# AI Pilot Charter: [pilot name]

> Working planning template for a bounded AI experiment. It does not approve production deployment and is not legal, privacy, security, compliance or safety advice. Adapt it with the responsible owners.

Charter version: [version]
Status: [draft, approved for controlled test, revised, closed]
Sponsor: [name and role]
Business owner: [name and role]
Technical owner: [name and role]
Data owner: [name and role]
Risk reviewer: [name and role]
Pilot operator: [name and role]
Gate decision maker: [name and role]
Planned decision date: [date]
Evidence location: [controlled repository]

## 1. Decision summary

Business decision this pilot will support:
[What will be decided after the pilot]

Why this decision matters now:
[Observed operating or commercial reason]

Pilot type:
[proof of concept, controlled pilot, other]

Requested authorization:
[The exact experiment this charter authorizes]

This charter does not authorize:
[Production launch, broad access, external actions or other exclusions]

## 2. Business problem and current baseline

Target workflow:
[Named workflow]

Workflow start:
[Verified trigger and required input]

Current steps:
[People, systems and decisions]

Successful terminal state:
[Observable completed business outcome]

Observed problem:
[Delay, cost, quality issue, risk, backlog or missed outcome]

Affected users or customers:
[Roles or groups]

Baseline evidence window:
[Dates or sample period]

Baseline source:
[System, report, sample or owner]

Baseline table:

| Measure | Current value | Unit | Source | Confidence or gap |
|---|---:|---|---|---|
| Volume | [value] | [unit per period] | [source] | [verified or unknown] |
| Cycle time | [value] | [time per unit] | [source] | [verified or unknown] |
| Accepted outcome rate | [value] | [% or ratio] | [source] | [verified or unknown] |
| Exception or rework rate | [value] | [% or ratio] | [source] | [verified or unknown] |
| Human effort | [value] | [time per unit] | [source] | [verified or unknown] |
| Direct cost | [value] | [currency per unit] | [source] | [verified or unknown] |

Unknowns requiring discovery:
[List]

## 3. Hypothesis and target job

Hypothesis:

If [bounded capability] supports [named users] with [target job], then
[critical metric] will change from [baseline] toward [acceptance threshold],
without exceeding [severe failure or risk limit].

Why this approach may fit:
[Evidence or operating judgment]

Alternative without AI:
[Rule, workflow, product or process alternative]

Pilot unit:
[One comparable unit of completed work]

Learning questions:

1. [Question]
2. [Question]
3. [Question]

## 4. Scope and exclusions

In scope:

- users: [roles or named group]
- workflow: [start to terminal state]
- systems: [approved systems and environments]
- data: [approved sources and classes]
- actions: [read, draft or bounded writes]
- volume: [maximum units, users or requests]
- channel: [approved interface]
- region or business unit: [boundary]

Out of scope:

- [Excluded workflow]
- [Excluded data]
- [Excluded action]
- [Excluded user or environment]

Scope change rule:
[Who may approve a change and how it is recorded]

Stop condition for scope drift:
[Condition that pauses the pilot]

## 5. Owners and responsibilities

| Responsibility | Named owner | Decision or deliverable | Backup |
|---|---|---|---|
| Business acceptance | [name] | [decision] | [name] |
| Technical delivery | [name] | [deliverable] | [name] |
| Data access and quality | [name] | [decision] | [name] |
| Evaluation design | [name] | [deliverable] | [name] |
| Risk review | [name] | [decision] | [name] |
| Pilot operation | [name] | [deliverable] | [name] |
| Incident handling | [name] | [decision] | [name] |
| Final gate | [name] | [go, revise, hold or stop] | [name] |

Escalation contact:
[Channel and expected handling]

## 6. Users, data, systems and permissions

Primary users:
[Roles, experience and access]

People affected by outputs:
[Customers, employees, partners or other groups]

Data inventory:

| Data or source | Purpose | Class or sensitivity | Access | Retention | Owner |
|---|---|---|---|---|---|
| [source] | [purpose] | [class] | [permission] | [period] | [owner] |

Prohibited data:
[List]

Known data limitations:
[Missing, stale, biased, inconsistent or incomplete data]

System and action inventory:

| System | Environment | Action | Identity | Minimum permission | Approval | Reversible |
|---|---|---|---|---|---|---|
| [system] | [test or controlled live] | [read, draft, write, send] | [identity] | [permission] | [role or none] | [yes, no, path] |

Default when no permission rule matches:
[Recommended: deny, record and escalate]

The pilot may not expand its own access or treat silence as approval.

## 7. Solution boundary

Deterministic steps:
[Rules or workflows that do not require model judgment]

AI judgment steps:
[Classification, extraction, generation or planning tasks]

Human decisions:
[Decisions that remain with named people]

Model and provider dependencies:
[Models, services, hosting and material third parties]

Knowledge sources and freshness:
[Approved sources and update rules]

State or memory:
[What persists, for how long and who may access it]

Required trace:

`input reference -> source evidence -> model output -> proposed action -> policy decision -> human decision -> execution receipt -> verified outcome`

## 8. Evaluation plan

Evaluation unit:
[Same unit used in the baseline]

Evaluation dataset source:
[Sanitized history, controlled live sample, synthetic cases or combination]

Dataset separation:

- development or tuning cases: [location and owner]
- final evaluation cases: [location and owner]
- holdout or later monitoring cases: [location and owner]

Required case groups:

- normal cases;
- edge cases;
- known historical failures;
- ambiguous or incomplete inputs;
- untrusted or adversarial inputs where relevant;
- dependency and tool failures;
- cases that should abstain;
- cases that require human escalation.

Review method:
[Automated checks, human rubric, domain review or combination]

Human reviewers and calibration:
[Roles, rubric and agreement process]

Evaluation metrics:

| Dimension | Metric | Baseline | Acceptance threshold | Severe failure limit | Evidence |
|---|---|---:|---:|---:|---|
| Business outcome | [metric] | [value] | [value] | [limit] | [artifact] |
| Process | [metric] | [value] | [value] | [limit] | [artifact] |
| Quality | [metric] | [value] | [value] | [limit] | [artifact] |
| Failure | [metric] | [value] | [value] | [limit] | [artifact] |
| Human effort | [metric] | [value] | [value] | [limit] | [artifact] |
| Safety or authority | [metric] | [value] | [value] | [limit] | [artifact] |
| Cost | [metric] | [value] | [value] | [limit] | [artifact] |

Reproducibility record:
[Model, prompt, tool, policy, data and configuration versions]

## 9. Human control and failure handling

Approval points:

| Action | Exact payload shown | Approver | Timeout | Rejection behavior | Receipt |
|---|---|---|---|---|---|
| [action] | [fields] | [role] | [duration] | [deny or revise] | [record] |

Escalation triggers:
[Confidence, ambiguity, risk, value, recipient or other conditions]

Human handoff:
[Owner, context transferred, acceptance and user communication]

Failure scenarios to test:

1. timeout before an external action;
2. timeout after a system may have accepted an action;
3. duplicate request;
4. partial integration failure;
5. invalid, missing or stale data;
6. model or provider unavailable;
7. reviewer unavailable;
8. harmful, prohibited or out-of-scope request.

Rollback or compensation:
[Path by action]

Kill switch:
[Owner, mechanism and verification]

Incident channel and record:
[Location and minimum fields]

## 10. Operating measurement

Pilot observation window:
[Dates, number of units or both]

Measurement source:
[Logs, system of record, reviewer record and cost source]

Operating metrics:

| Measure | Definition | Owner | Collection method | Review cadence |
|---|---|---|---|---|
| Accepted units | [definition] | [owner] | [method] | [cadence] |
| Human interventions | [definition] | [owner] | [method] | [cadence] |
| Exceptions | [definition] | [owner] | [method] | [cadence] |
| Detected failures | [definition] | [owner] | [method] | [cadence] |
| Silent failures found later | [definition] | [owner] | [method] | [cadence] |
| Cycle time | [definition] | [owner] | [method] | [cadence] |
| Cost per accepted unit | [definition] | [owner] | [method] | [cadence] |

