What a restaurant AI voice agent actually does
Answers the restaurant's phone line, takes to-go and delivery orders straight into the POS, books or changes reservations, and routes anything it can't confidently handle to a live host.
A restaurant AI voice agent sits on a dedicated phone number and runs a loop built for the specific chaos of a dinner rush: pick up the call, play an AI and recording disclosure, classify whether the caller wants to order, book a table, or ask a question, then pull live menu data and item availability through a POS tool call before confirming anything. For orders, it reads back every item and modifier, captures payment through keypad tones or a texted link instead of ever hearing a card number out loud, and fires the confirmed ticket straight to the kitchen display system. For reservations, it checks the live book through OpenTable, Resy, SevenRooms or Tock and either confirms or offers alternate times. Anything outside that script, a large party, a complaint, a catering inquiry, transfers live with the order context attached.
The category moved from simple answering machines to real order-taking systems in 2026: Loman announced a direct Toast integration in July that reads the live menu, takes payment mid-call, and fires tickets to the POS and kitchen display without a human touching either system. That's a meaningful shift, because it means the agent's mistakes now show up as real tickets and real charges, not just a voicemail transcript. ConverseNow, the enterprise QSR player used by Domino's franchisees, deepened its own POS reach with an April 2026 Deliverect partnership and is an Olo integration partner.
Where these agents still fail: multi-item orders with stacked substitutions ("the usual but sub fries for a side salad, no onions, extra ranch on the side") get mis-heard more often in noisy dining-room or kitchen audio, large-party and VIP reservation requests need a human judgment call the agent can't make, and the moment a caller speaks a card number instead of using the keypad, that call enters PCI scope the agent was built to avoid. Taylor v. ConverseNow, a 2025 case over a Domino's phone order, shows the real liability when recording and data handling go wrong.
AI capabilities involved
Real-time voice conversation and order-taking (speech-to-speech)
Parsing multi-item orders with modifiers and substitutions
Menu, allergen and hours FAQ grounded in your actual menu data
Deciding when to transfer a call to a human host
Who uses this
- Independent restaurants and 2-5 location groups missing 30%+ of calls during dinner rush with no host stand coverage
- Quick-service and pizza franchises wanting phone orders to post directly into Toast, Square or Clover without a cashier re-keying them
- Full-service restaurants with heavy weekend reservation and waitlist volume relative to staff on the phone
- Multi-location groups already running Toast or Olo deciding whether to add a voice layer on top
What the restaurant voice agent handles (and what it shouldn't)
| Task | Agent handles it? | Systems it touches | Human role |
|---|---|---|---|
| Phone orders with modifiers ("extra cheese, no onions") | Fully | Toast, Square or Clover POS, menu database | None unless an item is 86'd mid-call |
| Taking payment during the call | With approval | Keypad/DTMF capture or a texted pay-by-link, Stripe/Square | Oversees PCI setup; agent never hears or stores a spoken card number |
| Reservations and waitlist | Fully | OpenTable, Resy, SevenRooms or Tock | None unless the book is full and needs a judgment call |
| Hours, menu, ingredient and allergen questions | Fully | Menu knowledge base | None |
| Catering and private-event lead capture | With approval | CRM or email, calendar | Events coordinator follows up on qualified leads |
| Large party or VIP reservation requests | No | Reservation system | Manager makes the call |
| Order changes or cancellations after the ticket fires | With approval | POS and kitchen display | Confirms with the kitchen before canceling |
| After-hours overflow and callback scheduling | Fully | Voicemail or callback queue | None |
Off-the-shelf AI agents on the market
Real products you can sign up for today — with current 2026 pricing, honest pros and cons.
