What AI service actually feels like from the other side
A customer giving up isn't a saved contact. We need to stop counting it as one.
If we count avoided handoffs as success, an agent can waste ten minutes and still look good on paper.
That's shit service. Calling it containment doesn't improve it.
The question I keep coming back to is whether the thing can actually help. A friendly answer is worth very little when you're still stuck.
If I need an explanation, give me one. If I need something changed, either do it or get me to someone who can. Don't make me spend ten minutes discovering you were never allowed to help.
That's my objection to sympathy without action. Being told my frustration is understood gets old pretty quickly when nothing happens next.
There is plenty to like when it works. You can ask the small question, get an answer while you still need it and carry on. You don't have to explain the whole thing again after every handoff.
Those are useful improvements. They're also things we can look for in the actual service, without arguing about how clever the model is.
A model can sound sure and still be wrong — NIST identifies this risk too. That isn't a failure rate for your service; you still have to check its answers.
I want the service to know its limits and make them clear early. Permission belongs in the system, and the reply should reflect what the agent is actually allowed to do.
When it hands over, the next person needs the relevant history. Making the customer repeat everything means we've added work to their day. We shouldn't need a survey to notice that.
We do need evidence before claiming the task was resolved. If there's a changed record or completed action, check it. If there isn't, ask the customer and look for repeat contact about the same problem.
None of those measures is perfect. A quiet customer might be happy or might have given up. A transcript won't tell us everything they felt.
That's a reason to check more than the conversation ending. It isn't a reason to count the ending as success.
I can also overdo the argument for action. Some people want instructions and control over what happens next. If clear guidance gets their job done, good — the agent doesn't need to touch anything.
The point is whether the customer got what they needed. Take a conversation your dashboard calls resolved and read it from their end.
Would they call it resolved too?