Key takeaways
- Getting a voice agent to say hello takes an afternoon. Getting it through a noisy Saturday call with a party of nine and a nut allergy is the real build.
- A homemade agent makes you the owner of every part: phone line, speech, voice, instructions, handoffs, records, recording notices and fixes.
- Speech recognition is not equally accurate for every caller[2], so test with the voices and phone lines you actually get.
- The running cost is your manager's review time and on-call duty. Build only if someone can own both.
Hello is the easy part
The builder demos make it look like an afternoon's work: paste in the menu, pick a voice, connect a number, call it. That part really is quick. The work starts when a guest rings at 7:40 on a Saturday from a car with a weak signal, asks for a table for nine, mentions a nut allergy halfway through and talks over the agent twice. A restaurant phone agent earns its keep on that call, and it has to leave your host something usable when it hangs up.
A working agent is a chain of parts: a phone number or forwarding rule, speech recognition, a language model that decides what to say, a synthetic voice, your house information, rules for handing calls to staff, somewhere to keep recordings and transcripts, alerts for the team, and the greeting that tells callers they are speaking to an automated system and that the call is recorded. Build it yourself and you choose, pay for and keep every link in that chain working.
- Your team
- System
01Your team
Choose the parts
Number, speech, voice and model, each from a provider you pick and pay.
02Your team
Write the instructions
Hours, policies, menu facts, and the answers it must never guess.
03System
Wire the handoffs
Transfers, a backup when nobody picks up, and where records go.
04Your team
Test real calls
Kitchen noise, accents, interruptions, odd requests, bad signal.
05Your team
Go live and keep watch
Read transcripts, fix answers, track every provider's changes.
Example sequence. The last step never ends, which is the part most build estimates leave out.
The parts list, and who owns each
| Part | Build it yourself | Managed service |
|---|---|---|
| Phone line and forwarding | Buy a number or set up forwarding; handle outages | Forwarding rules agreed and tested with you |
| Speech and voice | Pick providers, tune for noise and accents | Chosen and maintained by the service |
| Instructions | Write prompts and keep them current | You approve the facts and boundaries; the service builds the script |
| Handoffs to staff | Build transfers, backups and a no-answer step | Set from the manager path you choose |
| Records and alerts | Store recordings and transcripts, write summaries, send alerts | Summary by email; recording and transcript in the account |
| Recording notice and disclosure | Write it and check it against each state's rules | A greeting with the notice, approved by you |
| Monitoring and fixes | Your evenings and weekends | The service's support, plus your weekly review |
Operators already know where this pinches. In the National Restaurant Association's November 2026 research, complexity was the most cited concern about AI at 46.1%, ahead of cost at 43.5% and customer acceptance at 41.9%[1]. Only 3.2% of operators were implementing AI projects companywide, while 20% were running small pilots, and training and education topped the list of support operators wanted, at 56.5%[1]. A build adds to every one of those pressures.
Where homemade agents break on restaurant calls
Most failures happen in the ear, not the brain. Speech recognition is not equally accurate for everyone: a 2020 study in PNAS tested five commercial speech recognition systems and found an average word error rate of 0.35 for Black speakers against 0.19 for white speakers[2]. Phone audio, kitchen clatter and cars make every system worse. A build that sounds perfect in a quiet office has to be tested against the guests who actually call you.
- Callback numbers. One misheard digit makes the whole request useless. Read every number back, and test it with callers who speak quickly.
- Dates. “Next Friday” said on a Thursday, “the 14th” without a month, “tomorrow” after midnight. Make the agent repeat the full date.
- Interruptions. Guests talk over greetings and change their minds mid-sentence. The agent has to stop talking and listen.
- Allergies. US law names nine major food allergens[3], and a guess about any of them is a liability. The rule is a handoff to a person, every time.
- Gaps in the instructions. Asked something you never covered, a model will produce a plausible answer unless it is told to say it doesn't know and take a message.
- Silence. When one provider in the chain slows down, the caller hears dead air and hangs up. Somebody has to notice.
Example test call for a homemade agent: from a mobile in a noisy room, say “Hi, yeah, can I get a table Friday, um, actually Saturday, for nine, maybe ten, around seven-thirty? One of us has a tree nut allergy.” A pass is a read-back of Saturday, the full date, nine or ten guests, 7:30, a name and a number, plus a handoff for the allergy. Anything less is a request your host can't act on.
Two existing guides do most of the groundwork for either path: preparing restaurant information for AI tools for the source material, and how to test an AI phone agent before going live for the test matrix. The handoff rules in when an AI phone agent should hand off to staff apply whoever builds it.
