vendor-ai-review

name	vendor-ai-review
description	Review AI vendor contracts (LLM API providers, ML libraries, AI services). Включает SaaS-MSA общие issues (см. contract-law/saas-msa-review) + AI-specific: training data rights, output ownership, model warranty, hallucination liability, rate limits, API stability, безопасность (запрет training на our data), exit strategy + portability. Post-2022 sanctions specifics.
argument-hint	[path к AI vendor contract]
user_invocable	true
ported_from	ai-governance-legal/vendor-ai-review
ported_at	"2026-05-19T00:00:00.000Z"
adaptation_category	B
last_legislative_update	2026-05

/vendor-ai-review

Назначение

Review AI vendor / API / library contracts. Differs from generic SaaS-MSA review (см. contract-law/saas-msa-review) — focus на AI-specific clauses.

Pre-flight

• Vendor identified: [имя + type: hyperscaler LLM API / Russian LLM / open-source / SDK / library]
• Sanctions exposure check: US/EU vendor with restricted РФ access? (post-2022)
• Use case clear: [для каких задач будем использовать AI]
• Data sensitivity: [какие данные будем sending к vendor]

AI-specific contract review points

A. Data rights — наши данные у vendor

Training on our data

Critical: Vendor может использовать наш input / output для training future models?

• Best: "Vendor shall not use Customer Data for training models" — explicit
• OK: Opt-out available + opt-out exercised
• Problematic: Default training on customer data with no opt-out
• Bad: Training explicit + irrevocable license к vendor

Examples in industry:

• OpenAI Enterprise: Не trains on data (default for enterprise)
• OpenAI consumer ChatGPT: Trains by default (opt-out exists)
• Anthropic Claude (API): Не trains by default
• Google Bard / Gemini Enterprise: Не trains by default
• YandexGPT API: Check current terms (varies)
• GigaChat: Check current terms

→ Push for explicit "no training" в SaaS-MSA.

Data retention

• Как долго vendor stores наши queries / responses?
• Recommended: short retention (30 days max)
• Logs минимум для дoзinformation troubleshooting

Data location

• Где stored (geographic)?
• 152-ФЗ ст.18 ч.5 (локализация — для ПДн РФ residents) — нужно processing в РФ
• Cross-border — уведомление РКН

B. Output rights — кому принадлежит result

Generated content ownership

Best: "Customer owns all rights в outputs" — explicit assignment

Common ambiguity: Outputs могут быть subject к terms ограничивающим use, even if "owned".

Restrictions на output use

• Можно ли использовать commercially?
• Можно ли redistribute?
• Можно ли train other models на outputs?
• Attribution required?

C. Model warranty (или disclaimer)

Что vendor warrants?

Common reality (almost все):

• AS-IS basis
• No warranty of accuracy / completeness
• No warranty of fitness for particular purpose
• Disclaimer hallucinations / errors

Что мы хотим:

• Reasonable level of accuracy для intended use
• Notification если material model changes
• Removal of harmful outputs (если known)

→ Negotiate carve-outs из broad disclaimers для material reliance scenarios.

D. Hallucination / error liability

Default: Vendor имеет broad disclaimer + cap on damages.

Risk shift:

• Customer absorbs full liability для use cases dependent on AI accuracy
• Customer must verify outputs before reliance
• High-stakes decisions (medical, legal, financial) — extra caution

E. Rate limits + API stability

• Rate limits: Что happens если exceeded? Soft cap / hard cap?
• API stability: Versioning policy — backwards compatible breaking changes?
• Deprecation notice: Sufficient lead time для migration?
• Uptime SLA: Realistic + remedies для downtime?

F. Pricing model

• Token-based (typical for LLM):
  • Input tokens / output tokens — separate rates
  • Predictability — high usage может escalate quickly
• Per-API-call: more predictable
• Subscription: flat fee — but cap usage
• Mixed: combo

→ Negotiate cap на monthly costs / requests.

G. Security + confidentiality

• Запрет on training: see (A) above
• Доступ vendor staff к customer data — restricted?
• Encryption: transit + at-rest
• Audit rights: can we audit security practices?
• Breach notification: how fast?

H. Compliance с 152-ФЗ (для ПДн)

Если будем sending ПДн в vendor:

• Vendor — accepts role как processor (DPA от data-protection pack)
• Локализация — vendor processes в РФ или есть legal basis для cross-border
• Уведомление РКН подано?
• Vendor agrees к РКН аудиту?

