Technology in the restaurant industry, told through one 102-second call.

POS and kitchen display, online ordering, self order kiosks, contactless and QR payments, loyalty and CRM, AI phone answering and drive thru, reservations and waitlist platforms, delivery aggregation middleware, workforce and scheduling, inventory and waste, and the analytics layer that ties all of that together. The interesting question is which of those categories you can actually hear working. A POS beeps, a kiosk dings, a kitchen display chirps. The phone is the one place where most articles list 'voice AI' as a trend, and almost no article tells you what it actually sounds like.

A 102.36 second AI-handled phone order recorded on a real production phone line and shipped in the PieLine repo at public/audio/dennys-order.mp3. The audio is transcribed with Deepgram multichannel diarization, the result is stored in src/components/voice-activity-data.ts, and that file is imported by the hero visualization on aiphoneordering.com. The call contains a greeting, a two item order (Lumberjack Slam plus Coke), clarifying questions on eggs and bread, an upsell, an accepted upsell with a strawberry modifier, a full readback, a POS write, a total of $34.11, and a 12:45 AM pickup time.

Because every restaurant operator has already read the trend lists. They know about AI, kiosks, and contactless. What they usually have not seen is what one unit of AI actually does in 102 seconds of real audio. The second-by-second walkthrough in this guide lets an operator compare what a modern restaurant voice stack handles to what their own phone line is doing tonight, and judge for themselves whether the gap is closeable in a quarter.

46 captions across 102.36 seconds, 30 from the AI and 16 from the customer. The file lives at src/components/voice-activity-data.ts and exports a typed object with duration, sampleRate, an envelopes object containing per-speaker amplitude arrays, and a captions array of speaker, start, end, and text fields. A comment at the top of the file says it is auto-generated from Deepgram multichannel transcription of public/audio/dennys-order.mp3.

The cheesecake upsell starts at 52.525 seconds with 'Before I finish up, would you like to add a sweet treat like a slice of New York style cheesecake?' and finishes at 59.44 seconds with the throwaway line 'It might make your Coke jealous.' The customer accepts at 62.465 seconds with 'Sure. Yeah.' and then adds a strawberry modifier at 65.985 seconds. That is a 10 second window from offer to accepted modification, and it is the part of the call that lifts the ticket from the base Lumberjack Slam plus Coke to the final $34.11.

'Placing your order now' is spoken at 89.12 seconds, 'Done' at 91.52 seconds. So roughly 2.4 seconds of wall clock between 'I am about to send this' and 'it is in.' The total wall clock from the customer finishing their first order sentence at 9.36 seconds to the POS write at 89.12 seconds is just under 80 seconds, including clarification, upsell, and full readback confirmation.

PieLine ships live integrations with Clover, Square, Toast, NCR Aloha, and Revel, and the homepage and llms.txt both claim over 50 POS integrations in total. Those five cover the overwhelming majority of independent and mid market installs in North America. The POS write in the Denny's recording is the last stage of the same pipeline, which is what lets 'Your total is $34.11' match what the kitchen prints for the cook.

The PieLine homepage ships a React calculator at src/app/page.tsx, lines 221 to 230, that takes calls per day and average ticket and returns monthly lost revenue. The fixed constant missedRate is 0.35. At 80 calls per day and a $35 ticket, the calculator returns $29,400 in monthly lost revenue. The Denny's call is the shape of what one recovered call looks like. Multiply that by the recovered share of 35 percent and that is the order of magnitude of the technology's effect, per location, per month.

The pickup time quoted in the transcript is 12:45 AM, which is one of the clearer signals that this kind of technology is not a daytime aid. Denny's, diners, late night Indian and Chinese operations, and 24 hour QSR chains are exactly where the technology pays back fastest, because those windows are when the line cook is on and a human phone answerer is not.

Yes. The audio file is served at /audio/dennys-order.mp3 from aiphoneordering.com because it sits in public/audio/ and Next.js serves the public directory verbatim. The hero clip on the homepage plays the audio with a dual channel envelope visualization driven by the same voice-activity-data.ts file this guide references. There is no paywall and no signup required to confirm what the technology sounds like.

A human taking the same order takes longer per call (closer to 3 minutes with upsell), handles one line at a time, and produces almost no structured data afterward. A dedicated phone host costs $3,000 to $4,000 per month fully loaded and still cannot pick up call two during call one. PieLine is documented as handling up to 20 simultaneous calls at $350 per month flat for the first 1,000 answered calls, which is the math that moves voice from a staffing problem to a technology problem.

Mylapore, an 11 location South Indian chain in the Bay Area, is rolling PieLine across all locations and projects roughly $500 in recovered revenue per location per day, roughly $2M per year at full rollout. Idly Express in Almaden runs more than 90 percent of calls end to end on the AI, with the remaining sliver transferred to staff with the conversation context attached. Amber India is onboarding. These are the reference installs cited on aiphoneordering.com and in the PieLine llms.txt.