Restaurant technology, evaluated by asking for the receipts

Any software, hardware, or service that reduces the cost, error rate, or staff attention required to convert a hungry person into a completed, paid order. That definition is deliberately narrow. It excludes category-sounding words like 'digital transformation' and 'connected experience' that do not tie to a measurable operational metric. If a vendor cannot point to which number on your P&L their product moves, and in which direction, their product is not restaurant technology yet; it is restaurant marketing material.

Because restaurant technology evaluations collapse the moment you ask to hear one working. The file at public/audio/dennys-order.mp3 in the aiphoneordering.com repo is a real AI-handled phone order, recorded in stereo with the customer on the left channel and the AI on the right. We transcribed it with Deepgram's nova-3 multichannel model and shipped the captions, per-channel 60 Hz amplitude envelope, and the regeneration script alongside the audio. The AI answers 'Hi. This is Denny on a recorded line.', takes a Lumberjack Slam with scrambled eggs and sourdough, upsells cheesecake with the line 'It might make your Coke jealous.', and closes at a $34.11 pickup order in 102.36 seconds. Any operator can download the file, run the script, and reproduce the transcript.

At a phone-ordering restaurant in 2026, the stack has seven layers. Layer one: the phone carrier that routes inbound calls (Twilio, Telnyx, your existing VoIP). Layer two: the speech-to-text service that transcribes the caller (Deepgram nova-3 multichannel in our case, alternatives include Whisper and Rev). Layer three: the reasoning model that understands the order and decides what to say next (a tuned LLM with a restaurant-specific prompt, menu, modifier tree, and rules). Layer four: the text-to-speech service that speaks back. Layer five: the POS integration that writes the finished order into Clover, Square, Toast, NCR Aloha, or Revel. Layer six: the payment processor when the order is a phone-paid pickup or delivery. Layer seven: the analytics plane that tracks concurrency, missed calls, upsell rate, and average handle time. Every row in a restaurant technology buyer's evaluation should ask a specific question of one of those seven layers.

Layer five, the POS integration. Speech-to-text is a commodity. Large language models are commodities. Text-to-speech is a commodity. The specific thing that blows up in production is writing the finished order, with modifiers, into the restaurant's POS. Half-and-half pizzas, spice levels, protein substitutions, allergy notes, and custom sushi rolls are not flat fields in a POS schema; they are nested modifier trees whose shape differs by POS vendor and sometimes by store. PieLine ships 50+ live POS integrations and 95%+ order accuracy because we spent more engineering on that layer than on any of the others. If a restaurant technology vendor will not let you place a live test call that lands in your own POS during the sales cycle, assume that layer is not ready.

PieLine's tested ceiling is 20 simultaneous calls per location. That is not a marketing number; it is the concurrency the infrastructure is qualified against and the graceful-degradation point beyond which calls queue rather than drop. Several competitors claim 'unlimited concurrent calls.' In practice concurrency is bounded by upstream speech-to-text rate limits, POS API rate limits, and the vendor's own compute. Any operator evaluating restaurant technology voice products should ask for the measured peak, what the behavior is at ceiling plus one, and whether the number was measured on a weekday afternoon or on a holiday rush. We publish ours and we will run a load test on demand.

Six items, each with a verifiable artifact the vendor must produce. One: a playable recording of a completed end-to-end order in the same cuisine as yours. Two: a word-level transcript of that recording generated by a named speech-to-text model, not edited by hand. Three: a screenshot of the resulting order inside the restaurant's POS, including every modifier. Four: the measured peak concurrency number and the behavior at ceiling plus one. Five: a printed price in dollars per month or dollars per call, not 'contact sales.' Six: a named reference customer in the same cuisine or format that will take a ten-minute phone call. A vendor that produces all six in twenty-four hours is real. A vendor that produces zero is a pitch deck.

One: public/audio/dennys-order.mp3 is in this repo and plays in any browser. Two: the Deepgram nova-3 multichannel transcript is in src/components/voice-activity-data.ts with 46 captions, 28 from the AI and 18 from the customer, at 60 Hz per-channel envelope sampling. Three: the pickup-order line 'Your total is $34.11, and your order will be ready for pickup at 12:45AM' is audible in the recording and present as a caption. Four: 20 simultaneous calls per location, published. Five: $350 per month for up to 1,000 calls, $0.50 per call beyond, money-back first month. Six: Mylapore (11-location South Indian chain in the Bay Area), Idly Express, Amber India, and China Village are our cuisine references; Mylapore's owner Jay Jayaraman speaks publicly on . We hit six of six on our own checklist, which is why we wrote it this way.

Yes, in the rollout, not in the artifact. A 100-location chain buys the same voice stack as a single Mexican shop, but the chain also buys a rollout playbook: POS version reconciliation across stores, menu-variance handling, multi-brand number routing, and a central analytics view. Independents install in a day. Chains install in a quarter. The receipts checklist above is identical for both buyer types; only the reference call is longer. Mylapore is rolling PieLine across 11 locations and projecting $500 per location per day in recovered phone-order revenue, roughly $2 million per year chain-wide once the rollout completes.

Anything that requires large hardware investment before it produces the first dollar and anything marketed with the word 'platform' before it has a single named live customer in your cuisine. Self-order kiosks are legitimate but have a 3 to 6 month payback once you count buildout and menu engineering. Robotic prep equipment works at scale but has a 6-month-plus time-to-first-dollar. Blockchain for supply chain, NFT loyalty, and 'generative AI menu development' without a measurable metric are the obvious overhypes. The boring truth is that most of an independent restaurant's 2026 technology ROI comes from plugging revenue leaks (missed calls, missed reservations, retyped orders, unreconciled delivery aggregator revenue) that the existing categories already address.

Because the phone channel is the leak that is already happening on the day you sign. Restaurants miss 30 to 40 percent of inbound calls during peak hours, each one is an order someone else captured, and the fix is a 10-minute phone-forward setup with no hardware, no staff training, and no change management. Week one of operation typically captures enough previously-lost orders to exceed the $350 monthly subscription. Every other category on a restaurant technology roadmap (loyalty, kiosks, KDS, robotics) requires staff behavior change or physical installation; the phone channel does not. That is why we rank it first and why Mylapore rolled it out before their other tech investments.