Most callers will not care. That is the honest answer, and it disappoints people on both sides of the argument.
The reaction operators brace for is a wave of offended regulars. What usually arrives is silence: orders keep landing, tickets keep printing, and nobody says anything about the voice on the phone one way or the other. The people who do react tend to react to something specific, and it is almost never the fact that the caller reached software.
What the reactions actually look like
Two surfaces show you real reaction. The first is your transcripts, specifically the calls that ended in a hangup during the conversation or in a transfer request. The second is your counter, because a customer annoyed enough to mention it will mention it to the person who hands them their bag, not to a survey.
Neither surface produces a clean percentage, and you should be skeptical of anyone who offers you one. What they produce is a list of moments where the call went sideways, with enough detail to fix the cause.
When operators read a month of those transcripts, the pattern is consistent. A small number of callers hang up in the first four seconds, usually before the agent has finished the greeting. A smaller number ask for a person and get one. A handful get stuck on a menu item the agent cannot parse. The rest complete their order and go about their day.
The complaint that is really about latency
The single most common source of irritation is not the voice. It is the gap.
A person on the phone fills dead air without thinking about it. They say "mhm" while typing, or "let me check on that." An agent that goes silent for two full seconds after a caller stops speaking sounds broken, and callers respond to broken by repeating themselves, which makes the problem worse. The caller then describes the experience as "the robot didn't understand me," when the agent understood fine and simply answered too late.
This is why response time is worth testing before you ever look at accuracy numbers. The threshold where a conversation stops feeling like a conversation is lower than most people guess, and the reasoning behind that is in sub-second latency for restaurant voice AI. Related, and just as irritating: an agent that talks over a caller who is still speaking, or that treats a mid-sentence breath as the end of a turn. That behavior is covered in interruption handling and endpointing.
Fix the timing and a meaningful share of what you would have logged as "customer dislikes AI" disappears.
The complaint that is really about a dead end
The second cluster is callers who wanted a person and could not get one.
This one produces the angriest calls and the worst reviews, and it is entirely self-inflicted. If your escalation rules require a caller to ask twice, or route to a phone nobody answers during a rush, you have built a trap. The caller's complaint will be about AI. The cause is your handoff policy.
Set the trigger low. A caller who says "can I talk to someone" should be transferred on the first request, without a clarifying question and without an attempt to solve the problem first. Decide in advance where that call lands when the store is slammed and nobody can pick up, because "rings out at the host stand" is not a policy. The options and their tradeoffs are laid out in human handoff and failover.
There is a version of this problem that comes from the other direction. Some operators set the trigger so low that the agent transfers on any hesitation, which puts the phone right back where it was. The setting you want is: transfer immediately on request, transfer on complaint or refund, transfer on anything involving a large catering order, handle the rest.
The complaint that is really about accuracy
The third cluster is the one worth taking seriously as a product problem rather than a sentiment problem.
A caller whose order arrives wrong does not file a complaint about automation. They call back, they ask for a remake, and the interaction costs you the food, the labor, and the goodwill. If that happens more after you turn the agent on than it did before, the tool is not ready and no amount of tone adjustment fixes it.
Track this from day one by checking tickets against transcripts for a couple of weeks rather than trusting a dashboard. The method is in improving phone order accuracy. Pay particular attention to how the agent handles callers with strong accents or heavy background noise, since those calls fail quietly and the caller rarely tells you why. Both cases are discussed in accent and speech recognition accuracy.
Why disclosure lowers the complaint rate
Callers figure it out. They always figure it out, usually within one exchange, and the ones who feel they were fooled are the ones who write about it afterward.
A greeting that says plainly what they have reached, in your restaurant's own words, and mentions that a person is available, costs you two seconds and removes most of the grievance. It also sets the caller's expectations correctly, which makes them speak more clearly and complete orders faster. The wording matters more than you would think, and the patterns that work are in custom greetings and brand voice and telling customers you use AI.
The alternative approach, a voice designed to pass as human without saying so, is a bet that nobody notices. People notice.
Reading your own numbers instead of ours
Here is the method, and it takes about twenty minutes a week for the first month.
Pull every call that ended without a completed order. Sort them into four piles: hung up during the greeting, asked for a person, agent failed to understand something, and wrong number or out of scope. Count each pile. The first pile is mostly people who called the wrong place or changed their mind, and it exists with human answering too. The second pile tells you whether your handoff is working. The third pile is your fix list, and it is usually five or six specific menu items.
Then ask two counter staff what customers have said. If nobody has said anything in three weeks, that is your answer.
Compare all of that against what the phone was doing before. If calls were ringing out during your dinner rush, the relevant comparison is not "AI versus a friendly human" but "AI versus nobody picking up," and the arithmetic on that is in the real cost of a missed restaurant phone call.
Where the objection is correct
Some restaurants should not do this, and it is worth saying so plainly.
If your call volume is genuinely small and the owner answers on the second ring while chatting with the caller by name, an agent removes something the customer is actually paying for. If your business is built on private events negotiated over the phone, most of those calls should reach a person anyway, and automation buys you very little. The pushback you get in those rooms is not resistance to change. It is accurate.
Where the objection stops being correct is any operation with calls ringing out at peak, a host stand phone pulling someone away from guests, or a voicemail box nobody clears. In those rooms the current caller experience is already bad, and the question is not whether callers prefer a person. It is whether they prefer an agent to a busy signal.
Run this test in your first month: pick the ten longest calls and the ten shortest, read them end to end, and mark each one as something you would be comfortable with a regular hearing. If more than one or two fail that bar, you have specific work to do, and the work is nearly always a menu correction, a timing setting, or a transfer rule rather than a decision to give up on the whole approach.