If you're weighing an AI phone agent, the most useful research isn't a vendor page. It's what other operators say when there's no salesperson in the room. Owners are blunt in the restaurant subreddits and in Facebook groups for franchisees, and the same themes come up in roughly the same order every time.
Here's what they raise, and what's actually true about each one.
The concern operators raise first: accuracy
The number-one worry is a wrong ticket. Owners have been burned by tech that demos well and fails on a busy Friday, and they're right to lead with it, because a bad ticket costs more than a missed call. A missed call is a lost order. A wrong ticket is a remake, a comp, a bad review, and a guest who now tells people your food came out wrong.
The honest answer is that accuracy lives in the POS write, not the voice. A system can understand a caller perfectly and still deliver a useless ticket if the modifiers land in a free-text note instead of on the item.
The right way to judge it is to place a hard, modified order yourself and see if it lands correctly. Not a cheeseburger. Order the thing your staff hates ringing in.
The second: does it sound like a robot
Nobody wants their regulars hitting a phone tree. This concern has aged faster than the others, because the voices really did get good. Modern agents talk at a normal pace, handle a caller interrupting mid-sentence, and don't force a menu of options at the top of the call.
What operators still complain about is timing rather than tone. An agent that pauses a beat too long before responding reads as broken even when the words are right, which is why latency matters as much as voice quality. Call the demo line from a noisy kitchen, not from your office, and listen for the gaps.
Your callers will judge this in the first five seconds, and they'll judge it against your host, not against a phone tree.
The third: what happens on the weird calls
Complaints, off-menu requests, an upset guest, a driver who can't find the door. This is where owners get specific, and the specificity is a good sign.
A workable agent solves what it can and hands the rest off cleanly with context, so a person picks up an informed call instead of a cold one. The failure mode owners describe is an agent that loops: it doesn't understand, asks the caller to repeat, doesn't understand again, and never escalates. That's worse than voicemail, because the guest spent ninety seconds finding out you couldn't help them.
Ask any vendor to show you a transcript of an escalated call, not just the escalation rate.
What a normal shift actually looks like
Operators who've run one for six months describe something less dramatic than either the pitch or the fear.
The phone rings during the 6pm push and nobody looks up. Orders print at the same rate they always did, mixed in with the online tickets. Two or three times a shift, a call rings through to the host stand because someone asked for a manager or the agent hit something it couldn't place. The host takes those, which is a smaller number of interruptions than answering every call at the stand.
The change owners actually notice isn't revenue. It's that the 11:30am catering call and the 5:45pm pickup order both got handled, instead of one of them going to voicemail because a person could only be in one conversation.
The complaint that shows up three months in
Almost nobody's second-quarter complaint is about the voice. It's drift.
A price went up in the POS and the agent kept quoting the old one. A limited-time item ended and the agent kept selling it. The kitchen 86'd wings on a Saturday and the agent took forty dollars of wing orders anyway, because the 86 was called out verbally and never entered anywhere a system could read.
That last one is worth sitting with, because it's not really a software failure. If your 86 process is a shout across the line, no integration can catch it. Getting value out of an agent usually means tightening how 86'ing actually happens first. Owners who report the fewest problems are the ones who assigned a single person to own menu edits and gave them ten minutes a week.
The numbers owners check, and the ones they should
The dashboards vendors ship lead with calls answered and containment. Answered calls will look great immediately, because the previous number was whatever your staff could get to.
The figures worth pulling monthly are narrower. Order accuracy, measured by reading tickets rather than trusting a summary. Abandonment during the agent conversation, which tells you if callers are hanging up on it. Escalation reasons broken into "by design" and "by failure," which is the actual work list. Containment on its own is easy to misread, because it climbs when an agent refuses to transfer.
Run the arithmetic yourself before you sign, too. If you take 60 calls a day and miss 15, and a recovered order averages $30 with a third of missed calls being real orders, that's roughly $150 a day in orders you weren't capturing. Compare that against plans starting at $250 a month and decide whether the gap is big enough to bother. If it isn't, the honest answer is that you don't need this yet.
Where owners say it wasn't worth it
Two patterns come up repeatedly, and both are about fit rather than product quality.
Low volume. A room taking twelve calls a day doesn't have a phone problem, it has a staffing preference. Automating twelve calls saves a few minutes and adds a monthly bill.
Relationship-heavy service. If a meaningful share of calls are regulars who expect a specific person, or private-event inquiries that turn into four-figure bookings, those calls should reach a human by policy. That's a legitimate configuration rather than a defeat, and it's the same logic behind routing catering leads to a person.
Why unfiltered accounts beat vendor pages
Operator communities are useful because the complaints are specific and the praise is grudging. Vendor pages are useful for confirming what a system connects to. Neither substitutes for the test.
So do the test. Call a demo line during your own dinner rush, order the way your worst regular orders, ask for something you don't sell, then ask for a manager. Three minutes. Then look at what reached the kitchen. If the ticket reads like a good employee entered it and the manager request routed to a person on the first ask, the concerns above are handled. If it doesn't, no amount of positive commentary from someone else's restaurant should move you.