Start from the sentence, not the settings
Open the conversation and find the exact sentence that was wrong. Not the general shape of the failure — the sentence. What the agent said tells you where it came from, and reading it next to the question is usually enough to sort it.
Then ask the same question yourself in the test chat. If the answer is wrong there too, you have something reproducible to fix. If it is right, the difference is usually context earlier in the customer's conversation, which is its own useful finding.
Cause one: it did not have the fact
The agent said it did not know, or gave a vague answer that dodged the question. Search your knowledge for the fact. If it genuinely is not there, the agent was right to say so — 'our hours are not listed in what I have here' is an honest answer to a website that does not list its hours, and it is a content problem, not a model problem.
The fix is a note: state the fact plainly, in one or two sentences, and re-ask. This is the most common cause by a distance, and the cheapest to fix.
Cause two: it was told something untrue
The agent stated something specific and confident — a price, an address, a policy — that is simply wrong. Look in the instructions under How it talks before you look anywhere else. A fact that was hand-typed into the instructions gets asserted with complete confidence and never gets corrected by a re-crawl.
Delete it from the instructions. If it is true, put it in knowledge instead, where it has a source and can be updated. Instructions should read like rules for behaviour, not like a fact sheet.
Cause three: nobody could take it
The customer asked for a person, or asked something no agent should answer, and the conversation went nowhere. Check whether asking for a human is switched on under Actions, and on the phone, whether a transfer number is set at all. With no number there is no transfer tool, so a caller saying 'representative' three times gets three polite non-answers.
This is worth checking on every agent you run, not just the one that failed. It is a single empty field and it is invisible until a real customer hits it.
Actions
Is asking for a human enabled? On a new agent it is off by default.
How it talks
Is there a transfer number, and is it a line somebody answers?
Business hours
If handover is limited to them, does the out-of-hours message say something useful?
Then re-ask
Every fix ends with the same question asked again in the test chat. A change you did not verify is a change you are guessing about, and knowledge in particular can surprise you — a note that reads perfectly can still lose to a page that mentions the same words more often.
If the answer is still wrong after a note that plainly states the fact, that is a retrieval problem rather than a content one, and it is worth telling us: it is the kind of thing we fix in the product rather than in your configuration.