AI Vendor Risk Assessment Template: A Scorecard for Vetting AI Vendors

An AI vendor risk assessment is a structured review of an AI tool before you buy it. It checks how the vendor handles your data, how secure they are, and whether they meet the rules that apply to your business.

This page gives you the full scorecard, not a checklist stub. You get eight scoring categories, a weighting method, a red-flag list, and the exact questions to send a vendor. Score each category from 0 to 5, apply the weights, and you get one number you can compare across vendors. The editable version is in the callout.

How the Scorecard Works

Score each of the eight categories from 0 to 5, multiply by the category weight, then add the weighted scores for a total out of 100.

A 0 means the vendor cannot answer or fails the check. A 5 means strong, documented evidence. Weights let the categories that matter most to your risk count for more.

• 0 = no answer or a clear failure.
• 1–2 = weak, vague, or verbal-only claims.
• 3 = adequate, with some documentation.
• 4–5 = strong, with certificates, contracts, or written evidence.
• Total score: 80–100 low risk, 60–79 medium risk, below 60 high risk (proceed only with mitigations).

Category Weights

Each category carries a weight so your total reflects real risk, not just a raw average.

These weights are a sensible default. Adjust them for your industry, then keep them consistent so vendors are compared fairly.

• Data handling and training use — weight 5 (highest).
• Security certifications (SOC 2, ISO 27001) — weight 4.
• Access controls and authentication — weight 3.
• Sub-processors and data location — weight 3.
• Model provenance and transparency — weight 2.
• Uptime and SLAs — weight 1.
• Compliance (EU AI Act, GDPR) — weight 4.
• Exit and data portability — weight 2. (Weights sum to a 0–100 scale when each 0–5 score is multiplied out.)

Category 1 — Data Handling and Training Use

Score how the vendor stores your data and whether they train their models on it.

This is the single most important category. A vendor that trains on your inputs can expose your confidential data.

• Does the vendor train its models on your data by default? (No = high score.)
• Can you opt out of training, and is that opt-out in the contract?
• Is your data encrypted at rest and in transit?
• How long is your data retained, and can you request deletion?
• Is your data logically or physically separated from other customers’ data?

Category 2 — Security Certifications (SOC 2, ISO 27001)

Score the vendor’s independent security certifications and audits.

Certifications are third-party proof, not marketing claims.

• Does the vendor hold a current SOC 2 Type II report?
• Will they share the report or an executive summary under NDA?
• Do they hold ISO 27001 or an equivalent certification?
• When was their last penetration test, and will they share results?
• Do they have a documented incident response and breach notification process?

Category 3 — Access Controls

Score how the vendor controls who can access your data and their systems.

Weak access control is a common path to a breach.

• Does the platform support single sign-on (SSO) and multi-factor authentication?
• Can you set role-based access so staff only see what they need?
• Are admin actions and data access logged?
• How does the vendor control its own employees’ access to your data?
• Can you revoke access quickly when a user leaves?

Category 4 — Sub-Processors and Data Location

Score which third parties the vendor shares your data with and where that data lives.

Every sub-processor is another party that can touch your data.

• Does the vendor publish a current list of sub-processors?
• Which foundation model providers do they route your data to?
• In which countries or regions is your data stored and processed?
• Will they notify you before adding a new sub-processor?
• Do sub-processor contracts pass through the same data protections?

Category 5 — Model Provenance and Transparency

Score how clearly the vendor explains which models power the product and how they behave.

You cannot govern what you cannot see.

• Which underlying models does the product use, and can that change without notice?
• Does the vendor document known limitations and failure modes?
• Do they test for bias and provide any evaluation results?
• Can you get consistent versions, or do models change silently?
• Is there human oversight for high-impact outputs?

Category 6 & 7 — Uptime/SLAs and Compliance

Score the vendor’s reliability commitments and their fit with the laws that apply to you.

Reliability and compliance are contract questions, not sales-call questions.

• Uptime: Is there a written SLA with an uptime target (for example 99.9%) and service credits?
• Uptime: Is there a public status page and history of incidents?
• Compliance: Will the vendor sign a Data Processing Agreement (DPA)?
• Compliance: Are they GDPR ready, with a lawful basis and data subject request support?
• Compliance: How do they address EU AI Act obligations for their risk tier, and can they provide documentation?

Category 8 — Exit and Data Portability

Score how easily you can leave the vendor and take your data with you.

Plan your exit before you sign, not after a problem.

• Can you export your data in a standard, usable format?
• What happens to your data when the contract ends, and how fast is it deleted?
• Is there vendor lock-in that makes switching costly or impractical?
• What are the contract termination terms and notice periods?
• Will the vendor certify data deletion in writing?

Red Flags — Stop and Reconsider

Treat any of these red flags as a reason to pause and dig deeper before you sign.

One red flag is not always a deal-breaker, but it always deserves a hard question.

• The vendor trains on your data with no clear opt-out.
• They will not share a SOC 2 report or any security documentation.
• They cannot or will not sign a DPA.
• They will not disclose their sub-processors or data locations.
• They have no written SLA and no status page.
• They cannot explain how you would export your data and leave.
• Answers are vague, verbal only, or change between calls.

The Exact Questions to Send a Vendor

Send these questions to any AI vendor and use their written answers to fill in the scorecard.

Ask for answers in writing so you have evidence for your records and any future audit.

• Do you train your models on our data by default, and can we opt out in the contract?
• Can you share your current SOC 2 Type II report or executive summary under NDA?
• Do you support SSO, MFA, and role-based access control?
• Can you provide your current list of sub-processors and the regions where our data is stored?
• Which underlying models power the product, and how do you notify us of changes?
• What is your uptime SLA, and where is your status page?
• Will you sign a DPA, and how do you support GDPR and EU AI Act obligations?
• How do we export our data and confirm deletion when we leave?