---
name: ai-system-inventory-builder
description: Build an AI system inventory (AI BOM / model register) for an SMB — every AI system in use, its owner, risk tier, data classes, vendor/model, human-oversight mechanism and impact-assessment status — the foundational artefact ISO/IEC 42001 and the EU AI Act both require.
version: 1.0.0
author: VantagePoint Networks
author_url: https://www.vpnetworks.co.uk
audience: IT Managers, AI/Data Governance Leads, CISOs, DPOs, Compliance Leads, MSP Consultants
output_format: Formatted Markdown register with a cover note, the inventory table, per-system detail cards, a risk-tier summary, a shadow-AI findings section, and a maintenance plan — ready to drop into an ISO 42001 AIMS or an EU AI Act compliance file.
license: MIT
last-reviewed: 2026-10
---

# AI System Inventory Builder

A Claude Code skill that turns "we think we use a few AI tools somewhere" into an auditable **AI system inventory** — the asset register that every AI-governance conversation now opens with. It's the first artefact an ISO/IEC 42001 auditor asks for, the Article-level evidence the EU AI Act expects, and the thing almost no SMB actually has.

## How to use this skill

1. Download this `SKILL.md` file.
2. Place it in `~/.claude/commands/` (macOS/Linux) or `%USERPROFILE%\.claude\commands\` (Windows).
3. In Claude Code, run `/ai-system-inventory-builder`. Describe what your organisation does and what AI you know about in plain English. Answer the clarifying questions. Receive a complete, tiered register.

## When to use this

- An auditor, client, or insurer has asked for your "AI system inventory" / "model register" / "AI bill of materials" and you have nothing to hand.
- You're starting an ISO/IEC 42001 programme and need the Clause 8 / Annex A.4 inventory before anything else makes sense.
- You're scoping EU AI Act exposure and need to know which systems fall into which risk tier.
- You suspect **shadow AI** — staff using tools nobody approved — and want a defensible first pass at what's actually in use.
- You're integrating an AI governance view into an existing ISO 27001 asset register and want them to line up.

## What you'll get

A single Markdown document containing:

- A **cover note** — scope, owner, date, review cadence, the frameworks this maps to
- The **inventory table** — one row per AI system, with the fields auditors expect
- **Per-system detail cards** for anything risk-tiered Medium or above
- A **risk-tier summary** mapped to EU AI Act tiers (prohibited / high / limited / minimal) and your own scheme
- A **shadow-AI findings** section — likely unsanctioned tools, inferred from the sectors and workflows you describe, with a discovery checklist
- A **gaps & next actions** list (missing owners, overdue impact assessments, unassessed vendors)
- A **maintenance plan** — how to keep it current, and the triggers that force an update

## Fields captured per AI system

| Field | Why it matters |
|---|---|
| System name & business purpose | Identity and intended use |
| Owner (named role) | Accountability — the single most-missed field |
| Organisational role | Provider / developer / deployer / user — drives which obligations apply |
| Vendor & model/version | Supply-chain and change tracking |
| Risk tier | Prioritises governance effort; maps to EU AI Act |
| Data classes consumed | Public / internal / confidential / special-category; lawful basis |
| Human oversight mechanism | Required for consequential decisions |
| Last impact assessment (AIIA) date | ISO 42001 A.5 evidence |
| Deployment status | Pilot / production / retired |
| Notes | Disclosures, DPAs, known limitations |

## Clarifying questions I will ask you

1. **What does your organisation do, and roughly how many staff?**
2. **Which AI tools do you *know* you use?** (Copilot, ChatGPT/Enterprise, Claude, Gemini, Perplexity, transcription tools, sector-specific AI, AI features inside existing SaaS)
3. **Do you build, fine-tune, or host any models yourself**, or only consume vendor AI? (sets your role)
4. **What data do these tools touch?** (general, PII, client-privileged, financial, special-category health)
5. **Any AI making or materially influencing decisions about people?** (hiring, credit, access, eligibility) — these tier high
6. **Regulator(s) / frameworks in play?** (ICO, FCA, ISO 42001 target, EU AI Act exposure, sector body)
7. **Do you sell into the EU**, or process EU residents' data? (EU AI Act applicability)
8. **Known shadow-AI concerns?** (teams you suspect are using unapproved tools)
9. **Existing registers to align with?** (ISO 27001 asset register, data map/RoPA, CMDB)
10. **Who owns this inventory, and how often will it be reviewed?** (quarterly recommended)

