90 Days Gen AI Risk Trial -Start Now
Book a demo
Free TemplateVendor Assessment

AI Vendor Evaluation Scorecard

A weighted 100-point scoring framework for evaluating AI vendors across 5 security domains. Score each vendor objectively and make defensible procurement decisions.

Updated March 2026 · 5 scored domains · GDPR Article 28, ISO 27001, SOC 2 aligned

100 pts
weighted scoring system
5 domains
security coverage
17 criteria
evaluation checkpoints
Free
to use and customise

Why AI Vendors Need Specialist Evaluation

Standard vendor security questionnaires were designed for SaaS and cloud infrastructure — not for AI. AI vendors introduce unique risks around data training, model inference, and AI-specific attack vectors that traditional vendor assessments miss entirely. Using this scorecard alongside your standard TPRM process ensures AI vendors are evaluated on the criteria that matter most.

63%
of enterprises use AI vendors without a DPA
Most organisations have not ensured GDPR-compliant data processing agreements are in place with their AI vendors, creating significant regulatory exposure.
41%
of AI vendors train on customer data by default
Without an explicit no-training clause, the default position of many AI services is to use customer prompts and data to improve their models.
78%
of vendor assessments miss AI-specific security criteria
Standard TPRM questionnaires do not cover prompt injection, model access controls, or AI output security — leaving critical gaps in vendor risk assessments.
3.5x
higher breach cost when AI vendor controls are absent
Data breaches originating from AI vendor weaknesses carry higher costs than average breaches due to the volume and sensitivity of data processed by AI pipelines.

The Vendor Evaluation Scorecard

Click each domain to expand the scoring criteria. Use the scoring guidance to allocate points and document justifications for each criterion.

Data security is weighted highest because AI vendors process your organisation's data in the model inference pipeline. Weaknesses here directly translate to data exposure risk.

Data encryption at rest and in transit8 pts

Verify AES-256 encryption at rest and TLS 1.2+ in transit. Request documentation showing encryption covers training pipelines, not just API transport. Score 8 if both verified with evidence, 4 if self-attested only, 0 if not confirmed.

No-training guarantee in contract8 pts

Contractual commitment that customer data (prompts, documents, outputs) is not used to train or fine-tune the vendor's models. This must be a contractual term, not a marketing claim. Score 8 if in writing in the DPA/contract, 0 if only in marketing materials or not available.

Data retention and deletion policy6 pts

Documented retention periods for all data categories processed (prompts, outputs, logs). Ability to request deletion and confirmation of deletion within a defined SLA. Score 6 if fully documented with deletion capability, 3 if partial, 0 if no documentation.

Customer data isolation5 pts

Verification that data from different customers is logically or physically isolated — particularly in multi-tenant inference environments. Score 5 if documented isolation architecture provided, 3 if attested, 0 if not addressed.

Data residency options3 pts

Ability to specify data processing location (e.g. EU-only for GDPR compliance). Score 3 if configurable with contractual commitment, 1 if regional deployment available but not contractual, 0 if no control.

How to Run a Vendor Evaluation

A scorecard produces useful outputs only when applied consistently. Follow these five steps to run an evaluation that produces defensible, comparable results across vendors.

1
Define the evaluation scope and data sensitivity
Before scoring, document what data the AI vendor will process (including data classification levels), what the use case is, and whether a DPA will be required. Higher data sensitivity should raise the minimum score threshold for approval.
2
Send the scorecard as a structured vendor questionnaire
Distribute the relevant sections to each vendor shortlisted. Request supporting evidence for all claims — SOC 2 Type II reports, ISO 27001 certificates, penetration test executive summaries, and draft DPAs. Set a clear response deadline.
3
Score each vendor independently before reconciling
Have two evaluators score each vendor independently against each criterion, then compare scores and resolve discrepancies. Independent scoring reduces bias and produces more defensible procurement decisions.
4
Verify high-stakes criteria with a technical review call
Do not rely solely on vendor self-attestation for critical criteria such as no-training guarantees, prompt injection controls, and incident notification terms. Schedule a 60-minute technical security review call with the vendor's security team to verify key claims.
5
Document final scores, decisions, and re-assessment dates
Record the final score for each vendor with scoring justifications. Document the approval decision, any conditional requirements, and the date for re-assessment (typically 12 months or when a material change occurs). Store documentation in your vendor risk management system.

Frequently Asked Questions

Monitor Your AI Vendors Continuously After Approval

A point-in-time vendor assessment captures your vendors' security posture on one day. Aona monitors all AI vendor usage across your organisation continuously — tracking which vendors employees are using, what data is being shared, and whether approved vendor controls are actually in effect.

Book a Demo