Consent, recording and disclosure are yours too
Building your own means writing your own greeting and owning the law behind it. Recording first: federal wiretap law lets a party to a call record it, or someone with one party's prior consent[4], but a number of states require every party's consent, including California, Florida, Illinois and Pennsylvania[5]. A recording notice at the start of the greeting is the simple way to cover callers wherever they are.
Outbound calls are stricter. On February 8, 2024 the FCC ruled that calls made with AI-generated voices are “artificial” under the Telephone Consumer Protection Act, so they face the same consent rules as other artificial or prerecorded voice calls[6]. FCC rules also require an artificial or prerecorded voice message to state clearly at the beginning who is responsible for the call, and to give a phone number for that business[7]. A homemade agent that rings guests back to confirm bookings needs all of that designed in; the FCC ruling on AI voices covers what it means for restaurants, and state AI disclosure laws covers the inbound side.
Card numbers are the third trap. PCI guidance says card security codes may not be stored after authorization, even encrypted, and a call recording that captures one must be securely deleted or made unrecoverable[8]. The cleanest design is not to take card numbers by voice at all: text the guest your ordering link and let your payment page do its job.
The monthly cost of running it yourself
Build estimates usually stop at launch. The recurring cost is the hours after it: reading transcripts, correcting answers, updating hours, retesting after a provider changes something, and being the person who gets the message when it breaks. Value that time at what the person doing it earns. The national median wage for food service managers was $33.36 an hour in May 2025[9].
| Input | Example |
|---|---|
| Calls the agent answers per day | 50 |
| Minutes to review each checked call | 2 |
| Share of calls you review (%) | 30% |
| Reviewer's hourly pay | $33.36 |
| Calls a month | 1,085 |
| Staff hours on these calls | 36 |
| Hours handed off | 11 |
| Monthly value of that staff time | $362 |
How it is worked out: Calls a day × 21.7 working days × minutes per call ÷ 60 × share handled = hours; hours × hourly cost.
Assumptions
- Hourly pay is the May 2025 national median wage for food service managers, before benefits.[9]
- Example: 50 calls a day, with 30% of transcripts read in the first months and two minutes to read each one and note a fix.
- Fixing prompts, retesting and handling outages come on top and are not counted here.
Example inputs. Use your own call volume and the pay of whoever will own the agent.
Example: on these inputs, review alone takes about 11 manager hours a month, roughly $360 at the median wage, before a single prompt is rewritten. A managed service does not make review go away; you should still read summaries every week. What it removes is the maintenance underneath: provider changes, outages, transfer logic and recording storage.
Then there is the on-call duty. Each provider in a homemade chain updates its service on its own schedule, and a change to the speech or voice layer can alter how the agent behaves on a Saturday you didn't plan to spend at a laptop. Somebody has to get the alert, know which part failed and switch the line back to the host stand while it is fixed. With a managed service, that person still exists on your side, but their job is to send a message, not to debug the stack.
Side by side
| Build it yourself | Managed service | |
|---|---|---|
| Time to the first real call | Building, then weeks of testing | Approve the answers, test calls, then forward |
| Who fixes it on a Friday night | You | The service's support |
| Control over wording | Every word, and every mistake | You approve the greeting and the answers |
| Handoffs and backups | Built and tested by you | Set from the manager path you choose |
| Recording notice and disclosure | Your legal homework | Built into a greeting you sign off |
| Call records | Wherever you store them | Summary by email, recording and transcript kept |
| Your own systems | Anything you can code | Signed webhooks to the systems you run |
For most single restaurants the right-hand column is the better fit. A group with developers on staff may weigh it differently.
- Answered2 events
- Request captured3 events
- Passed to staff2 events
- Confirmed by staff2 events
Activity
- Call
Kitchen hours
Caller asked when the kitchen closes tonight. Ava gave the approved time: 10 pm.
- Call
Table request · party of 8
Saturday, 7:30 pm, birthday. Saved as a request for the host to confirm; no table promised.
- Text
Ordering link sent
Caller wanted pickup and agreed to a text. Ava sent the restaurant’s official ordering link.
- Call
Allergy question → manager
Guest asked about a tree-nut allergy. Ava did not answer it and took a callback for the manager on duty.
- Web form
Catering inquiry · 40 guests
Office lunch next Thursday, drop-off, vegetarian options wanted. Sent to the catering manager.
- Call
Missing item on a delivery
Guest said a side was missing. Ava logged the order name and passed it to the manager with a summary.
- Call
Host confirmed party of 8
Host called the guest back and confirmed Saturday at 7:30 pm. Status updated in the inbox.
- Chat
Private room question
Website chat: rehearsal dinner for 22 in March. Event request created for the events lead.
