How Should Businesses Evaluate AI Vendors?

When evaluating an AI vendor, my recommendation is to treat the assessment as a combination of security review, privacy review, operational risk review, and vendor management review, not simply a technology purchase. Many organizations now use AI-specific due diligence questionnaires that extend traditional cloud and SaaS vendor assessments.

1. Data Handling and Privacy

These are usually the first questions procurement, legal, and security teams should ask:

Data Usage

• Is customer data used to train foundation models or improve services?

• Can model training be disabled contractually and technically?

• Are prompts, uploaded files, embeddings, and outputs segregated from training datasets?

• Are opt-in and opt-out controls available?

Data Residency

• Where is data processed?

• Where is data stored?

• Can data remain within specific countries or regions?

• Can the vendor guarantee regional processing?

Data Retention

• How long are prompts retained?

• How long are outputs retained?

• How long are logs and telemetry retained?

• Can retention periods be customized?

• Is immediate deletion available?

Data Ownership

• Who owns prompts, outputs, uploaded content, generated content, and fine-tuned models?

• What rights does the vendor retain after processing?

2. Security Architecture

AI systems can introduce new attack surfaces beyond traditional SaaS applications.

Core Security Controls

• Is data encrypted in transit and at rest?

• Which encryption standards are used?

• How are encryption keys managed?

• Is customer-managed encryption supported?

Access Control

• Does the platform support:

  • SSO

  • SAML

  • OAuth

  • MFA

  • RBAC

  • Just-in-time access

Operational Security

• How is privileged access controlled?

• Are employee activities logged and monitored?

• Is production access restricted?

• Is access reviewed regularly?

Security Assurance

Request evidence such as:

• SOC 2 Type II reports

• ISO 27001 certification

• ISO 27701 certification

• Penetration test summaries

• Vulnerability management processes

• Security whitepapers

3. AI-Specific Security Risks

Traditional vendor reviews often overlook these AI-specific concerns.

Model Security

• How is prompt injection mitigated?

• How is jailbreak resistance tested?

• What protections exist against model manipulation?

• How are hallucination risks addressed?

Output Controls

• Are content safety filters implemented?

• Can organizations configure safety policies?

• Is high-risk content detected and blocked?

Data Leakage Prevention

• Can users retrieve information from other tenants?

• How is tenant isolation enforced?

• What safeguards prevent inadvertent disclosure through generated outputs?

4. Third-Party and Subprocessor Risks

Many AI vendors rely on multiple supporting services.

Key questions include:

• Who are all subprocessors?

• What data does each subprocessor receive?

• Are subprocessors disclosed publicly?

• How are new subprocessors approved?

• How are customers notified of changes?

Additionally ask:

• Can the vendor provide a current subprocessor inventory?

• What diligence is performed on subprocessors?

• Do subprocessors meet equivalent security standards?

5. Transparency and Documentation

Transparency is critical for informed risk decisions.

Ask whether the vendor can provide:

• Architecture documentation

• Security documentation

• Privacy documentation

• Data flow diagrams

• Model cards

• Responsible AI documentation

• Incident response procedures

• Change management policies

Useful questions include:

• How are model updates communicated?

• Can customers defer major changes?

• How are breaking changes handled?

• How is model performance monitored over time?

6. Compliance and Regulatory Support

Organizations operating in regulated industries should review:

Privacy Regulations

• GDPR

• CCPA/CPRA

• HIPAA

• Local data protection laws

Industry Requirements

• Banking regulations

• Healthcare requirements

• Government procurement standards

• Critical infrastructure obligations

Ask:

• Is a DPA available?

• Are SCCs available for international data transfer?

• Can the vendor support regulatory audits?

7. Human Access to Customer Data

This area is often underestimated.

Questions should include:

• Can vendor employees access customer prompts?

• Under what circumstances?

• Is customer approval required?

• Is access logged?

• Is access time-limited?

• Can organizations restrict human review entirely?

8. Deployment Options

Different risk profiles require different deployment models.

Public Cloud AI

Suitable when:

• Strong contractual controls exist

• Data classification permits cloud processing

• Security requirements can be met

Private or Dedicated Environments

Consider:

• Single-tenant deployments

• Dedicated compute

• Private networking

• Virtual private cloud configurations

On-Premises or Self-Hosted

For highly sensitive workloads, ask:

• Is self-hosting supported?

• Can models run in isolated environments?

• What operational responsibilities remain with the customer?

9. Incident Response and Resilience

Important questions include:

Security Incidents

• How quickly are customers notified?

• What constitutes a reportable incident?

• What are notification timelines?

Business Continuity

• Are backups maintained?

• What are the recovery objectives?

• What service availability commitments exist?

• Is disaster recovery regularly tested?

10. Vendor Governance and Long-Term Viability

AI adoption is not a one-time procurement decision.

Review:

• Vendor financial stability

• Product roadmap

• Security governance program

• Responsible AI governance

• Regulatory monitoring

• Independent audits

Ask:

• Who is accountable for AI governance?

• Is there a responsible AI committee?

• How are emerging risks assessed?

• How are customers informed of material changes?