Use the same fictional source material
Try one ordinary chat, one question that requires reading a saved note, and one proposed action you can decline. Some models can chat but cannot call the household tools. Also check who is billed and which provider receives the selected context after the switch. Save a sample note with a unique detail and use it for each connection. Keep the account, record visibility and request wording the same so differences in answers are easier to interpret.
Compare three different abilities
Ask an ordinary conversational question, a question that requires reading the sample note, and a request for a proposed change. Inspect where the answer came from and whether the assistant can show the proposed action without executing it. Use a test action outside any standing permission so the rehearsal actually requires approval. A model that handles conversation well may not support household tools or the input type you need.
Check cost and destination
Confirm whether requests use the household’s managed balance or a person’s provider account. Identify the provider that receives selected context after the switch. Keep the previous setup details available until the new connection handles the representative jobs; do not move private records into a test merely to make it feel realistic.