- Call
Catering manager followed up
Catering manager called back, confirmed the date and sent menu options. Marked as quoted.
- 46Calls answeredExample
- 12Requests sent to staffExample
- 3Catering and event leadsExample
- 5Handed to a managerExample
Example only. Fictional business, people and numbers; not a recording or a live account.
When building it yourself makes sense
Some operators should build. A group with a developer on staff, its own ordering or booking system with an API, and an unusual workflow no service supports may get exactly what it wants from a build, provided someone owns it on call. That describes few restaurants. AI use rises with company size: 30% of US firms with at least 250 employees used AI in April 2026, against 17% of firms with fewer than 20[10], and 9 in 10 restaurants have fewer than 50 employees[11]. Most kitchens do not have a spare engineer.
A managed SI Employee covers the same ground without the build. You approve the house information and the manager path; it answers forwarded calls, takes reservation requests for your host to confirm, sends catering details to your catering manager, texts the ordering link when the caller agrees and hands complaints and allergy questions to a person. Groups that want call data in their own systems can receive each call by signed webhook (integrations). The three-restaurant group example shows how one manager path can serve three numbers.
If the attraction of building is mostly the price, read the limits of free AI receptionist tools first. Free credits and a weekend of building can look like the same thing as a working phone line until the first busy Friday.
Five questions before you start building
Who answers for it at 9 pm on a Saturday?
Name the person, and their backup. If the answer is “whoever notices”, nobody does.
Who reads the transcripts every week?
Someone who knows the menu and the policies, with time blocked in the schedule, not squeezed in after close.
Can you keep the greeting compliant?
Recording notice, automated-assistant disclosure and any outbound-call consent, for every state your guests call from.
Is there anything to connect to?
If your reservation or ordering system has no API, the agent can only take requests, which is what a managed service already does.
What does a bad week cost?
Price a weekend of missed bookings and wrong answers. If that number is larger than a year of a service, the build is not saving money.
Frequently asked questions
How long does it take to build an AI receptionist for a restaurant?
A first version that answers calls can take an afternoon. Testing it against real restaurant calls (noise, accents, interruptions, allergy questions, party bookings) and fixing what fails is what takes weeks, and monitoring never stops.
Can a no-code voice agent take restaurant reservations?
It can collect party size, date, time, name and number. Confirming a table needs a working connection to your reservation system and someone who owns it when it fails; without one, it is taking requests for your host to confirm.
Is it legal to put a homemade AI agent on our phone line?
Answering inbound calls is generally fine with a clear greeting and a recording notice where state law requires consent[5]. Calling guests back with an AI voice is treated as an artificial-voice call under the TCPA and needs consent[6].
What does a managed service still need from us?
Approved house information, the questions to ask for reservations and catering, a manager path with a backup, a round of test calls before forwarding, and someone who reads the call summaries each week.
Can we start with a managed service and build later?
Yes. Forwarding is a setting you can change, and the house information document you write is yours either way. Running a service first also shows you which calls a build would really need to handle.
References
National Restaurant Association (2026). Research Insight: The Future of Restaurant Technology (press release, November 5, 2026) (opens in a new tab) ↑
Proceedings of the National Academy of Sciences (via PubMed Central) (2020). Racial disparities in automated speech recognition (Koenecke et al.) (opens in a new tab) ↑
US Food and Drug Administration (2026). Food Allergies (opens in a new tab) ↑
Legal Information Institute (Cornell Law School), US Code (2026). 18 U.S. Code § 2511 - Interception and disclosure of wire, oral, or electronic communications prohibited (opens in a new tab) ↑
Digital Media Law Project, Berkman Klein Center (2026). Recording Phone Calls and Conversations (legal guide, no longer updated) (opens in a new tab) ↑
Federal Communications Commission (2024). FCC Makes AI-Generated Voices in Robocalls Illegal (news release) (opens in a new tab) ↑
Legal Information Institute (Cornell Law School), eCFR text of FCC rule (2026). 47 CFR 64.1200 - Delivery restrictions (opens in a new tab) ↑
PCI Security Standards Council (2018). Information Supplement: Protecting Telephone-Based Payment Card Data (v3.0) (opens in a new tab) ↑
US Bureau of Labor Statistics (2025). Table 1. National employment and wage data from the Occupational Employment and Wage Statistics survey by occupation, May 2025 (opens in a new tab) ↑
Federal Reserve Bank of Minneapolis (analysis of Census Bureau BTOS) (2026). AI adoption in business grows steadily but unevenly (opens in a new tab) ↑
National Restaurant Association (2025). National Statistics: Restaurant Industry Facts at a Glance (opens in a new tab) ↑