---
name: ai-vendor-assessment
description: Evaluate an AI vendor (product or API) against data-handling, security, training-data posture, licensing, regulatory fit, and operational-risk criteria — producing a recommendation, risk register, and contract checklist.
version: 1.0.0
author: VantagePoint Networks
author_url: https://www.vpnetworks.co.uk
audience: IT Managers, DPOs, Compliance Leads, Procurement, CTOs, CISOs, MSP advisors
output_format: Formatted Markdown assessment with scored matrix across 8 dimensions, risk register, contract-review checklist, alternative comparison, and recommendation memo.
license: MIT
last-reviewed: 2026-04
---

# AI Vendor Assessment

A Claude Code skill for the IT manager or DPO asked to sign off on "can we use <AI tool> with our data?" — producing a defensible answer in under a week rather than outsourcing it to a consultancy.

## 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-vendor-assessment`. Describe the vendor and intended use. Answer the clarifying questions. Receive the assessment.

## When to use this

- A team wants to adopt a new AI product and you need to evaluate it before procurement.
- Your insurer, regulator, or client has asked for a register of AI tools in use with a risk assessment for each.
- A shadow-AI tool has surfaced and you need a rapid post-hoc assessment.
- You're standing up an AI-use policy and need a baseline vendor-evaluation template.
- A vendor refresh / renewal is coming and your posture on their new AI features needs a fresh look.

## What you'll get

A single Markdown document containing:

- **Vendor & product summary**
- **Scored assessment matrix** across 8 dimensions
- **Data-flow diagram description** (where data travels, where it rests, for how long)
- **Training-data posture** (is your data used to train; can it be opted out)
- **Regulatory fit** (UK GDPR, sector-specific, territorial restrictions)
- **Risk register** with severity and mitigation
- **Contract / DPA review checklist**
- **Alternative comparison** (shortlist of equivalents)
- **Recommendation memo** (approve / approve with conditions / decline / defer)

## Clarifying questions I will ask you

1. **Vendor and product name?**
2. **What will it be used for?** (primary use case, users, data types processed)
3. **What data classes are involved?** (public, internal, confidential, restricted)
4. **Is personal data processed?** (yes — which categories / no)
5. **Is special-category personal data involved?** (health, biometric, political etc.)
6. **Regulatory / sector constraints?** (FCA, SRA, ICAEW, DSPT, PCI, HIPAA-equivalent)
7. **Deployment model?** (SaaS, private cloud, self-hosted, on-prem)
8. **Jurisdiction requirements?** (UK only, EU, US acceptable)
9. **Vendor-provided docs you have?** (privacy notice, DPA template, security white paper, SOC 2, ISO 27001 cert)
10. **Known alternatives being considered?**
11. **Who's the internal champion?** (team requesting the tool)
12. **Urgency?** (research only / near-term adoption / already live — post-hoc)

