AI customer support sounds simple.
Give an AI your documentation, let users ask questions, and let the AI answer.
That was roughly what I thought when we started building AI support into FeedLog.
Then we actually started testing it.
And I realized the hardest part isn't getting AI to answer.
It's deciding when it should answer at all.
The happy path is easy
When a user asks:
How do I export my data?
And the answer clearly exists in your knowledge base, everything works beautifully.
User asks → AI finds the right information → user gets an answer.
Done.
But real users don't always ask questions your documentation can answer.
They ask things like:
Why was I charged twice?
My account stopped working. Can you fix it?
I followed the documentation and it still doesn't work.
Now the AI has a decision to make.
Should it answer?
Should it ask another question?
Should it admit it doesn't know?
Or should it bring in a human?
That's where AI support gets interesting.
A wrong answer is worse than no answer
With a normal chatbot, a hallucination is annoying.
With customer support, it can become a real problem.
If the AI confidently gives the wrong billing information or invents a feature that doesn't exist, you've saved zero support time.
You've probably created another support ticket.
So we're starting to think about AI support less as:
"How many questions can AI answer?"
And more as:
"How many questions can AI safely resolve?"
Those are very different metrics.
The goal isn't to remove humans
I don't think good AI support means humans disappear.
The better model might be:
- AI handles repetitive, well-documented questions.
- AI recognizes when information is missing.
- AI understands when a conversation is becoming complicated.
- Humans step in only when they're actually needed.
The interesting part is the handoff between those two.
We're still figuring this out while building FeedLog.
But one thing has become pretty clear:
A good AI support agent shouldn't just know how to answer. It should know when not to.
If you're already using AI for customer support, I'd be curious to know: how much do you actually trust it to handle on its own?
Top comments (1)
the wrong answer being worse than no answer is the real insight here. if a bot admits it does not know and hands off, that is a fine outcome. the failure mode is confidence with no signal attached, same answer whether it is sure or guessing.