I. Exit / portability

• Termination rights: какие events trigger? Срок уведомления?
• Data export: при выходе можем export все наши data?
• Format: machine-readable?
• Fine-tuned models: ownership / portability?
• Transition support: vendor cooperates в migration?

J. Post-2022 sanctions specifics

Если vendor — US/EU

• Access restrictions — мы можем technically not access (geofencing)
• Payment restrictions — банк может block transactions
• Sanction-driven termination — vendor may unilaterally terminate
• No guarantee of continued availability

→ For material reliance — risky. Either backup vendor (Russian) или self-hosted (open-source).

Если vendor — Russian

• Generally OK от sanctions perspective
• Quality / capability — variable (YandexGPT / GigaChat reasonable for Russian content; might fall behind US frontier models for specialized tasks)
• Vendor lock-in risk reduced

Self-hosted open-source

• Llama / Mistral / Yi / Qwen — generally accessible
• Operational cost higher
• Quality control полностью our (good and bad)

K. AI-specific compliance representations

Vendor представляет:

• Compliance с applicable laws (152-ФЗ если processing ПДн)
• Training data licensed properly (для generative AI — copyright concerns)
• Не trained on illegal / sensitive data в нарушение
• No backdoors / surveillance built-in (для post-2022 РФ vendors — particularly relevant)

Workflow

Шаг 1. Identify vendor + scope

## Vendor

- **Name:** [...]
- **Type:** [LLM API / ML SaaS / library SDK / иное]
- **Country:** [...]
- **Sanctions exposure:** [Y/N]
- **Use case:** [для чего будем использовать]
- **Volume estimated:** [requests / data]
- **Data sensitivity:** [ПДн / sensitive / public]

Шаг 2. Apply checklist (A-K above)

Для каждого section — review contract language + flag risks.

Шаг 3. Risk summary

## Risk findings

### 🔴 Critical

1. [Vendor trains on customer data by default — must change to opt-out]
2. [Sanctions exposure без backup — material risk]
3. ...

### 🟡 Medium

1. [Rate limits not transparent]
2. [Indemnity caps too low]

### 🟢 OK

[List]

Шаг 4. Recommended changes

## Recommended contract changes

### Required (won't sign without)

- [ ] "Vendor shall not use Customer Data for training" — added
- [ ] Customer owns all rights в outputs — added
- [ ] Backup arrangement или termination right if material API change

### Strongly desired

- [ ] Reduce data retention к 30 days
- [ ] Increase indemnity cap к [X]
- [ ] Audit rights expanded

### Nice to have

- [ ] Negotiate volume discount
- [ ] Add SLA для new model versions

Шаг 5. Output

# AI Vendor Review — [vendor + product]

## Bottom line

- **Sign / Don't sign / Sign with changes:** [...]
- **Critical issues:** [count]

## Issues detail

[From Step 3]

## Required changes

[From Step 4]

## Alternative vendors considered

- [Russian alternative 1]: pros / cons
- [Open-source alternative]: pros / cons

## Strategic recommendation

[Use this vendor / Use alternative / Hybrid (this for X, alternative for Y)]

Что НЕ делает

• Не negotiates contract (это business + outside-адв)
• Не делает security audit vendor (это infosec team)
• Не replaces outside-адв legal opinion для high-stakes deals

Attribution

Adapted from ai-governance-legal/vendor-ai-review by Anthropic (Apache 2.0).

Категория B:

• Universal AI vendor concerns preserved
• 152-ФЗ ПДн processor obligations — РФ-specific
• Post-2022 sanctions overlay — РФ-specific risk
• Russian vendor alternatives (YandexGPT / GigaChat) — РФ-specific

Original copyright: © 2026 Anthropic PBC, Apache 2.0.

---

⚠ Юридический disclaimer

Данный skill — техническая платформа, не оказывает юридических услуг по ст.2 ФЗ-63 «Об адвокатской деятельности». Outputs не заменяют консультацию лицензированного юриста / адвоката ФПА.

AI может галлюцинировать, выдавать устаревшие нормы, неверно интерпретировать факты. Material decisions (увольнение, M&A, налоговый спор, IP litigation) — обязательно engage outside-адв ФПА с релевантной специализацией.

Проект ru-legal и его contributors не несут ответственности за решения, принятые на основе outputs системы. Использование — на свой риск.