AI phone ordering that fills every table's takeout slot
A voice agent that answers every takeout call, takes orders straight into the POS, and upsells like a trained host, even at the dinner rush.
How the work was scoped
- Industry
- Restaurants & Hospitality
- Duration
- 9 weeks
- Cooperation model
- Fixed price
Client name withheld under NDA. Engagement details are shown to the extent our agreement permits.
Direct POS Order Injection
Injects confirmed takeout orders and payments directly into the POS without staff intervention.
The agent reads the Toast menu live, so an item eighty-sixed at six o'clock isn't offered at six-oh-one. Orders are built as structured line items while the caller talks, then written into the POS with payment attached as one transaction. The kitchen sees a ticket identical to one a staff member would have keyed, with the call recording linked for disputes.
- Live Toast menu: eighty-sixed items disappear at once
- Structured line items assembled during the call
- Tickets identical to staff-keyed ones, recording attached
The menu the agent reads from the POS: prices, tonight's specials, two dishes eighty-sixed and the exact wording used per dish, beside the ticket it wrote with payment in the same transaction and the recording linked.
Which upsell wording works: taken by wording and by hour across the last ten offers, what each added, the attach rates the suggestions come from, and the one-offer rule.
AI Upsell Engine
Suggests popular drinks, appetizers, and desserts based on order contents and live specials.
Upsells come from the venue's own attach-rate data, so it offers what actually sells alongside what's already in the cart. It suggests once, takes no for an answer and never repeats within a call. That rule came from the owner, after hearing what a persistent upsell does to a regular. Specials are pulled from the POS, so nothing is offered that the kitchen has stopped making.
- Suggestions driven by the venue's own attach rates
- One offer per call; no is final
- Specials read from the POS, never hard-coded
Dietary & Modifier Handler
One call turn by turn, each caller turn scored, with the modifiers matched to the POS tree, the one read back, and the tree nut allergy repeated verbatim, written to its own field and held until staff acknowledged it.
Understands complex order modifications, allergies, and custom preparation notes accurately.
Modifiers are matched against the POS's own modifier tree, so a custom request either maps to something the kitchen can actually make or is read back for confirmation. Allergy mentions are treated as a distinct class: they're repeated to the caller verbatim, written into the ticket's allergy field (never the notes), and flagged for staff acknowledgment before the ticket fires.
- Modifiers matched to the POS's own tree
- Allergies read back verbatim and written to their own field
- Staff acknowledgment required before an allergy ticket fires
What we were brought in to do
The restaurant lost takeout revenue every Friday because staff couldn't answer the phone and run the floor. We built a voice agent that takes orders straight into the POS.
A single venue doing about forty percent of its revenue as takeout, with one phone and no one free to answer it between six and nine on a Friday. The owner had counted a month of missed calls off the phone bill before commissioning anything. The other half of the cost was invisible in that number: every answered call was a server leaving a table mid-service.
AI & Integration
Where the old way broke
During peak hours, phone orders went to voicemail and walk-in service suffered whenever staff stopped to take a call.
Peak-hour calls went to voicemail, and takeout customers who reach voicemail order elsewhere instead of leaving a message. Worse, the calls that were answered came out of floor service: the venue was choosing between the order on the phone and the table waiting to pay, and losing money whichever it chose.
We deployed a natural-sounding voice agent that answers instantly, reads the live menu, takes and upsells orders, and drops them directly into the venue's POS with payment.
What we built together
- 01
Integrated with the POS for live menu, pricing, and order injection
POS integration went first because an agent that can't see the live menu will eventually sell something the kitchen has stopped making.
- 02
Tuned the agent's voice and upsell prompts per location
Each location gets its own voice and upsell prompts, tuned to what that venue's customers tend to add to an order.
- 03
Handled modifiers, allergies, and pickup times reliably
Modifiers match the POS's own modifier tree, and allergies are read back verbatim and written to their own field for staff acknowledgment.
- 04
Monitored and refined against real call recordings
Real recordings were reviewed weekly, with the owner deciding what had gone wrong. That's where the offer-once upsell rule came from.
Phase by phase
Phase 1: POS API Integration
Live POS Injection Engine
Integrated live menu data, pricing rules, and direct order injection APIs into the venue's POS system.
- POS Integration API
- Menu Sync Engine
- Payment Injection Bridge
Phase 2: Conversational Scripting
Voice AI & Upsell Prompt Engineering
Configured natural-sounding voice prompts trained to handle dietary allergies, item modifiers, and intelligent host upselling.
- Voice Prompt Architecture
- Upsell Script Matrix
- Modifier Parser
Phase 3: High-Noise Audio Tuning
Speech Recognition & Noise Filtering
Tuned speech-to-text noise cancellation filters to accurately process phone calls made from noisy environments during dinner rushes.