Loman
Independent restaurants and small groups already running Toast
Demo only
From $199/mo flat; reported at around $299/mo per location for the full assistant
Custom for multi-location groups
Pros
- +Direct Toast integration reads the live menu and fires tickets to the POS and kitchen display
- +Takes payment mid-call instead of transferring to a human for checkout
- +Vendor reports up to 22% more phone revenue and 75-100 extra monthly orders per location for one customer
Cons
- −Pricing is reported inconsistently across sources; confirm the real tier before signing
- −Deepest integration is with Toast; other POS systems may need custom work
Slang.ai
Restaurants that want an established vendor with bilingual support
Demo only
Core $399/mo, Premium $599/mo (adds bilingual support and custom branding)
Custom
Pros
- +$68M in funding and 2,000+ restaurant locations by vendor count
- +Premium tier adds bilingual call handling out of the box
- +Established player with a longer integration track record
Cons
- −Priced per location flat, which adds up fast for multi-unit groups
- −Premium's bilingual and branding features aren't needed by every restaurant
ConverseNow
Multi-unit QSR and pizza franchises with enterprise-scale call volume
None; enterprise sales only
Contact sales; enterprise QSR pricing, not published
Deep POS integrations via an Olo partnership and an April 2026 Deliverect partnership
Pros
- +Built for high-volume quick-service and pizza chains, including Domino's franchisees
- +Vendor claims up to 25% same-store sales lift and 20%+ higher average ticket
- +Deliverect partnership extends reach into delivery-order aggregation
Cons
- −Enterprise-only: no self-serve pricing or small-restaurant tier
- −The subject of Taylor v. ConverseNow, a 2025 CIPA class action over call recording and card data at a Domino's location
Hostie
Restaurants comparing quotes across multiple vendors before committing
Demo only
Reported tiers conflict across sources: either $199/$399/$599 or $199/$299/$449 plus a $0.75/call overage
Not published
Pros
- +Runs a widely cited 500,000-call study claiming answer rates rose from 64% to 97% with its AI
- +Targets the same small-to-mid restaurant segment as Loman and Slang
Cons
- −Pricing is genuinely unverified: get a written quote before assuming either published tier is current
- −The 36%-missed-at-peak figure comes from the vendor's own study, not an independent source
The AI stack
A restaurant voice agent needs a realtime voice layer, a POS and reservation integration layer, a way to take payment without ever hearing a card number, and logging to catch mis-orders.
Voice (speech-to-text + text-to-speech, realtime)
Carries the actual phone conversation with low enough latency to feel natural during a live order
ElevenLabs Agents
$0.08/min on every self-serve plan, $0.16/min burst over concurrencyDIY builds prioritizing voice quality
Gemini 3.8 Live
Audio in ~$3/M tokens, audio out ~$12/M tokens (roughly $0.005-$0.018/min)Teams standardized on Google's ecosystem
Vapi orchestration + provider stack
$0.05/min platform fee, providers billed at cost (roughly $0.08-$0.13/min all-in)Custom builds and agency work
Our pick: Vapi or Retell for custom builds where you need provider flexibility; a packaged vendor (Loman, Slang) if you don't want to touch the voice stack at all.
Telephony
Routes the actual phone call to the agent and handles transfers to a live host
Vendor-native number (Loman, Slang, Hostie)
Included in the monthly planBuy-SaaS
Twilio or Telnyx via Vapi/Retell
Twilio inbound ~$0.008/min, Telnyx ~$0.0055/minCustom builds
Our pick: Keep your existing main line for humans and route a dedicated secondary number to the agent, so a dropped agent call doesn't take down your whole phone system.
Integrations / tools (POS, reservations, payment)
Lets the agent check real menu availability, book real reservations, and take payment without storing card data itself
Direct POS integration (Toast via Loman, Olo via ConverseNow)
Included on some vendor plans, custom work otherwiseBuy-SaaS
DTMF/keypad payment capture or pay-by-link (Stripe, Square)
Standard payment processing feesAny path taking payment by phone
Our pick: Never let the agent accept a spoken card number. Use DTMF capture or text a secure payment link every time.
Observability & evals
Catches mis-heard orders and missed upsells before they become a pattern of wrong tickets
Vendor dashboards (order accuracy, call volume)
IncludedBuy-SaaS
Langfuse or a logged transcript review
Free up to 50K units/mo, then $29+/moCustom builds
Our pick: Review a sample of real call recordings weekly against a short mis-order checklist, whichever path you take.
