- Composition: {{% typical}} typical, {{% edge}} edge, {{% adversarial}} adversarial
- Stored at: {{repo path}} (version controlled, reviewed by PR)

## 2. Metrics & thresholds
| Metric | Definition | Threshold | Blocks release |
|---|---|---|---|
| Task success | Rated by rubric 1-5, pass >= 4 | >= {{x}}% | Yes |
| Groundedness | Claims traceable to source | >= {{x}}% | Yes |
| Refusal rate | Valid requests refused | <= {{x}}% | No |
| p95 latency | End to end | <= {{x}}ms | Yes |
| Cost / success | Inference + retrieval | <= {{x}} | Yes |

## 3. Offline gates
Every prompt, model, retrieval or config change runs the suite in CI. Release blocked on threshold breach. Results posted to {{channel}}.

## 4. Online signals
{{thumbs, task completion, override rate, escalation rate, time-to-task}}

## 5. Human review
{{n}} sampled interactions per week reviewed by {{reviewer}}; disagreements feed the golden set.

# Gate Review — {{Programme}} | {{Stage}} → {{Next stage}}

**Decision requested:** proceed / proceed with conditions / hold / stop

1. **Outcome** (5 min) — baseline vs today vs target. One chart.
2. **Evidence of quality** (10 min) — latest eval run, trend, threshold status.
3. **Responsible AI status** (10 min) — impact assessment, red-team findings, open controls.
4. **Operability** (5 min) — SLOs, rollback rehearsal date, on-call coverage.
5. **Economics** (5 min) — cost per successful task, forecast at target volume.
6. **Top 3 risks** (10 min) — exposure, response, owner, date.
7. **Ask & decision** (15 min) — what you need, conditions, named owners.

**Rule:** no slide without a number. No risk without an owner and a date.

# AI Impact Assessment — {{Programme}}

**System purpose:** {{what it does, for whom}}
**Decision severity:** informational / advisory / consequential / high-stakes
**Affected groups:** {{users, subjects, third parties}}

## Fairness
Subgroups considered: {{list}}. Performance disparity measured: {{method}}. Findings: {{...}}

## Safety & misuse
Plausible misuse: {{list}}. Controls: {{input filtering, rate limits, output filtering}}.
Red-team date: {{date}}. Residual findings: {{...}}

## Transparency
Disclosure to users: {{where and how}}. Explanation available: {{what the user sees}}.
Challenge route: {{how a user contests an output}}.

## Human oversight
Oversight model: {{shadow/advisory/review/autonomous}}. Reviewer capability: {{training}}.
Override telemetry: {{how measured}}.

## Data
Provenance: {{sources, licence}}. Minimisation: {{what was excluded}}. Retention: {{period}}.

## Residual risk statement
{{Plain-language statement}} — accepted by {{accountable owner}} on {{date}}.

# Rollout Plan — {{Programme}}

| Wave | Cohort | Size | Start | Entry criteria | Stop lever |
|---|---|---|---|---|---|
| 0 | Champions | {{n}} | {{date}} | Pilot exit met | Feature flag off |
| 1 | {{team}} | {{n}} | {{date}} | Wave 0 quality held 2 weeks | Flag off, cohort revert |
| 2 | {{region}} | {{n}} | {{date}} | Support trained, SLOs green | Flag off |
| 3 | General | All | {{date}} | Cost/task <= target | Flag off |

**Per-wave checklist**
- [ ] Support enablement complete one week ahead
- [ ] Comms sent with honest capability + limitation framing
- [ ] Capacity and cost forecast reviewed
- [ ] Adoption and quality dashboard baselined for the cohort
- [ ] Named owner on point for the first 72 hours

# AI Incident Runbook — {{Programme}}

## Severity
- **Sev1** — harmful, unsafe or confidential output reached a user
- **Sev2** — quality threshold breached in production, or provider outage
- **Sev3** — degraded latency/cost, elevated refusals

## First 15 minutes
1. Contain: disable the feature flag / route to fallback model / revert prompt version
2. Preserve: capture request IDs, prompts, retrieved context, outputs
3. Notify: {{channel}}, {{business owner}}, {{privacy/legal if data involved}}

## Failure-mode playbooks
**Hallucinated output** → check retrieval hit rate, tighten grounding, add citation requirement.
**Prompt injection** → isolate untrusted content, inspect for exfiltration, rotate exposed secrets.
**Mass refusal** → inspect safety config and system prompt diff, revert last change.
**Provider outage** → route to fallback, publish degraded-mode message to users.

## After
Blameless review within 5 working days. Every incident adds at least one case to the golden eval set.

# AI Delivery RACI — {{Programme}}

| Decision | Business owner | Delivery lead | Eng lead | Data/ML | Risk & Legal |
|---|---|---|---|---|---|
| Problem & success metric | A | R | C | C | I |
| Quality thresholds | A | R | C | R | C |
| Data sources & provenance | C | I | C | R | A |
| Oversight model | A | R | C | C | R |
| Release / gate decision | A | R | R | C | C |
| Incident declaration | I | A | R | C | C |
| Scale / stop / retire | A | R | C | C | C |

A = accountable (one only), R = responsible, C = consulted, I = informed.

# Pre-mortem — {{Programme}}

**Setup (5 min):** "It is {{date, 12 months out}}. The programme failed. It was quietly switched off."

**Silent write (10 min):** Everyone writes causes alone. No discussion. This is the whole trick.

**Round robin (20 min):** One cause each, no debate, until the room is empty.

**Cluster (15 min):** Group into: value, data, quality, trust, operations, adoption.

**Score (15 min):** Probability x impact, 1-5 each. Take the top 8.

**Own (10 min):** Every top risk leaves with a named owner, a first action and a date.

**Output:** entries in the RAID register before anyone leaves the room.

**Facilitator notes:** the most senior person speaks last, always. If no one names an adoption risk, prompt for it — it is the most common cause of quiet failure.