- Noise Filter Benchmarks
- Audio Model Tuning
- Speech Recognition Audit
Phase 4: Multi-Site Launch & Analytics
Restaurant Deployment & Call Logger
Deployed across multi-location restaurant sites with real-time call recording, transcript logging, and ticket analytics.
- Multi-Site Launch Clearances
- Call Log Dashboard
- Revenue Impact Tracking
The weekly review of real recordings with the owner's verdict on each, and the call where a second upsell to a regular became the offer-once rule.
Operational results after launch
+100%
Phone orders captured
+14%
Average ticket
0
Peak-hour missed calls
Phone orders captured is measured against the telephony log of inbound calls, so unanswered calls are in the denominator. Average ticket compares agent-taken orders with staff-taken orders over the same period, because the historic average had a different mix. Missed calls is a count, not a rate.
Client name withheld under NDA. Figures are approximate, drawn from the engagement’s own reporting.
About our collaboration
- 01
A cross-functional team of 5 worked on a fixed price basis over 9 weeks, covering Voice ordering agent, POS integration, Menu automation. We ran two-week increments, each one shippable, reviewed live before it merged. Decisions were recorded as they were made, so the reasoning survived the people who made it.
The agent was tuned against real call recordings weekly, and the owner, not a metric, decided what had gone wrong. The upsell rule (offer once, take no for an answer) came from that review in week two, after the owner heard what a persistent second offer sounded like to a regular.
What we'd carry into the next one
- 01
100% phone order capture eliminated lost takeout revenue during peak Friday dinner rushes.
The recovered revenue was almost entirely the peak hour: off-peak calls had always been answered, so the agent added nothing there and everything between six and nine.
- 02
Intelligent upsell prompts increased average ticket size by +14%.
The ticket rise came from consistency more than persuasion: the agent offers on every order, where staff offered when they remembered and had time.
- 03
Direct POS injection freed floor staff to focus on in-person dining guests.
The floor benefit was the one the owner hadn't priced in: nobody leaves a table to take an order now, and that showed up in table turn before it showed up anywhere else.
The order is built while the caller talks, not worked out afterward.
Step through a pickup order: the modifier that has to be read back, the allergy that gets its own field, the one upsell and its no, and the ticket that waits for a person before the kitchen sees it. Switch tabs, use the arrow keys once one is focused, or play the call.
Two dishes become two line items the moment they're said, priced from the live POS menu. Nothing waits to be parsed from a transcript after the call.
A worked example with an invented caller and menu. Offsets into the call are illustrative, not measured.
From a ringing phone to a ticket the kitchen can make
The POS integration went in first: an agent that can't see the live menu will eventually sell something the kitchen has stopped making. Everything else reads from it and writes back to it.
- 01 · TriggerThe takeout line ringsAnswered instantly, every call, including the six-to-nine rush when nobody on the floor could pick up.
- 02 · IngestSpeech to text, tuned for noiseRecognition tuned and benchmarked against calls made from noisy places during the dinner rush.
- 03 · EngineConversation, modifiers, upsellLine items built during the call; modifiers matched to the POS tree; one offer from the venue's own attach rates.
- 04 · StateLive POS menuPrices, specials and eighty-sixed items read live, so nothing is offered that the kitchen has stopped making.
- 05 · DeliverTicket into the POSOrder and payment written as one transaction, recording linked. Allergy tickets wait for staff acknowledgment.
What the agent won't get wrong on a busy Friday
An allergy never fires unseen
Allergy mentions are repeated to the caller verbatim, written to the ticket's allergy field (never the notes), and held for staff acknowledgment before the ticket fires.
No free-text guesses in the kitchen
Modifiers are matched against the POS's own modifier tree. A request that doesn't map to something the kitchen can make is read back to the caller for confirmation.
Only what the kitchen is making
Menu, prices and specials are read live from the POS, so an item eighty-sixed at six isn't offered at one minute past. Every ticket links its call recording for disputes.
Losing orders to a phone nobody can answer at the rush? Scope your build in 3 minutes.
Scope your buildNearby engagements
Web PlatformsHospitality POS that keeps working when the internet doesn't
A venue system built as a desktop terminal that owns its own data, so tabs stay open, dockets keep printing, and a dropped line is an inconvenience instead of a closed bar.
Restaurants & Hospitality · 26 weeks
Mobile AppsA delivery app built around fast decisions
A food delivery app where browsing, ordering, and live order tracking are built around quick decisions and a smooth checkout.
Restaurants & Hospitality · 14 weeks
AI & AutomationAn AI agent that qualifies leads the moment they land
A conversational agent that engages new inquiries instantly, qualifies budget and intent, books viewings, and hands warm leads to agents with full context.
Residential Brokerage · 8 weeks
Let's talk
Running a large platform, shaping a first MVP, or getting a product ready for a funding round? Tell us where you are. We'll shape the process around it, and stay with you after launch.
