Enterprise + Sovereign AI

What should an AI vendor risk assessment cover?

An AI vendor risk assessment should cover seven domains: security controls and certifications, data handling and residency, the model and subprocessor supply chain, availability and resilience, contractual protections, financial and operational stability, and exit portability. AI adds two risks traditional assessments miss — upstream model dependencies and training-data governance — so score those explicitly rather than burying them in a general questionnaire.

Key facts

Security controlsSSO/SCIM, RBAC, encryption with BYOK, SIEM audit export
ResidencyRegion-locked planes: EU, GCC, APAC, US; verify store-level mapping
SubprocessorsPublished list with change notification in the DPA
Deployment modelsCloud, VPC, on-prem, air-gapped — reduces shared-tenancy risk
AvailabilityAutomatic upstream failover; 99.9% uptime on paid plans; 4-hour Enterprise support SLA
Compliance postureSOC 2 Type II and ISO 27001 readiness in progress (not yet certified); GDPR and PDPL alignment
Contract termsTerms at /legal/terms; DPA with audit rights, deletion and breach notice
StatusLive API with 30+ models; component health published

TL;DR

  • Score upstream model dependencies separately — they are a distinct concentration risk.
  • Require a current subprocessor list with locations and change-notification terms.
  • Verify residency at store level, including logs, vectors and backups.
  • Test availability claims against the SLA and the status page history.
  • Assess exit before signing: export formats, deletion evidence and transition help.

How it works, step by step

  1. Define the risk domains and a scoring rubric before sending the questionnaire.
  2. Collect evidence: control descriptions, region maps, subprocessor list, DPA, SLA, status history.
  3. Assess the model supply chain: which upstreams serve each model, and what happens on deprecation.
  4. Review data governance: logging defaults, training use, retention and deletion.
  5. Evaluate resilience: failover design, region strategy, capacity and incident process.
  6. Score residual risk and document mitigations or acceptance decisions.
  7. Schedule periodic reassessment and trigger reviews on material changes.
1Define the riskdomains and ascoring rubric2Collect evidence:controldescriptions,3Assess the modelsupply chain: whichupstreams serve4Review datagovernance: loggingdefaults, training5Evaluateresilience:failover design,6Score residual riskand documentmitigations or

Original data

Automatic upstAvailabilitySOC 2 Type II Compliance postureLive API with StatusSource: Plugsky facts table · updated 2026-09-25

Try it yourself

Open the sovereign AI readiness score →

The seven domains

  • Security — encryption, access control, vulnerability management, penetration testing.
  • Data — classification handling, residency, retention, training exclusion, deletion.
  • Supply chain — subprocessors and upstream model providers, with locations and change terms.
  • Availability — architecture, failover, SLA, incident history.
  • Contract — liability, indemnity, audit rights, breach notice, exit assistance.
  • Vendor stability — financial health, key-person risk, roadmap coherence.
  • Exit — portability, formats, deletion certificates and knowledge transfer.

AI-specific risks to score explicitly

Two risks are unique to AI vendors. Model dependency: the model you rely on may be served by an upstream provider that changes terms, retires a version or alters routing. Ask for the fallback chain and deprecation notice period. Data reuse: confirm whether prompts, completions or embeddings can influence training, and how that is enforced and audited. Plugsky documents automatic failover and a DPA covering subprocessors and deletion, and publishes a model catalogue with 30+ models; validate the specifics that matter to your risk committee.

Scoring and residual risk

Score each domain on evidence strength, not vendor confidence. A control description with sample outputs is stronger than a policy statement. Record residual risk after mitigations, assign an owner, and set review dates. Where certifications are in progress — as with Plugsky's SOC 2 and ISO 27001 readiness — attach a milestone and monitor it, or compensate with private deployment and BYOK.

Ongoing monitoring

  • Subscribe to subprocessor change notifications and status updates.
  • Review audit evidence at least annually.
  • Track model deprecations affecting your workloads.
  • Re-test deletion and access reviews on a schedule.
  • Re-score when the vendor changes regions, models or ownership.

Honest comparison

CapabilityPlugskyHyperscaler AI platformBuilding in-house
Subprocessor transparencyPublished list with change notificationPer-service subprocessor pagesYou vet each supplier
Model supply chainFallback chains across 30+ modelsVendor-controlled catalogueYou manage every model
Residency evidenceRegion-locked planes and workspace selectionRegion configuration per serviceYour design
Audit exportSIEM connectorsNative cloud auditCustom pipelines
Deployment isolationVPC, on-prem, air-gapped optionsShared cloud by defaultYou own the stack
CertificationsSOC 2 / ISO 27001 readiness in progressCompleted audits in many regionsYour own programme

Frequently asked questions

What makes AI vendor risk different from SaaS risk?

Two factors: upstream model dependencies you do not contract with directly, and data reuse questions around prompts, completions and embeddings. Both need explicit assessment.

How do we assess upstream model providers?

Ask for the mapping of models to upstream providers, their locations, the fallback chain, and the deprecation notice process. Treat concentration on a single upstream as a scored risk.

What evidence should replace a missing certificate?

Control descriptions, penetration-test summaries, architecture diagrams, configuration exports, audit samples and contractual remedies. Document residual risk after mitigations.

How often should we reassess?

At least annually, and on material change: new regions, new subprocessors, model deprecations, ownership changes or significant incidents.

Can we reduce risk with deployment choice?

Yes. Private endpoints, on-prem and air-gapped tiers remove shared-tenancy exposure and give you control over data and keys.

What exit clauses matter most?

Export formats for data and prompts, transition assistance, deletion certificates, and continued access during migration. Negotiate them before signature.

Where can we verify availability claims?

Compare the SLA at /legal/sla with the incident history shown on /status and with your own monitoring during a pilot.