AI Voice Agent for Restaurant Orders: Guide Fish Audio
Plan restaurant phone orders and reservations with menu lookup, allergen warnings, payment boundaries, barge-in, human transfer and call QA.
An AI voice agent for restaurant orders answers the phone, reads the menu, captures a bounded order or reservation and hands the result to your POS or order-management system. Fish Audio supplies transcription, streaming text to speech, a consistent voice and API access. Your restaurant systems own the menu, prices, availability, modifiers, prep times, payments and allergy information. There is no native Fish Audio POS integration, and the agent must not invent a dish, price or allergen claim.
Phone orders are a high-noise, high-consequence workflow. The agent's job is to be clear, repeat the order and stop when the request leaves the menu.
What should the agent handle?
A safe first scope:
- Menu questions: items, descriptions, available sizes, modifiers and current price.
- Pickup or delivery orders: item, quantity, modifiers, name, phone and requested time.
- Reservations: party size, date, time, seating preference and contact details.
- Hours and location: opening times, address, parking and accessibility notes.
- Order status: whether the kitchen has accepted the order and the estimated prep time.
Keep these out of the automated flow:
- Allergy safety guarantees.
- Refund or complaint decisions.
- Large or catering orders that need kitchen approval.
- Payment card capture.
- Anything requiring a manager's judgment.
For the support-side equivalent, see AI voice agents for customer support. For delivery exceptions after an order, see AI voice agents for ecommerce.
What data does the agent need?
Connect the agent to narrow, read-only endpoints where possible:
| Data | Use | Boundary |
|---|---|---|
| Menu and modifiers | Read items, sizes and prices | No invented items or prices |
| Availability | Mark sold-out items | Re-check at confirmation |
| POS or order API | Create and read orders | No direct database access |
| Prep-time feed | Give an estimate | Label it as an estimate |
| Reservation system | Read and hold tables | No overbooking |
| Allergen source | Read approved warnings | Never guarantee safety |
| Store hours and location | Answer basic questions | Read from one approved source |
Cache stable content such as the menu, but re-check availability and price before confirming an order.
What should an order call sound like?
Agent: Thanks for calling Maple Street Kitchen. Pickup or delivery tonight?
Caller: Pickup. Can I get the chicken bowl with no sesame?
Agent: The chicken bowl is available. I have to tell you the kitchen handles sesame and cannot guarantee an allergen-free preparation. Would you like me to connect you with a staff member to discuss that?
Caller: I'll just take it as is, but no extra sauce.
Agent: One chicken bowl, no extra sauce. Anything else?
Caller: A large miso soup.
Agent: One large miso soup. Your total is $21.40, ready in about 20 minutes. Pay at pickup. Can I confirm the name and phone number?
The agent reads the warning, records the caller's decision without endorsing it, repeats the order and gives an estimate.
How should allergen questions be handled?
Never say a dish is safe. Read the approved allergen information, state that cross-contact is possible in the kitchen and offer a transfer to a trained staff member. Do not improvise ingredient lists or accept a caller's assertion that a dish "should be fine."
Record that the warning was read and whether the caller chose to continue. If the caller asks for a guarantee, stop and transfer. This is one of the few places where a short call is better than a completed order.
How should payments be handled?
Keep card data out of the conversation. Use one of these patterns:
- Pay at pickup or on delivery.
- Send a secure payment link by SMS.
- Transfer to a staffed line with an approved payment terminal.
- Use a telephony provider's secure payment capture flow.
The agent should not repeat, store or log a card number, security code or bank credential. If a caller offers a card number, interrupt and redirect to the approved payment path.
How should the agent handle noise and interruptions?
Restaurant callers are often in cars, kitchens or crowded rooms. Confirm each item and quantity, spell unusual names and read the order back before committing. When the caller interrupts, stop speaking and listen; the real-time text to speech guide covers barge-in and streaming.
If the caller cannot be understood after one clarifying attempt, offer a transfer or a callback. Guessing an item creates a remade order, a refund and a worse experience than a short hold.
What are the common failure modes?
| Failure | Safe fallback |
|---|---|
| POS API times out | Take the order details and offer a callback or transfer |
| Item sells out during the call | Re-check availability and offer alternatives |
| Caller asks for an allergy guarantee | Read the warning and transfer to staff |
| Caller is unintelligible | Confirm once, then transfer or offer callback |
| Price or total is unclear | Read the confirmed total from the POS response |
| Caller wants a refund or complains | Transfer to a manager |
| Kitchen prep time changes | Give a new estimate and label it clearly |
Log the POS response with the order. If the order was created but the confirmation failed, keep the order and tell the caller it is confirmed.
Which metrics should you track?
| Metric | What it tells you |
|---|---|
| Order completion rate | Whether the flow converts calls into orders |
| Order accuracy rate | Whether items, modifiers and quantities are correct |
| Remake or refund rate | The cost of misheard orders |
| Average order value | Whether the agent handles modifiers and upsells correctly |
| Abandonment rate | Whether waits, noise or confusion lose callers |
| p95 time to first audio | Whether the conversation feels natural |
| Transfer rate | Whether allergy and exception boundaries are working |
Review transcripts for the items that are misheard most often, then add pronunciation notes and confirmation prompts.
How much does restaurant-order speech cost?
Natural speech is roughly 150 words per minute, close to 900 characters. One Fish Audio credit covers about 1,000 characters, so agent speech starts near one credit per minute before repetitions, holds or longer menu reads.
| Pack | Price | Credits | Approximate agent speech |
|---|---|---|---|
| Starter | $9 | 1,000 | About 16 hours |
| Creator | $29 | 4,000 | About 66 hours, plus commercial rights |
| Studio | $99 | 15,000 | About 250 hours, plus API access |
Telephony, POS or reservation software, payment processing and staff time remain separate. The TTS API pricing guide converts characters into a full budget.
Key takeaways
- Fish Audio supplies speech; your menu, POS and reservation systems own the facts.
- Never guarantee an allergen-free meal.
- Keep card data out of the voice conversation.
- Repeat the order and confirm the total before committing.
- Re-check availability and price at confirmation.
- Transfer when noise, allergy or an exception makes the call unsafe.
Frequently asked questions
Can a voice agent take restaurant orders?
Yes, when it is connected to your menu, availability and POS or order-management APIs. Fish Audio supplies speech; it does not provide a menu database, POS integration or order routing on its own.
Can the agent guarantee an allergen-free meal?
No. The agent should read the approved allergen information, state that cross-contact is possible and transfer allergy questions to a trained staff member. It must never promise that a dish is safe.
Should the agent take card payments?
Do not collect card numbers in the voice conversation. Use pay-at-pickup, pay-on-delivery, a secure payment link or a staffed payment flow, and keep the agent outside the cardholder-data boundary.
How should the agent handle a noisy call?
Confirm each item and quantity, slow down, repeat the order back, and offer a transfer when the caller cannot be understood. A short confirmation is cheaper than a remade order.
How much does restaurant-order speech cost?
At roughly 150 words per minute, one minute of agent speech is about one Fish Audio credit. POS, telephony, payment processing and staff time are separate costs.
Related guides
- AI voice agent for ecommerce - order status, returns and delivery exceptions.
- AI voice agent for appointment booking - reservations and calendar boundaries.
- AI voice for customer support - escalation and interruption handling.
- IVR voice generator - phone menus, hold messages and voicemail.
- Telephony text to speech API - PCM, SIP and phone audio boundaries.