Ask HN: The agent lied to you, how will you handle it?

3 points by yeeyang 2 months ago | 5 comments
I was looking through the logs today, and saw this user asked our Agent to update their credit card phone number. Thing is, they'd already tried before, but the ID check failed, so customer service told them they had to handle it in person.

Anyway, our Agent asked the user for more ID info, and should have called customer service again to give it another shot. But nope! It kinda just... hallucinated? Basically, the Agent lied to the user.

The logs clearly show it never even tried to make the call. It just told the user something like, "Hey, I called again for you, but they still can't do it. You should probably try another way."

And because it didn't actually make a call, our review system didn't catch anything wrong – no call record, right? So, the human review step got skipped, and that bogus reply went straight to the customer.

Man, it seems like these Agents can get "lazy" or avoid tricky tasks, just like people, huh? I'm scratching my head trying to figure out how to stop this from happening again. Got any ideas?

  • samchon 2 months ago
    If there are some prepared functions, I do not provide it as tools directly. Instead, make the orchestration to pass a function seleting agent. This strategy makes success rate of function calling higher.

    Github: /wrtnlabs/agentica

    Here is my repo, and if you click principles chapter, then you can see the detailed story of it (my karma is lower, so cannot direct link to website)

    With this strategy, I could accomplish a properly working shopping agent that is composed by 289 functions.

    • yeeyang 2 months ago
      much appreciated!
    • gregjor 2 months ago
      Don’t trust important business processes or customer relationships to half-baked tech.

      This kind of thing tells your customer very clearly how little your company cares about them. Calling a chatbot an “agent” doesn’t make the experience better for the customer.

      • yeeyang 2 months ago
        We use this method to help users solve problems, and there are many successful cases.
        • tailspin2019 2 months ago
          FWIW I’m tentatively bullish on AI (for specific use cases) but I have never once… ever… ever… had a useful interaction with a customer service chatbot.
      • 2 months ago