AI and restaurant technology
AI phone ordering for restaurants: a practical evaluation guide
A welcoming voice is the start of a useful phone agent. The harder test is whether it takes the right order, knows when information is missing and gives the caller a dependable next step.
Separate answering a call from completing an order
An answering assistant may provide hours and directions or send a link. An ordering agent also has to understand menu choices, validate the request, record an order and explain what happens next. Ask a provider which of those actions is actually connected to your restaurant’s systems.
Make the acceptance point explicit. A caller saying “yes,” an order appearing at the register and payment being confirmed are separate events. The agent should report the event that actually happened. It should never fill a missing confirmation with confident-sounding reassurance.
Jooule’s phone agent uses restaurant-scoped menu and order tools. Its available actions depend on configured telephony, AI and ordering services. Payment links and staff transfer need their respective supported connections. Plan a controlled rollout for the actual restaurant instead of treating a sample call as universal readiness.
Give the greeting a job
Use a short greeting that identifies the restaurant, makes the assistant’s role clear and invites the caller’s request. A useful opening could be: “Thanks for calling. I’m the restaurant’s ordering assistant. How can I help?” Adapt the wording to the restaurant’s character, then listen to how it sounds aloud.
Ask one necessary question at a time. Acknowledge the answer and continue without repeating the whole order after every item. When a detail is ambiguous, offer a narrow clarification: “Did you mean the small or the large?” Long speeches make it harder for the guest to correct the detail that matters.
Jooule can configure the restaurant’s greeting and tone. Callers can request a supported language change, subject to the configured language and voice support. Test the change with menu names, numbers and addresses; fluency in the greeting does not prove that an order in that language is correct.
Make uncertainty visible before the kitchen receives it
Prepare the same required choices, prices and availability the restaurant uses for other ordering. A caller can say “the usual,” ask for an unavailable dish or request a combination the menu does not support. The agent should check its tools, ask for the missing detail or explain the limit.
An ingredient question deserves particular care. A menu label or an inferred recipe is not proof of how food is prepared or whether cross-contact can occur. Define an escalation route for allergy concerns and requests the available information cannot resolve. Do not let a conversational answer become an invented safety guarantee.
Keep each restaurant’s identity and menu separate. In a demonstration, ask about an item known only to another restaurant and confirm that the agent does not borrow it. Repeat this after a language change: switching language must not switch the business, prices or ordering rules.
Offer a relevant extra once, then respect the answer
A useful suggestion fits the order and the moment. Offer an available side after the main item is clear, state any relevant price through the configured menu and stop when the caller declines. Do not interrupt a correction, an allergy discussion or an urgent request to reach a person with a sales pitch.
Jooule exposes a setting for whether the agent offers an extra suggestion. Evaluate the entire exchange: a recommendation should not silently enter the basket, and a refusal should not prompt the same suggestion in different words. The guest should know which items they actually accepted.
Finish with a concise read-back of consequential details: items and important choices, fulfillment method, timing and the payment next step. Ask for correction before creating an irreversible misunderstanding. Keep confirmation wording tied to the system’s returned state.
Rehearse payment failure and reaching a person
For payment links, test the complete path: the right recipient receives the right order’s link, the supported provider returns the result and the register reflects it. A sent message is not a completed payment. An abandoned or failed payment needs an accurate explanation and an agreed restaurant follow-up process.
A human handoff also has two ends. Configure the destination, call it during a rehearsal and test what happens if nobody answers. Tell callers what the assistant can actually do next. Do not promise a transfer merely because a transfer tool was requested.
Verify your recording, retention and disclosure arrangements with the people responsible for the restaurant’s policies and applicable requirements. Use designated test identities and the agreed payment test environment while rehearsing. A test plan should not accidentally place real kitchen work or charge a real customer.
Review each call before expanding the rollout
Use a small sequence that grows in difficulty. Write the expected menu, total, order state and handoff result before calling. After each call, compare what the guest heard with the saved order and tool results. Fix a failure and repeat that case before moving on.
Score ordering accuracy, appropriate clarification, successful correction, honest uncertainty, payment state and staff handoff separately from warmth. Track elapsed call time and service cost without inventing a target that ignores your menu. A passing sample lowers uncertainty; it cannot establish that future calls will never fail.
- Simple pickup with no extras.
- Required choice followed by a correction.
- Two similar dishes with different modifiers.
- Unavailable item and unsupported special request.
- Requested supported language switch midway through ordering.
- Allergy question requiring staff judgment.
- Suggested extra accepted, then a separate call where it is declined.
- Payment link sent but not paid; then a successful test payment.
- Human transfer answered and unanswered.
- Dropped call or repeated request: inspect whether duplicate work was created.
