After the approval, can anyone show me what the system actually did?

The moment.

The demo was going well. The vendor's AI tool had read a customer file, drafted a reply, and a person had approved it. Green tick. Then someone asked to change the recipient and run it again. It sent. The tick stayed green. Nobody could say whether the first approval covered the second send, or where the record of it lived.

What you would see.

The brochure: the vendor's documents say the product can do something. That tells you what they meant to build. The settings: you open the product and look at how it is configured. That tells you what it would do in general. The test you run yourself: in your own system, one real transaction, then you open the record and read what happened. Only the third kind is proof. The brochure is a promise. The settings are a plan. The test is a fact.

A good answer to the recipient test: the second send is stopped, or goes out with a fresh approval, and there is a record showing who approved what and when. A bad answer: the tick stays green, the email is gone, and the explanation is a sentence from the vendor.