Reference architecture
The hardest engineering problem is accurate order capture under noisy real-world audio while keeping payment data out of the AI's path entirely: a mis-heard modifier produces a wrong ticket, and a card number that leaks into a transcript turns a convenience feature into a PCI and wiretap liability at once.
Call arrives
Telephony (Vapi/Retell/Twilio, or vendor-native)The call routes to the agent's dedicated number, with a recorded-call and AI-disclosure notice played first
Speech-to-text and intent classification
STT plus reasoning modelReal-time transcription feeds intent detection: order, reservation, FAQ, or complaint
Menu and POS lookup
Toast, Square or Clover integrationThe agent checks live item availability and pricing before confirming anything, catching 86'd items
Order confirmation
Reasoning modelThe agent reads the full order back, including modifiers, before closing out
Payment capture
DTMF masking or a texted pay-by-linkCard digits are entered by keypad or on a hosted page, never spoken or transcribed
Ticket fires to the kitchen
POS/KDSThe confirmed order posts directly to the kitchen display, the same as if a host rang it in
Human handoff
Host stand or manager lineLarge parties, complaints and anything low-confidence transfer live with full order context attached
Logging and eval
Observability toolEvery call and order is logged, with a weekly sample reviewed for mis-orders and missed upsells
Estimated cost per request
~$0.11-$0.18 per minute of call; roughly $0.30-$0.60 in model, voice and telephony costs per completed phone order at a typical 2-3 minute call
Cost calculator
Drag the sliders to model your actual usage. The numbers update in real time so you can stress-test economics before writing a single line of code.
Estimate your monthly cost at your own call volume and location count.
Estimated monthly cost
$344
≈ $4,128 per year
Calculator notes
- The vendor plan line assumes a $299/mo flat rate per location, in the middle of Loman and Slang's published ranges; ConverseNow and other enterprise players are quote-based and not reflected here.
- Model, voice and telephony cost applies on every path, whether you buy SaaS or build custom, since someone has to pay for the underlying reasoning and realtime audio.
- A DIY build swaps the flat vendor fee for a $0.08-$0.31/min all-in rate on Vapi, Retell or Bland (see build_vs_buy), which can be cheaper at low volume but ships without POS, payment and reservation integrations built in.
- Add standard card-processing fees on top of any path once you start taking payment by phone.
Build it yourself with vibe-coding tools
Vapi gets you a working restaurant phone agent in a few days: a dedicated number, an assistant with POS and reservation tools, and a transfer path to the host stand for anything it shouldn't handle alone.
Time to MVP
2-5 days for a first version, 2-4 weeks to get payment handling and POS sync production-ready
Total cost to MVP
$0-$100 (a Vapi account plus provider usage; no platform minimum until you need Pro-tier volume)
You'll need
Starter prompt
Create a Vapi assistant on a dedicated phone number using Claude Sonnet 5.5 as the model, with a knowledge-base tool loaded with your current menu, hours and allergen information, and an HTTP tool connected to your POS (Toast, Square or Clover) for live item availability and order submission. Give the assistant this system prompt: 'You are the phone assistant for {Restaurant Name}. Start every call with: "Thanks for calling {Restaurant Name}, this call may be recorded and you are speaking with an AI assistant, how can I help?" Handle three intents: taking a to-go or delivery order from today's menu, checking item availability with the POS tool before confirming anything, and reading the full order back with every modifier before closing it out; booking or changing a reservation for parties up to 6 using the reservation tool; and answering hours, menu, ingredient and allergen questions only from the knowledge base. Never ask for, repeat, or accept a spoken card number. When payment is needed, say "I will text you a secure payment link" or trigger the keypad-entry tool instead. For parties over 6, catering requests, complaints, or anything you are not at least 90% confident about, say "Let me connect you with someone who can help" and transfer immediately. Never invent a menu item, price, or hours that are not in the knowledge base. Keep responses short, warm and specific.' Add an OpenTable or Resy tool for reservations, route every transfer to the host stand's phone line, and log every call transcript and order to a Google Sheet or Postgres table for a weekly accuracy review.
