A caller two minutes into a fifty-dollar order stops and says, "Hold on. Who am I actually talking to?" What your agent says back is a decision you made weeks earlier, whether or not you noticed you were making it.
Most operators land on a name because it sounds warmer than the alternative. "Thanks for calling Vesuvio, this is Mia, are you picking up or having it delivered?" That opening tests well in a room full of managers. It also creates a set of obligations that a nameless greeting doesn't, and those obligations come due on exactly the calls you care most about: the confused one, the angry one, and the one that turns into a review.
What a name changes about the first thirty seconds
Give the agent a human first name and most callers will assume a person until something tells them otherwise. That assumption does real work. People speak more naturally to a name. They interrupt less, they use complete sentences, and they don't do the slow, loud, over-articulated thing that callers do when they think they're talking to a machine. Speech recognition gets measurably easier when the caller stops performing for the microphone.
So the case for a name is not sentimental. It's that a name produces better audio, and better audio produces fewer order errors.
The cost lands later in the same call. Someone who has been talking to "Mia" for two minutes and then works out that Mia is software feels a little foolish, and mild embarrassment sours into irritation much faster than plain confusion does. You've traded a smoother opening for a sharper correction. Whether that's a good trade depends entirely on how often the correction happens, which depends on how often your agent hits something it can't finish.
The kind of restaurant where a name pays off
If your calls are short and repetitive, the correction moment barely exists. A wing shop taking sixty orders between six and eight has callers who want to say their order, hear a total and a time, and hang up. Almost nobody asks who they're speaking to, because nobody is engaged enough with the interaction to wonder. A name here is nearly free upside.
The picture inverts at a full-service room fielding private-event questions, dietary conversations, and callers who want to know whether the chef can do something off-menu. Those calls run longer, wander further, and hit the edges of what the agent knows. Every one of those edges is a moment where the caller re-evaluates who they're talking to, and a human name makes each re-evaluation slightly worse. A neutral greeting sets a lower expectation that the call can actually meet.
Run the count before you decide. Pull a week of calls, mark the ones that lasted over ninety seconds or included a question outside ordering, and see what share of your volume they are. Under ten percent, name it. Over a third, don't.
The answer that decides everything
Whatever you call the agent, one line matters more than the name: what it says when a caller asks directly whether it's a person.
The only workable answer is an immediate, unhedged yes-it's-automated, followed by an offer to get a human. "I'm an automated assistant. Want me to get someone from the restaurant on the line?" No pause, no deflection, no staying in character. An agent that dodges that question, or that says something coy like "I'm here to help with your order," produces the single worst version of the call and the review that follows it. We go through the broader version of this in telling customers you use AI, and the regulatory side sits in AI disclosure laws for voice agents.
Write this response yourself. Don't accept a vendor default and don't assume it exists. Call your own number during the demo, ask the question flatly, and listen to what comes back. Our demo test list has the rest of what to try.
If your agent handles that question well, a name is a reasonable choice. If it doesn't, the name is the least of your problems.
Names that cause trouble
Three patterns to avoid, learned the boring way.
A name that matches a real employee's name. Sooner or later a caller will ask for Sarah, mean the actual Sarah who works Tuesdays, and get the agent insisting it can help. Check your roster first, and check it again when you hire.
A name that's obviously a bot name. "Ava," "Aria," "Nova," and the rest of that family have become recognizable enough that some callers clock them instantly, which gives you the correction cost without the natural-speech benefit. If you want a name, pick one that sounds like it belongs to someone who could plausibly work at your restaurant in your neighborhood.
A name in the wrong register for the room. A steakhouse with white tablecloths and a phone host named Zoey has a small consistency problem that nobody can articulate but several people will feel. The name is part of the same set of choices as tone and pacing, which is the territory of choosing a voice persona and custom greetings and brand voice.
What your staff will do with it
The part nobody plans for: your own team starts using the name.
Within about two weeks, someone will say "Mia took this one" while reading a ticket, and the shorthand sticks. That's mostly harmless and occasionally useful, since it gives staff a clean way to talk about where an order came from. It becomes a problem in one specific situation, which is a caller escalating to a person and hearing an employee refer to the agent by name as though it were a coworker. That confirms the caller's suspicion that they were misled, even if nobody misled them.
Brief your team on one rule: on a live transferred call, it's "the automated system," never the name. Internally, call it whatever you want. Training staff on voice AI handoffs covers the rest of that conversation.
The case for skipping the name
Plenty of restaurants run without one and lose nothing. "Thanks for calling Vesuvio, are you ordering for pickup or delivery?" is a perfectly good opening. It doesn't claim to be a person, so there's nothing to correct later, and it starts the caller on the question you actually need answered.
The measurable cost is small: slightly more machine-directed speech from callers in the first few seconds, which shows up as marginally lower recognition accuracy on the first item. Whether you can see that in your own numbers is doubtful. Whether you can see a bad review complaining about being tricked is not, and that failure mode only exists on the named side.
So here's the decision rule. Name it if your calls are short, your correction answer is honest and instant, and no employee shares the name. Skip it otherwise. Then check your reviews for the next sixty days, since that's where the real verdict shows up, and handling negative reviews about AI is what to do if one lands.