## Output template

```markdown
# AI Vendor Assessment — <vendor> / <product>

**Date:** <date> · **Prepared by:** <role> · **Sponsor:** <role>
**Status:** DRAFT / FINAL · **Recommendation:** <Approve / Approve with conditions / Decline / Defer>

## 1. Vendor & Product Summary
| Field | Value |
|---|---|
| Vendor legal name | |
| Product | |
| Version / tier | |
| Deployment | SaaS / private cloud / self-hosted |
| Primary jurisdictions | |
| Data residency options | |
| Established | <year> |
| Funding / ownership | |

**Intended use case at <firm>:** <one paragraph>
**Users in scope:** <N> (list roles)
**Data classes processed:** <public / internal / confidential / restricted>

## 2. Scored Assessment (0-5 per dimension)
| Dimension | Score | Notes |
|---|---|---|
| Data handling & residency | <> | |
| Training-data posture | <> | |
| Security & certifications | <> | |
| Privacy & DPA | <> | |
| Vendor stability | <> | |
| Regulatory fit | <> | |
| Interoperability & portability | <> | |
| Commercial terms | <> | |
| **Overall** | <avg> | |

- 0-2 = do not approve without major remediation
- 3 = approve with conditions
- 4 = approve
- 5 = exemplary

## 3. Data Flow
<Narrative of where data travels.>
- **Ingress:** <how data enters — browser, API, email, file upload>
- **At rest:** <where stored, how encrypted, for how long>
- **In transit:** <TLS versions, IP allowlisting>
- **Processing location:** <country, region, sub-processors>
- **Outbound:** <what leaves the vendor's systems, under what conditions>
- **Logs / metadata:** <what's retained as metadata vs. content>

Diagram / simple text flow:
```
User ──[TLS 1.3]──> Vendor API (<region>) ──> Model inference (<region>)
                                          ──> Log store (<region>, <N> days)
                                          ──> Training-data pipeline? (opt-in / opt-out / default-on)
```

## 4. Training-Data Posture (the critical one)
| Question | Vendor answer | Evidence | Acceptable? |
|---|---|---|---|
| Is our data used to train base models? | | <doc ref> | |
| Is our data used to train fine-tuned models? | | | |
| Can we opt out? How? | | | |
| Is opt-out default or opt-in? | | | |
| What does the vendor consider "training"? (evaluation, RLHF, improvement all count differently) | | | |
| Retention of inputs / outputs for any vendor purpose? | | | |
| Are contractor humans reviewing prompts? | | | |

## 5. Security & Certifications
| Certification / control | Status |
|---|---|
| SOC 2 Type II | <Yes — date / In progress / No> |
| ISO/IEC 27001 | |
| ISO/IEC 27701 (privacy) | |
| ISO/IEC 42001 (AI management) | |
| Pen-test cadence | <annual / on-demand / none> |
| Vulnerability disclosure programme | |
| Customer-managed keys (BYOK) | |
| SSO / SAML / SCIM | |
| Audit logs exportable | |
| Sub-processor list (public? notified on change?) | |

## 6. Privacy & Data Protection Act / UK GDPR Fit
| Aspect | Finding |
|---|---|
| Vendor is Data Processor for your content | <Yes / No / Mixed> |
| DPA available | <Yes — acceptable / Yes — needs negotiation / No template, vendor-drafted> |
| International transfer mechanism | <UK-IDTA, UK addendum, adequate-third-country, BCRs, SCCs> |
| Retention (content, metadata, logs) | |
| Data-subject-rights handling | |
| Breach notification SLA | <N hours> |
| DPIA requirement triggered? | <Yes / No / Already conducted> |

## 7. Regulatory / Sector Fit
- **FCA:** <applicable? operational-resilience impact if vendor fails, concentration risk>
- **ICO:** <registration, data protection implications>
- **SRA (if legal):** <privilege preservation, client confidentiality>
- **ICAEW (if accountancy):** <professional-standards alignment>
- **Cyber Essentials / ISO 27001 scope:** <does the vendor fall within your scope? what evidence do you need from them?>
- **Export controls / sanctions:** <any vendor operations in restricted jurisdictions?>

## 8. Risk Register
| # | Risk | Severity | Likelihood | Mitigation | Owner |
|---|---|---|---|---|---|
| R1 | Training-data opt-out not documented | High | Low | Get written confirmation before contract | DPO |
| R2 | Sub-processor in non-adequate jurisdiction | Med | Med | Require prior-notification clause | DPO |
| R3 | No customer-managed keys | Med | — | Accept or defer purchase | CTO |
| R4 | Vendor may change training posture at will | Med | Med | Annual re-review; contract audit right | Procurement |
| R5 | Token-usage billing could spike on misuse | Low | Med | Set spend cap + alerting | Finance |

## 9. Contract / DPA Review Checklist
- [ ] DPA in place, UK-IDTA or addendum if international transfers
- [ ] Processing scope and purpose limitation clear
- [ ] Sub-processor list current + change-notification clause
- [ ] Training-data use explicitly addressed (not absence of language)
- [ ] Retention periods for content / metadata / logs
- [ ] Deletion / portability on termination
- [ ] Security-measures annex (Art. 32)
- [ ] Breach-notification SLA (<N> hours)
- [ ] Audit right (reasonable, proportionate)
- [ ] Liability cap acceptable
- [ ] Termination for convenience available
- [ ] IP ownership of prompts and outputs clear
- [ ] Confidentiality mutual
- [ ] Governing law (England & Wales typical)

## 10. Alternative Comparison
| Option | Data residency | Training opt-out | Certifications | Commercial | Verdict |
|---|---|---|---|---|---|
| <Vendor A — this> | | | | | |
| <Vendor B> | | | | | |
| <Vendor C> | | | | | |
| Self-host equivalent | | | | | |

## 11. Recommendation Memo
<One-paragraph clean recommendation.>

**Decision requested of:** <role>
**Recommended:** <Approve / Approve with conditions / Decline / Defer>
**Conditions (if approve-with):**
1. <condition>
2. <condition>
**Blockers (if decline):**
1. <blocker>
2. <blocker>
**Target revisit date (if defer):** <date>

## 12. If Approved — Deployment Conditions
- [ ] Acknowledged by <role>
- [ ] AI Use Policy updated to include this tool
- [ ] Tool added to SSO / audit logging
- [ ] Owner named; monthly usage review scheduled
- [ ] First-quarter review date set
```

## Example invocation

**User:** "Sales team wants to use <Vendor X> for AI-powered prospecting — it ingests our CRM notes and drafts outreach emails. We're an FCA-authorised firm. Vendor offers SaaS only, US-headquartered, no UK data residency option."

**What the skill will do:**
1. Ask the 12 questions, pressing on: what CRM data is ingested (personal data of prospects = yes), whether there's a US transfer mechanism, whether the vendor uses customer data to train, FCA operational-resilience implications.
2. Produce the assessment likely scoring 2-3/5 overall: data residency fails for FCA sensitivity (no UK option), training-data posture unclear without vendor confirmation, no CMK, limited certifications.
3. Risk register flags: sub-processor exposure, international transfer compliance, concentration risk (SaaS only), training-data default.
4. Recommendation likely **Decline** or **Defer** until vendor answers 3-4 specific written questions in the contract.
5. Alternative comparison suggests: Microsoft Dynamics + Copilot (UK region, DPA stable) or HubSpot AI (check data residency tier), or a self-hosted approach if the use case is narrow.

## Notes for the requester

- **The training-data question is the one that kills most AI vendor assessments.** Many vendors have evasive language ("we may use aggregated, anonymised data"). Insist on specifics in writing.
- **"SOC 2 Type II" is not a free pass.** It tells you about controls, not about privacy or AI-specific posture. ISO 42001 (AI management) is the emerging differentiator.
- **Check the sub-processor list, not just the vendor.** OpenAI, Anthropic, Azure OpenAI — each has different privacy postures. "Built on OpenAI" may change your assessment materially.
- **Decline is a valid recommendation.** "Come back when you have UK residency" is kinder than six months of wrangling with a tool that was never going to work.
- **Good looks like:** a signed memo in under a week, a Yes/No/Conditional decision with reasons, a contract-review checklist the sponsor can hand to legal, and a reviewable register entry for next year's re-assessment.

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