An AI vendor should be disqualified when it lacks basic controls that protect your data and cannot show a credible plan to close the gap. Missing encryption, no independent audit evidence, no access controls, or no commitment on whether your data is used to train models are the clearest deal-breakers.
Context matters, though. A gap in a low-risk use case differs from the same gap in a tool that touches customer data or financial records. This article explains how to separate fatal gaps from fixable ones, which controls to demand, how to verify them, and when to negotiate instead of walking away.
What separates a deal-breaker control gap from a fixable one?
A deal-breaker gap exposes your data or operations to harm that the vendor cannot remediate within your risk tolerance or timeline. A fixable gap is one the vendor can close through a documented, dated, contractually binding commitment without changing how your data is handled in the meantime.
Use three tests to classify any gap in your AI vendor risk assessment:
- Data sensitivity: Does the AI tool process personal, financial, or confidential data? The more sensitive the data, the fewer gaps you can accept.
- Blast radius: If the control fails, how many systems, customers, or processes does it affect?
- Remediation speed: Can the vendor fix the gap before you go live, and will they commit to that in writing?
A missing control that fails all three tests is a disqualifier. A missing control that involves low-sensitivity data, is contained, and is easy to fix belongs in a negotiation, not a rejection.
Which missing security controls are automatic red flags for AI vendors?
Automatic red flags are missing basics that no serious vendor should lack: no encryption of data in transit and at rest, no independent attestation such as SOC 2 or ISO 27001, no role-based access control, no incident response process, and no breach notification commitment. Any one of these should pause your approval.
Treat these as your core AI vendor disqualification criteria:
- No encryption: Data moves or sits unprotected. This is a fundamental failure, not a configuration detail.
- No independent audit evidence: A vendor with no SOC 2 report, ISO 27001 certificate, or comparable assessment asks you to trust claims alone.
- Weak access management: No multi-factor authentication, no least-privilege access, and no logging of who touches your data.
- No incident response or breach notification terms: You cannot manage a risk you learn about weeks late.
- Unclear subprocessors: The vendor cannot tell you which third parties handle your data and where.
- No data deletion process: You cannot get your data removed when the contract ends.
Vague answers deserve as much scrutiny as missing ones. If a vendor cannot explain a control in plain terms, assume it does not operate reliably.
Which AI-specific controls must a vendor have before approval?
Beyond standard security, an AI vendor must control how your data feeds models, isolate your data from other customers, and protect against AI-specific attacks. Without these, a vendor with strong general security can still leak your information through the model itself.
Require clear answers on these AI data protection controls before you approve:
- Training data use: The vendor must state whether your inputs train its models. Your default position should be that they do not, unless you opt in.
- Tenant isolation: Your prompts, outputs, and fine-tuned models must stay separated from other customers.
- Prompt injection and output safeguards: Ask how the vendor tests for manipulated inputs and prevents the model from exposing sensitive data.
- Model and data provenance: The vendor should document where training data and underlying models come from, including any third-party foundation models.
- Retention limits: Prompts and outputs should have defined retention periods you can control.
- Human review and monitoring: Know who at the vendor can read your inputs and under what conditions.
A vendor that cannot answer the training data question clearly fails a basic AI vendor due diligence check. Do not accept “we take privacy seriously” as an answer.
How can you verify a vendor’s security controls actually exist?
Verify controls by asking for evidence, not assurances. Request current audit reports, review them for scope and exceptions, test claims through a tailored security questionnaire, and confirm that contract terms match what the vendor says. A control exists only when you can see proof that it operates.
Follow this sequence for a reliable check:
- Request the reports. Ask for the latest SOC 2 Type II report or ISO 27001 certificate with its statement of applicability. A SOC 2 AI vendor claim means little without the report itself.
- Read the scope. Confirm the audit covers the actual AI product you will use, not just the vendor’s corporate IT or a different service.
- Check the dates and exceptions. Look for reports older than twelve months and for noted control failures.
- Send an AI vendor security questionnaire. Add AI-specific questions on training data, isolation, and model safeguards to your standard list.
- Test where you can. Ask for penetration test summaries, a live demo of access controls, or a short technical session with the vendor’s security team.
- Match the contract to the claims. Write the controls into the data processing agreement so they carry legal weight.
Contradictions between the questionnaire, the audit report, and the contract signal a risk by themselves. Resolve them before you sign.
Should you walk away or negotiate when controls are missing?
Walk away when the missing control is a core protection, the data is sensitive, or the vendor refuses to commit to a fix. Negotiate when the gap is contained, the vendor shows credible intent, and you can bind them to a dated remediation plan with consequences.
Negotiate if you see all of these conditions:
- The gap is not on your automatic red flag list.
- The vendor acknowledges the gap openly and offers a specific timeline.
- You can limit exposure in the meantime, for example by restricting the tool to non-sensitive data.
- The contract includes audit rights, remediation deadlines, and the right to exit if the vendor misses them.
Walk away if the vendor dodges questions, offers only verbal promises, or treats a basic control as optional. How a vendor responds to your scrutiny predicts how it will behave after an incident. Because AI tools often drive real spend, weigh the security decision alongside cost and value. Our strategic portfolio management approach helps you compare vendors and initiatives against business priorities, so you do not approve a risky tool just because it looks cheap.
How It’s Value helps with AI vendor risk and spend
We help you make AI vendor decisions on facts instead of sales claims. Security gaps and cost overruns often come from the same root problem: nobody owns the decision, and nobody sees the full picture across finance, IT, and procurement. We close that gap.
Here is how we support you:
- We build clear governance with defined roles and decision rights, so you know who approves a vendor and on what criteria.
- We connect finance, IT, procurement, and the business, so risk, cost, and value enter the same conversation early.
- We bring transparency to technology and cloud spend, which helps you spot unmanaged AI tools and compare vendors on a like-for-like basis.
- Through our FinOps services, we help you control the usage-based costs that AI platforms often create.
- We turn vendor assessments into decision-ready insight, not just another report.
Do you want to approve AI vendors with confidence and keep their costs under control? Get in touch with us and we will show you how to build a vendor approval process that works for you all.