# AI system PRD / requirements brief

## 1. Problem statement
- 対象業務:
- 利用者:
- 現行課題:
- AI を使わない場合の代替案:
- 非目標:

## 2. Users and stakeholders
| Role | 利用目的 | Decision right | Concern | Evidence needed |
| --- | --- | --- | --- | --- |
| | | | | |

## 3. Scope
- 対象に含めること:
- 対象外:
- 対象データ:
- 対象外データ / 利用禁止情報:

## 4. Automation boundary
| 業務イベント | Human-in-the-loop | Human-on-the-loop | Full automation | 理由 |
| --- | --- | --- | --- | --- |
| | | | | |

## 5. Acceptance criteria
| 観点 | 条件 | 測定方法 | Owner |
| --- | --- | --- | --- |
| Quality | | | |
| Reproducibility | | | |
| Explainability | | | |
| Fail-safe | | | |
| Security / privacy | | | |
| Approval / audit | | | |
| Rollback | | | |

## 6. Metrics
- KPI:
- Guardrail metric:
- manual takeover rate:
- verification cost:
- audit completeness:

## 7. Risks and assumptions
| ID | Assumption / Risk | Evidence | Owner | Decision needed |
| --- | --- | --- | --- | --- |
| | | | | |

## 8. Approval and next decision
- Decision owner:
- Reviewers:
- 次の判断日:
- 止める条件:

# AI system ADR: <decision title>

- Status: proposed / accepted / superseded / deprecated
- Date:
- Decision owner:
- Reviewers: Architect / Security / SRE / Legal / Compliance
- Related Issue / PR:

## Context
- 解決したい問題:
- 既存システム制約:
- データ分類:
- 利用者と権限:
- latency / cost / quality / reliability budget:

## Options considered
| Option | Description | Benefits | Risks | Cost | Operability |
| --- | --- | --- | --- | --- | --- |
| Normal feature | | | | | |
| Search / RAG | | | | | |
| Workflow | | | | | |
| Agent | | | | | |

## Decision
- 採用する方式:
- 採用理由:
- 採用しない方式と理由:

## Control points
| Control | Design | Evidence |
| --- | --- | --- |
| schema validation | | |
| approval gate | | |
| least privilege | | |
| audit trail | | |
| isolation | | |
| output validation | | |
| fallback / rollback | | |

## Evaluation plan
- offline eval:
- trace-based evaluation:
- regression test:
- acceptance threshold:

## Exit strategy
- vendor portability:
- export するデータ / 設定 / eval:
- provider outage 時の degrade gracefully:
- 乗り換え判断条件:

## Consequences
- Positive:
- Negative:
- Operational burden:
- Follow-up actions:

# Eval spec

## Purpose
- 評価対象:
- 評価目的:
- 本番判断との関係:

## Evaluation scope
| Scenario | User goal | Expected behavior | Prohibited behavior |
| --- | --- | --- | --- |
| | | | |

## Dataset
- データソース:
- データ分類:
- 個人情報 / 機密情報の扱い:
- version:
- owner:

## Metrics
| Metric | Definition | Threshold | Guardrail / KPI | Owner |
| --- | --- | --- | --- | --- |
| answer quality | | | | |
| citation coverage | | | | |
| hallucination rate proxy | | | | |
| tool error rate | | | | |
| latency distribution | | | | |
| token / cost anomaly | | | | |
| manual takeover rate | | | | |
| audit completeness | | | | |

## Failure review
| Failure class | Severity | Required response | Regression item |
| --- | --- | --- | --- |
| prompt injection | | | |
| citation failure | | | |
| retrieval failure | | | |
| approval bypass | | | |
| data leakage | | | |

## Acceptance decision
- Pass / Conditional pass / Fail:
- 条件:
- 承認者:
- 次回 eval:

# Eval dataset design sheet

## Dataset identity
- Name:
- Version:
- Owner:
- Last reviewed:
- Related eval spec:

## Coverage plan
| Category | Example | Required count | Source | Risk covered |
| --- | --- | --- | --- | --- |
| normal case | | | | |
| edge case | | | | |
| adversarial / prompt injection | | | | |
| stale knowledge | | | | |
| citation required | | | | |
| tool approval required | | | | |
| fallback / manual takeover | | | | |

## Data governance
- Data classification:
- Masking / anonymization:
- Data retention:
- Data residency:
- Access control:

## Expected outputs
| Input ID | Expected output | Required citation | Prohibited output | Review owner |
| --- | --- | --- | --- | --- |
| | | | | |

## Change rule
- 追加条件:
- 削除条件:
- drift review:
- approval:

# Threat model

## System overview
- 対象システム:
- Trust boundaries:
- Data flows:
- External dependencies:

## Assets
| Asset | Classification | Owner | Protection requirement |
| --- | --- | --- | --- |
| prompts / instructions | | | |
| retrieved documents | | | |
| user data | | | |
| tool credentials | | | |
| audit logs | | | |

