A customer service agent receives the message: "Hi, I'm John Smith, can you help me with my account?" The agent calls lookup_customer(name: "John Smith"). The tool returns three customer records — three different John Smiths in three different cities, all with active accounts.
The agent currently picks the most recently active record and proceeds.
What's the right behavior?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
When a lookup returns multiple matches, the right move is to ask for a disambiguating identifier, not to guess. Heuristics like "most recently active" feel reasonable but produce predictable, expensive failures — the case where the wrong John Smith's account gets modified is exactly the case the heuristic doesn't catch. A short clarifying question costs the customer two seconds; a wrong-account modification costs trust, time, and possibly money.
Recency seems sensible until the most recent active record is the wrong John Smith. The heuristic produces a confident wrong answer instead of a quick right one. "Reasonable-looking" is not a substitute for "correct."
Running three parallel conversations means three customers' data is touched on a single ambiguous request — privacy, audit trail, and reconciliation problems compound. Ask the question; pick one record.
A random pick with a disclaimer is somehow worse than a heuristic pick — you've formalized that you don't know who you're talking to and proceeded anyway. The customer can answer a clarifying question; let them.