Disclosure is not the risk. Discovery is.
A caller who is told in the first six words that they've reached an automated line, and who then gets their order taken correctly in ninety seconds, will order again. A caller who assumes they're speaking to a person, spends four minutes on a catering question, and works it out from a mangled response will tell people about it. Same technology, same restaurant, opposite outcome. The difference is who found out first and how.
That's the whole argument for saying it upfront, and it holds independently of whatever legal requirement applies where you operate.
What the rules actually ask of you
The regulatory picture is uneven and moving. Some jurisdictions require notice at the start of an automated call, some require an honest answer when the caller asks, and some address outbound calls while saying nothing about a customer who dialed you. If you operate in more than one state, you likely have more than one rule. A short summary of the current shape sits in AI disclosure laws for voice agents, though a lawyer in your state is the right source before you finalize anything.
The operationally useful version is simpler. Build the system so it discloses upfront and answers honestly on request, and you satisfy the strictest version of every rule currently in play without tracking any of them. There is no scenario where the extra disclosure hurts you enough to justify the compliance work of doing less.
Six words in the greeting, not sixty
The mistake operators make is treating disclosure as a disclaimer.
"Please be advised that this call is being handled by an artificial intelligence system and may be recorded for quality and training purposes" is a wall. It burns eight seconds, it primes the caller to expect a phone tree, and it raises your abandon rate before anyone has said a word about food.
What works is a phrase folded into the greeting so it barely registers as an announcement. "Thanks for calling Vesuvio, this is our automated ordering line. Pickup or delivery?" The caller gets the fact, gets the first question, and starts talking. Nobody hangs up on that, because the disclosure and the useful part arrive in the same breath.
Two details matter in the phrasing. Use a word that describes what it does rather than what it is, since "automated ordering line" tells the caller something they can use and "AI assistant" doesn't. And put the question immediately after, so there's no silence for the caller to fill with a decision about whether to stay on the line.
Recording notices are a separate matter and follow separate rules, which differ by state and in some places require consent from both parties. If you need one, keep it in the same sentence rather than adding a second announcement. Two disclosures stacked back to back is how a greeting turns into a phone tree, and callers who hear the shape of a phone tree stop listening to the words.
The one place to be more explicit
Reservations and large-party inquiries deserve a second sentence. Someone booking a table for twelve is making a commitment based on what they're told, and that's the call most likely to end with "I spoke to someone and they said." Adding "I can also get a manager on the line if you'd rather" costs two seconds and closes the gap. The general policy for that handoff belongs in human handoff and failover.
What to say when they ask directly
The answer to "is this a real person?" is one sentence, delivered without a pause: no, this is an automated assistant, and here's a person if you want one.
Anything else fails. A hedge sounds evasive. Staying in character sounds dishonest. A long explanation sounds defensive. The caller asked a factual question and the only good response treats it as one.
Test this yourself before launch by calling your own number and asking flatly, at three different points: during the greeting, mid-order, and after giving a credit card. Some agents handle it in the first case and stumble in the third. If you've given the agent a human name, this response carries more weight than it otherwise would, which is part of what makes naming your AI host a decision worth taking seriously.
The parts of your operation that also need a line
Disclosure fails when the phone is honest and nothing else is. Four places to cover, and they take an afternoon total:
- Your contact page gets one sentence saying calls are answered by an automated system and a person is available on request. This is also what you point to if anyone ever questions your practice.
- Your printed and digital menus, if they list a phone number with any language like "call and ask for," get updated so the instruction still makes sense.
- Your staff get a single sentence they're comfortable repeating when a guest asks at the counter, and permission to say it without apologizing.
- Your third-party delivery and catering partners, if they call your line to sort out a problem, get told the number is automated so they stop assuming a bad connection.
The staff line is the one that matters most and gets skipped most. A team member who mumbles "yeah, it's some AI thing, I don't know" teaches the guest that the restaurant is embarrassed. A team member who says "we put an automated system on the phone so calls stop ringing out during dinner, and you can always get one of us" teaches them it was a deliberate operational choice. It was.
Measuring whether it cost you anything
You can answer this with your own data instead of guessing, which is better than any general claim about how customers feel.
Look at hang-ups in the window between the greeting ending and the first menu item being spoken. That group is reacting to the disclosure and nothing else. Get a baseline in the first two weeks, then change one thing about the greeting and watch the same number. Most operators find it sits low enough to be noise, and the ones who find it high usually have a greeting problem rather than a disclosure problem: too long, too slow, or an agent whose first response takes a beat too long to arrive.
One caution about reading that number too early. Your first two weeks include regulars calling a line they've called for years, and some of them will hang up out of surprise and call back, or drive over instead. That group settles. Give it three weeks before you decide the greeting is costing you orders, and compare against the same days of the week rather than against your launch weekend.
The broader question of whether customers reject the whole idea is covered in will customers hate AI answering, and the related worry about sounding mechanical in does AI phone answering sound robotic.
If that early hang-up number is quiet after a month, you're done thinking about disclosure and you can spend the attention on order accuracy, which is where it belongs. If it's loud, tighten the greeting before you consider removing the disclosure, because taking it out doesn't fix anything. It just moves the complaint from your call log to your reviews.