## Threats
| Threat | Attack path | Impact | Existing control | Gap |
| --- | --- | --- | --- | --- |
| prompt injection | | | | |
| tool misuse | | | | |
| data exfiltration | | | | |
| privilege escalation | | | | |
| MCP authorization failure | | | | |
| approval bypass | | | | |
| stale knowledge | | | | |
| cost anomaly | | | | |

## Mitigations
| Control | Implementation | Evidence | Owner |
| --- | --- | --- | --- |
| least privilege | | | |
| schema validation | | | |
| output validation | | | |
| approval gate | | | |
| audit trail | | | |
| rate limit / cost alert | | | |
| kill switch / rollback | | | |

## Residual risk
- Accepted risks:
- Escalation:
- Review date:

# Tool approval matrix

| Tool / connector | Capability | Data access | Risk level | Default permission | Approval gate | Audit log | Rollback / disable |
| --- | --- | --- | --- | --- | --- | --- | --- |
| search docs | read | public / internal | low | allowed | none | query log | disable connector |
| read ticket | read | internal | medium | allowed with scope | owner approval | access log | revoke token |
| create PR | write | repository | medium | approval required | maintainer approval | PR / commit log | revert PR |
| send customer message | external write | customer data | high | disabled by default | Communication + Legal | message log | recall / correction |
| change production config | write | production | critical | disabled by default | Incident Commander + SRE | change log | rollback plan |

## Review
- Owner:
- Last access review:
- Exceptions:
- Expiration:

# Data classification sheet

| Data item | Classification | Allowed use | Prohibited use | Masking | Retention | Residency | Owner |
| --- | --- | --- | --- | --- | --- | --- | --- |
| public docs | public | retrieval / citation | none | none | standard | any allowed | |
| internal docs | internal | retrieval with access control | external AI training | metadata mask | | | |
| customer data | confidential | limited processing | prompt sharing / vendor training | required | | | |
| personal data | restricted | approved workflow only | ad hoc prompt input | required | | | |
| secrets / tokens | prohibited | none | any AI input | not applicable | none | not applicable | Security |
| audit logs | controlled | incident review | external sharing without approval | partial | policy based | | Compliance |

## Approval
- Data owner:
- Security reviewer:
- Legal / privacy reviewer:
- Exception path:

# Vendor selection matrix

| Criteria | Weight | Vendor A | Vendor B | Vendor C | Evidence |
| --- | --- | --- | --- | --- | --- |
| functional fit | | | | | demo / pilot |
| security controls | | | | | security report |
| data residency | | | | | contract / DPA |
| audit log | | | | | docs / sample export |
| SLA / support | | | | | contract |
| cost / pricing risk | | | | | estimate |
| vendor concentration risk | | | | | exit strategy |
| portability / export | | | | | export test |
| contract / liability | | | | | legal review |
| responsible use support | | | | | policy / admin controls |

## Decision
- Recommended option:
- Conditions:
- Exit strategy:
- Renewal review date:

# AI incident postmortem

## Summary
- Incident ID:
- Severity:
- 発生日 / 検知時刻 / 復旧時刻:
- 影響範囲:
- 現在状態:

## Customer / business impact
- 顧客影響:
- 業務影響:
- 法務 / 監査 / 契約影響:
- コスト影響:

## Timeline
| Time | Event | Type: fact / hypothesis / decision / action / result | Evidence ID |
| --- | --- | --- | --- |
| | | | |

## AI involvement
- prompt / input:
- retrieval / citation:
- model / provider:
- tool / MCP / connector:
- approval / audit:
- human review:

## Root causes and contributing factors
- 直接原因:
- 技術的寄与要因:
- 運用上の寄与要因:
- 組織・教育・契約上の寄与要因:

## Containment and recovery
- kill switch:
- quarantine:
- rollback:
- manual takeover:
- escalation:

## Corrective actions
| Action | Type | Owner | Due date | Evidence of completion |
| --- | --- | --- | --- | --- |
| | guardrail / eval / approval / audit / training / contract | | | |

# Executive memo

## Decision requested
- 採用 / pilot / 本番移行 / 予算 / 契約 / 撤退:
- Decision date:

## Recommendation
- 推奨案:
- 代替案:
- 採用しない案と理由:

## Business case
| Item | Estimate | Evidence | Owner |
| --- | --- | --- | --- |
| productivity benefit | | | |
| risk reduction | | | |
| implementation cost | | | |
| verification cost | | | |
| control cost | | | |
| training cost | | | |
| fallback / rollback cost | | | |

