Ask the vendor how they detect AI feature failures, how fast they notify you, how you roll back, how they find root cause, and what you can do in the meantime. Get this incident process in writing before purchase, since you will need it on the day it matters.
Failures to plan for
An AI feature can fail in several ways, and each needs a different response.
| Failure | Example | Response you need |
|---|---|---|
| Wrong output | A bad fix proposal or wrong priority | Reject, report, track |
| Systematic drift | Accuracy drops after a model change | Rollback or disable |
| Data exposure | Information reaches the wrong party | Incident process, notice |
| Manipulation | Planted text steers behavior | Containment, patch |
| Outage | Model service is unavailable | Graceful fallback |
The questions
- How do you detect that the feature is misbehaving, and who watches?
- How soon would we be told about an issue that affects us, and through what channel?
- Can we disable the feature ourselves, immediately, without a ticket?
- Can we roll back to an earlier behavior, or pin a version?
- What do you provide after an incident, such as a written review?
- How do we report a bad output, and what is the typical response?
- Does the product still work if the AI part is off?
The last question is quietly the most important. A product that degrades to its non AI core is safer than one that stops.
Check the paperwork against the product
Ask where the process is documented and who owns it. A named contact and a documented channel are better than "reach out to your account team". Your legal and privacy teams should review notification terms; this post is not legal advice.
Run a small drill
Ask the vendor to walk you through a sample scenario, such as a wrong fix that merged. Where would the evidence be, and how would you revert? Then try the disable step in your own tenant. A drill takes under an hour and shows whether the answers are real.
Use the frameworks as prompts
The NIST AI Risk Management Framework, which is voluntary, organizes AI risk work under Govern, Map, Measure, and Manage, and the Manage function covers responding to risks once they appear. The OWASP Top 10 for LLM Applications 2025 names the failure modes above, from prompt injection to excessive agency.
Keep your own record
When something goes wrong, capture the inputs, outputs, tool version, and time. If the vendor's log is the only record, you depend on them to explain themselves.
Try it yourself
Click through it at your own pace, or open it full screen.
Watch how changes are surfaced over time and consider how you would notice a shift in an AI feature.
Where Safeguard fits
Safeguard offers policies, gates, and guardrails, and keeps autonomous remediation off by default with every category on Manual, which limits what a faulty proposal can do.
Quick check
Can I run a product with the AI feature off?
Ask. It should still work, with less assistance.
Who should hold the incident record?
You, not only the vendor.
Sources
- NIST, AI Risk Management Framework: https://www.nist.gov/itl/ai-risk-management-framework
- OWASP Top 10 for LLM Applications 2025: https://genai.owasp.org/llm-top-10/
Last reviewed: 19 October 2026.
