Where it genuinely helps
Drafting a first reply for an agent to review. Summarising a forty-message thread before handover. Detecting language and routing accordingly. Suggesting a category. Each of these saves real time and none of them can embarrass you, because a person approves the output.
Where it is risky
Answering customers directly with no human in the loop. The failure mode is not gibberish — modern models write beautifully. The failure mode is a confident, fluent, plausible answer that is wrong, sent in your name, in writing.
Ground it in your own content
An assistant that answers from your knowledge base is far safer than one answering from general training. If the answer is not in your content, the correct behaviour is to say so and offer a human, not to improvise.
Keep the key and the data yours
Bring your own API key so usage runs on your account, on your server, against your data, under your terms — rather than being metered per resolution by a vendor.
A sensible rollout
Enable one capability. Review its output for a fortnight. Keep drafts as suggestions requiring approval. Only after that consider letting it answer chat directly, and only for questions your knowledge base genuinely covers.