Search "AI customer service agent" and almost everything you find is selling the same thing: a voice receptionist that answers your phone. That's one real category, and if phone calls are your problem, fair enough. But the phrase covers more ground than that, and nobody shopping for one seems to be told which problem they're actually solving. Here's what to check before you buy, whichever channel you need covered.
Three different products share one search term
A chatbot answers from a script. It's fast and cheap, and it's useless the moment a question falls outside its FAQ — it either loops or hands off blind.
A voice receptionist answers your phone, books appointments, and takes messages. That's most of what currently ranks for "AI customer service agent" — a real, packaged product, built for one channel.
An agent reads what comes in, works out what it's looking at, and acts within limits someone set for it — reply, hold, or hand to a person. It's the least marketed of the three because it's the hardest to package: what it should do depends on your business, not a vendor's template.
None of these is "better". They solve different problems. The mistake is buying whichever one ranks first for the phrase and assuming it covers what you actually need answered.
Five questions to ask before you buy one
Which channel does it actually cover? Phone, email and chat fail differently. A tool that's excellent on the phone tells you nothing about how it handles an angry email thread. Ask for the specific channel, not the category.
What happens when it doesn't know the answer? Good: it stops and asks. Bad: it guesses confidently. Ask the vendor to show you an example of it saying "I'm not sure" — if they can't, assume it doesn't.
Who sees its mistakes before your customer does? If the answer is "nobody, it just sends," that's a decision about your business's voice you're outsourcing to a script you haven't read.
What's it actually measuring? Tickets closed and calls answered are activity. Hours you got back, and complaints that didn't happen, are the outcome. Ask which one the pricing page is counting.
What do you walk away with if you leave? The conversation history, the templates it learned, the configuration — is that yours, or does it reset to zero with a new vendor?
What I built, as one data point
I run an agent called Willow. She's not a voice product — she doesn't touch the phone, and I'd be overclaiming if I let that blur. What she does: she reads every inbound customer email, classifies it as routine, complex, needing escalation, or junk, and either fills in a matching reply template or drafts one from scratch. Every draft follows the same shape — acknowledge, answer, anticipate the next question, invite the next step — and anything with distress language, a complaint, or a pricing commitment gets escalated to me before it goes anywhere. She signs every reply "Paul Seville", never her own name, because the relationship is with me, not a bot answering on my behalf.
That's the shape of the second category above: narrow, accountable to a human on anything that matters, and honest about what it doesn't cover. It's also why I'm not an AI automation agency. I have an AI team. — one specialist with that kind of backup, not a script pretending to be a department.
If you're looking at a packaged tool
Run the five questions above against whatever's in front of you, phone product or not. If the sales page can't answer the escalation question in one plain sentence, that's your answer about how it handles the moment that actually matters.
If you're not sure which category your business even needs, a Systems Health Check is a faster way to find out than reading another vendor's FAQ page.