Paste this into Vapi
Follow-up prompts (run in order)
- 1
Add a tool that checks today's 86'd (sold-out) items before confirming any order
- 2
Add a waitlist text follow-up when the dining room is full
- 3
Add a catering-lead capture flow that emails the events coordinator
- 4
Add a Spanish-language fallback for bilingual callers
Expected output
A phone line that takes to-go orders straight into your POS, books reservations, answers FAQ calls, and texts a payment link instead of taking a card number out loud.
Known gotchas
- !DTMF payment capture must be wired correctly, or card digits can land in the call transcript and pull the whole call into PCI scope
- !Noisy kitchen or dining-room background audio increases mis-heard modifiers; a mandatory order read-back before closing out cuts this significantly
- !POS menu syncs can lag, so the agent can quote an item that just sold out unless an 86'd-items check runs on every order
- !Large party and group bookings need a human judgment call the agent can't make alone
- !Call recording triggers all-party-consent laws in states like California, so the disclosure line at call start isn't optional
- !Self-hosting the voice stack means you own uptime: a dropped call during dinner rush is a lost order, not just a bad review
Compliance & risk reality check
A phone-ordering agent carries more compliance weight than a text chatbot because it records calls and, if you're not careful, touches card data.
AI voice disclosure
FCC Declaratory Ruling 24-17 holds that AI-generated voices are "artificial" under the TCPA, and California's BOT Act, Utah's AI Policy Act and Maine's LD 1727 all push toward telling callers they're talking to AI.
Mitigation: Play an AI and recording disclosure in the first sentence of every call, and answer "are you a bot?" truthfully if asked.
Call recording and CIPA wiretap risk
Taylor v. ConverseNow (N.D. Cal., Aug 11, 2025) let a CIPA class action proceed after a caller to a Domino's location alleged ConverseNow's AI recorded her name, address and card details without consent; Ambriz v. Google separately established a "capability" test where a vendor merely capable of using call data for training can be an unlawful eavesdropper.
Mitigation: Disclose recording at the start of every call, contractually bar the vendor from training on your call data, and prefer zero-retention settings where offered.
PCI DSS for in-call payments
Any card number spoken into the call pulls the STT, LLM, recordings and transcripts into PCI DSS 4.0.1 cardholder-data scope, fully mandatory since March 31, 2025.
Mitigation: Use DTMF/keypad capture or a texted pay-by-link instead of ever letting the agent hear or transcribe a card number.
Outbound calls and texts
TCPA consent rules apply the moment the agent calls a customer back or sends a marketing text, even though inbound-only order-taking is largely outside the TCPA's "making a call" rules.
Mitigation: Get written consent before any outbound confirmation, waitlist or marketing call or text, and honor opt-outs immediately.
Menu and allergen accuracy
An agent that answers an allergen question wrong carries the same liability risk that got Air Canada held liable for its chatbot's invented refund policy.
Mitigation: Ground every allergen and ingredient answer in your actual written menu data, never let the model guess, and route uncertain allergy questions to a human.
Build vs buy: the real math
3-6 weeks for a RapidDev build wired into your POS and payment flow; 2-5 days for a basic Vapi or Retell version
Custom build time
$13K-$25K for an agency build (up to $35K for multi-location plus POS and payment integration); $0.08-$0.31/min all-in for DIY on Vapi, Retell or Bland
One-time investment
Roughly 4+ locations, or call volume climbing into the low thousands of minutes a month, is where flat per-location SaaS fees start costing more than a custom build's amortized year-one cost
Breakeven vs buying
A single location on Loman's $199-299/mo plan spends $2,400-$3,600 a year, far below a $13K+ agency build. At 4 locations, Slang's Core plan at $399/location runs $19,152 a year before any POS add-ons, which starts to approach a one-time custom build you then own outright. A DIY Vapi stack at roughly $0.12/min blended and 1,000 minutes a month per location costs about $120/month per location in platform, model and telephony costs alone, cheaper than vendor flat fees at low volume but shipping without the POS, payment and reservation integrations a vendor builds in by default.