## Output template

```markdown
# AI System Inventory — <organisation>

**Version:** 1.0 · **Date:** <date> · **Owner:** <role> · **Review cadence:** quarterly
**Maps to:** ISO/IEC 42001 (Clause 8, Annex A.4/A.8) · EU AI Act risk tiering · ISO 27001 asset register

## 1. Scope & method
<What's in scope, how systems were identified, what "AI system" means here.>

## 2. Inventory
| # | System | Purpose | Owner | Role | Vendor/Model | Risk tier | Data classes | Human oversight | Last AIIA | Status |
|---|--------|---------|-------|------|--------------|-----------|--------------|-----------------|-----------|--------|
| 1 | M365 Copilot | Drafting, summarising | IT Manager | Deployer | Microsoft | Limited | Internal, Confidential | Review before send | 2026-09 | Production |
| 2 | <…> | | | | | | | | | |

## 3. Risk-tier summary
| Tier (EU AI Act) | Count | Systems | Governance priority |
|---|---|---|---|
| Prohibited | 0 | — | n/a (must not deploy) |
| High | <n> | <…> | AIIA + oversight + logging mandatory |
| Limited | <n> | <…> | Transparency obligations |
| Minimal | <n> | <…> | Baseline policy only |

## 4. Per-system detail (Medium+ risk)
### <System name>
- **Purpose / intended use:** …
- **Why this tier:** …
- **Data & lawful basis:** …
- **Human oversight:** …
- **Open actions:** …

## 5. Shadow-AI findings
Tools likely in use but not formally sanctioned, inferred from your workflows:
- <e.g. staff using personal ChatGPT for drafting>
Discovery checklist: review SSO/OAuth grants, egress/proxy logs for AI domains, expense claims for AI subscriptions, browser-extension inventory, and a short anonymous staff survey.

## 6. Gaps & next actions
- [ ] <System> has no named owner
- [ ] <System> impact assessment overdue
- [ ] <Vendor> training-data posture not assessed (run /ai-vendor-assessment)

## 7. Maintenance plan
- Reviewed quarterly by <owner>.
- Update triggers: new tool adopted, model/version change, new data class, pilot→production, new affected population, or an AI incident.
- Change log appended below.
```

## Example invocation

**User:** "30-person independent brokerage, FCA-regulated. We use M365 Copilot, a couple of people use ChatGPT, and our CRM just added an 'AI assistant'. No idea if anything's high-risk. Targeting ISO 42001 next year."

**What the skill will do:**
1. Ask the 10 questions, drilling on: whether any AI touches client financial data or influences advice/eligibility (FCA sensitivity), EU exposure, and the CRM vendor's AI feature.
2. Produce the inventory with Copilot and the CRM assistant as deployer-role systems, ChatGPT flagged as likely shadow-AI pending an approved-tool decision, each risk-tiered.
3. Flag the CRM "AI assistant" as needing an `/ai-vendor-assessment` (unknown training-data posture) and an AIIA before it touches advice workflows.
4. Add a shadow-AI discovery checklist tuned to a small FCA firm, and a gaps list ready to become the first month of the ISO 42001 programme.

## Notes for the requester

- **The owner field is the one that fails audits.** "The business" is not an owner. Name a role.
- **Shadow AI is the real inventory.** The tools you approved are the easy half; budget time to actually look (OAuth grants, proxy logs, expenses) rather than trusting the approved list.
- **Tier by consequence, not by hype.** A chatbot that drafts marketing copy is minimal-risk; a model that screens CVs or scores applicants is high-risk even if it feels mundane.
- **Keep it alive.** An inventory is a quarterly habit, not a one-off document. A register last touched 18 months ago is worse than none — it reads as governance theatre.
- **Good looks like:** when a client questionnaire asks "list the AI systems you use and how you govern them", you attach one current PDF and move on.

---

### Related: runtime evidence for AI agents

An inventory tells you *what* AI you run. For autonomous or semi-autonomous **agents** acting in production, governance also needs *what each agent actually did* — a tamper-evident, exportable runtime record you can hand to an auditor or regulator. That runtime-assurance layer for AI agents is what **[Agent Audit](https://agentaudit.co.uk)** addresses, and it pairs naturally with this inventory under ISO 42001 A.6/A.9. See <https://agentaudit.co.uk>.

---
*VantagePoint Networks · <https://www.vpnetworks.co.uk> · Authored by Hak · Free under the MIT licence*