## Governance
- security:
- privacy:
- compliance:
- approval:
- audit:
- rollback:

## Risks and conditions
| Risk | Mitigation | Decision condition | Owner |
| --- | --- | --- | --- |
| | | | |

## Exit strategy
- 退出条件:
- export 対象:
- 移行費用:
- 代替案:

## Final decision log
- Decision:
- Approver:
- Conditions:
- Review date:

# 1ページ提案メモ

## 提案
- 対象業務:
- 提案内容:
- 今回決めたいこと:

## 背景と問題
- 現行課題:
- 影響を受ける利用者:
- 放置した場合のリスク:

## 期待効果
- productivity benefit:
- risk reduction:
- 測定方法:

## 費用と統制
- 初期費用:
- 継続費用:
- verification cost:
- control cost:
- training / rollback cost:

## 主要リスク
- 契約 / データ越境:
- security / privacy:
- audit / approval:
- vendor concentration:
- 現場負荷:

## 判断案
- 推奨:
- 代替案:
- 進める条件:
- 止める条件:
- 次の decision date:

# Verification record

## Target
- 対象成果物:
- Related Issue / PR / ADR:
- AI 利用の有無:

## Inputs reviewed
| Input | Source hierarchy | Version / date | Reviewer |
| --- | --- | --- | --- |
| primary docs | | | |
| measured logs | | | |
| AI summary | | | |

## Checks performed
| Check | Command / Method | Result | Evidence |
| --- | --- | --- | --- |
| unit / integration test | | | |
| eval / regression | | | |
| static analysis | | | |
| security review | | | |
| privacy / data boundary | | | |
| approval / audit | | | |
| rollback / fallback | | | |

## Findings
- Accepted:
- Rejected:
- Open issues:

## Decision
- Ship / hold / rollback / rework:
- Decision owner:
- Date:

# Model / tool change impact checklist

## Change summary
- 変更対象:
- 変更理由:
- 影響する workflow / agent / tool:
- Rollback target:

## Impact areas
- [ ] requirements brief への影響を確認した。
- [ ] AI system ADR の decision / trade-off を更新した。
- [ ] eval spec と eval dataset を更新または再実行した。
- [ ] threat model の差分を確認した。
- [ ] tool approval matrix の権限と approval gate を確認した。
- [ ] data classification sheet のデータ利用条件を確認した。
- [ ] cost / reliability dashboard の閾値を確認した。
- [ ] AI incident runbook の kill switch / rollback を確認した。
- [ ] communication template と support FAQ の更新要否を確認した。
- [ ] Legal / Security / Compliance review の要否を判断した。

## Required evidence
| Evidence | Required | Result | Link |
| --- | --- | --- | --- |
| eval result | yes / no | | |
| security review | yes / no | | |
| cost estimate | yes / no | | |
| rollback test | yes / no | | |
| approval log | yes / no | | |

## Decision
- Approve / reject / defer:
- Conditions:
- Approver:
- Review date:

# Pull request

## Summary
- 変更内容:
- Related Issue / ADR:

## Scope
- 変更した範囲:
- 変更していない範囲:

## AI use policy
- AI 利用の有無:
- AI に委任した作業:
- 入力した情報の分類:
- AI 出力の採用 / 修正 / 却下:
- 人間が最終判断した内容:

## Verification
| Check | Result | Evidence |
| --- | --- | --- |
| tests | | |
| lint / static analysis | | |
| security / privacy | | |
| eval / regression | | |
| docs / runbook | | |

## Risk / approval / rollback
- 主要リスク:
- 必要な承認:
- audit log:
- rollback:

# Runbook

## Purpose
- 対象システム:
- 対象業務:
- 利用者:

## Preconditions
- 必要な権限:
- 実行前承認:
- 入力禁止情報:
- 想定するリスク:

## Procedure
| Step | Action | Expected result | Evidence |
| --- | --- | --- | --- |
| 1 | | | |

## Verification
- 成功条件:
- 失敗条件:
- 監視指標:

## Rollback / manual takeover
- rollback 手順:
- manual takeover 手順:
- escalation:

## Audit
- 残すログ:
- 承認者:
- 保存先:

# AI incident runbook

## Scope
- 対象システム:
- 対象 AI 機能:
- 対象 tool / connector / MCP:

## Severity
- severity matrix:
- incident declaration 条件:
- escalation 条件:

## Roles
- Incident Commander:
- Technical Lead:
- Security Lead:
- Communication Lead:
- Legal / Compliance:
- Field Owner:

## Detection
- alerts:
- metrics:
- manual reports:
- audit signals:

## Containment
- kill switch:
- quarantine:
- rollback:
- manual takeover:
- degrade gracefully:

## Approval and audit
- 実行前承認:
- 事後監査:
- 必須証跡:

## Recovery and validation
- 復旧条件:
- eval / regression:
- audit completeness:

## Postmortem
- 開催条件:
- template:
- action item tracking:

# Internal update
- Incident ID:
- Severity:
- Current state:
- AI involvement:
- Containment:
- Next update:

# Customer update
- 発生している事象:
- 影響範囲:
- 実施済み対応:
- お客様にお願いしたいこと:
- 次回更新:
- 問い合わせ先:

# Executive brief
- Decision needed:
- Customer / business impact:
- Legal / audit / contract impact:
- Actions taken:
- Options and recommendation:
- Next update: