For buyers
You are being delivered an AI agent. Five questions before you sign it off.
A supplier delivers a chatbot or an agent that will refund, write to your customers or change your records. The demo always goes well. What matters is what it does when a customer insists or someone poses as a manager. Here are the questions to ask before accepting it.
What CrashMire shows you
- What may the agent do on its own, and what goes through a person? Ask for the list of its tools and the limits written in the code, not only in its instructions.
- What did it do facing a pushy customer, a fake manager and an instruction hidden in a message? Ask for replayed conversations, not a summary.
- What happens when the AI model changes? Ask who runs the test again, and when.
- Where is the log of its acts? Every refund, sending or change must be traceable and explainable.
- Who tested it, on what date, on which version? A test by whoever built the agent is not worth an independent one.
A crash test documents what was checked, on a date, on one version of the agent. It is neither a certification nor a guarantee that no incident will occur.
The proof, public
- The proposal →
Testing AI agents before handing them tools, with public criteria.
- A sample report →
What the diligence report holds: faults, frequencies, fixes, re-test.
- The method →
Public and versioned: what is measured, how, and its limits.