Comparison rule:
[How pilot units match the baseline]

Evidence integrity:
[How exclusions, missing data and changes are recorded]

## 11. Schedule, resources and budget

Checkpoints:

| Checkpoint | Evidence required | Owner | Planned date | Exit condition |
|---|---|---|---|---|
| Charter approval | [artifact] | [owner] | [date] | [condition] |
| Data ready | [artifact] | [owner] | [date] | [condition] |
| Evaluation ready | [artifact] | [owner] | [date] | [condition] |
| Controlled test ready | [artifact] | [owner] | [date] | [condition] |
| Pilot review | [artifact] | [owner] | [date] | [condition] |
| Final gate | [artifact] | [owner] | [date] | [condition] |

Required people and availability:
[List]

Required environments and systems:
[List]

Budget:

| Cost area | Approved limit | Basis | Owner | Actual |
|---|---:|---|---|---:|
| Discovery and design | [value] | [basis] | [owner] | [value] |
| Data preparation | [value] | [basis] | [owner] | [value] |
| Build and integration | [value] | [basis] | [owner] | [value] |
| Evaluation and review | [value] | [basis] | [owner] | [value] |
| Model and infrastructure | [value] | [basis] | [owner] | [value] |
| Support and contingency | [value] | [basis] | [owner] | [value] |

Dependencies and assumptions:
[List]

## 12. Decision gates, evidence pack and sign-off

Decision rules:

### Go

Use when:
[Critical thresholds pass, severe failures stay within limits and owners accept residual risk]

Next authorized scope:
[The next bounded scope, not blanket production approval]

### Revise

Use when:
[The hypothesis remains plausible and a specific change can be retested]

Required revision:
[Change, owner, evidence and new date]

### Hold

Use when:
[Data, access, ownership, compliance or dependency evidence is missing]

Condition to resume:
[Blocker, owner and proof]

### Stop

Use when:
[The outcome lacks value, representative evaluation fails or risk exceeds the boundary]

Closure actions:
[Remove access, preserve evidence, notify users and close costs]

Required evidence pack:

- approved charter and change log;
- baseline sources;
- data and system inventory;
- evaluation dataset description;
- evaluation results by case group;
- severe failure and incident record;
- human intervention record;
- cost and resource record;
- model, prompt, tool and policy versions;
- decision rationale;
- access removal or next-scope approval.

Sign-off:

| Role | Name | Decision | Date | Conditions |
|---|---|---|---|---|
| Sponsor | [name] | [decision] | [date] | [conditions] |
| Business owner | [name] | [decision] | [date] | [conditions] |
| Technical owner | [name] | [decision] | [date] | [conditions] |
| Risk reviewer | [name] | [decision] | [date] | [conditions] |
| Gate decision maker | [name] | [decision] | [date] | [conditions] |

## Production boundary

A successful pilot does not approve production. Before broader authority, run a separate production readiness review covering access, reliability, monitoring, support, incidents, change management and operating ownership.

## Primary references

- Brazilian Government Unified AI Guide: https://www.gov.br/governodigital/pt-br/infraestrutura-nacional-de-dados/inteligencia-artificial-1/publicacoes/guia-unificado-de-inteligencia-artificial-para-o-setor-publico-1/guia-unificado-de-inteligencia-artificial-para-o-setor-publico
- NIST ARIA pilot evaluation report: https://www.nist.gov/publications/assessing-risks-and-impacts-ai-aria-pilot-evaluation-report
- NIST AI Risk Management Framework Core: https://airc.nist.gov/airmf-resources/airmf/5-sec-core/
- OpenAI evaluation best practices: https://developers.openai.com/api/docs/guides/evaluation-best-practices
- UK Government guidelines for AI procurement: https://www.gov.uk/government/publications/guidelines-for-ai-procurement/guidelines-for-ai-procurement