Skip the DIY — RapidDev builds the production version
A Lovable MVP gets you a demo. Production needs auth that doesn't leak data, AI calls that don't bankrupt you, observability when models drift, and code you can audit. That's what we ship through our AI app development service.
Discovery call (free)
30 minWe map your exact restaurant AI voice agent use case: who uses it, target volume, AI model choice, integrations, compliance scope. You get a detailed scope document and fixed-price quote within 48 hours.
AI-accelerated build
3-6 weeks for a RapidDev build wired into your POS and payment flow; 2-5 days for a basic Vapi or Retell versionOur engineers use Claude Code, Lovable, and custom tooling to ship 3–5x faster than agencies. You see weekly progress in a staging environment — not a black box.
Launch + handoff
1 weekWe deploy to your infrastructure, transfer the GitHub repo, set up CI/CD and monitoring, and train your team. You own 100% of the source code, prompts, and model configurations.
What you get
Timeline
3-6 weeks for a RapidDev build wired into your POS and payment flow; 2-5 days for a basic Vapi or Retell version
Investment
$13K-$25K for an agency build (up to $35K for multi-location plus POS and payment integration); $0.08-$0.31/min all-in for DIY on Vapi, Retell or Bland
vs SaaS
ROI in Roughly 4+ locations, or call volume climbing into the low thousands of minutes a month, is where flat per-location SaaS fees start costing more than a custom build's amortized year-one cost
30-min call. Fixed-price quote within 48 hours. No commitment.
Frequently asked questions
How much does an AI voice agent for restaurants cost?
Vendor plans run $199-$599/mo flat per location (Loman, Slang.ai), with enterprise QSR platforms like ConverseNow quote-based. A custom RapidDev build runs $13K-$25K upfront (up to $35K for multi-location or payment-integration work), plus $0.08-$0.31/min in ongoing voice and model costs. A DIY Vapi or Retell version costs $0.08-$0.31/min with no monthly minimum.
How long does it take to set up?
A vendor plan connected to your existing POS and menu takes 1-2 weeks. A custom agency build takes 3-6 weeks because it has to be wired into your exact POS, reservation system and payment flow. A basic Vapi version can answer calls within 2-5 days, though production-ready payment handling takes 2-4 weeks.
Can RapidDev build this for us?
Yes. RapidDev has shipped 600+ apps and offers a free 30-minute consultation to scope your specific POS, payment and call-volume needs before you commit to anything.
Will customers know they're talking to AI?
They should, and California's BOT Act, Utah's AI Policy Act and Maine's LD 1727 all push toward disclosure. The safest design states "you're speaking with an AI assistant" in the first sentence of the call and truthfully answers "are you a bot?" if asked.
Can it take payment over the phone safely?
Only if it never hears a spoken card number. The standard approach uses DTMF keypad capture or texts a secure pay-by-link, which keeps the call out of PCI DSS cardholder-data scope entirely; letting the agent transcribe a spoken card number does not.
Can it connect to my POS like Toast or Square?
Loman has a direct Toast integration that reads the live menu and fires tickets to the kitchen display, and ConverseNow connects through Olo and Deliverect. A custom build can connect to Toast, Square, Clover or any POS with an API, including ones no vendor supports yet.
What happens when the agent doesn't understand an order?
A well-built agent always reads the full order back before closing out, which catches most mis-heard modifiers, and transfers live to the host stand for anything it's not confident about rather than guessing and firing a wrong ticket.
Can it handle reservations too, not just orders?
Yes. Most vendors and custom builds connect to OpenTable, Resy, SevenRooms or Tock to check the live book and confirm or offer alternate times, while routing large-party and VIP requests to a manager.
Want the production version?
- Delivered in 3-6 weeks for a RapidDev build wired into your POS and payment flow; 2-5 days for a basic Vapi or Retell version
- You own 100% of the code
- AI cost monitoring built in
30-min call. No commitment.
