An AI phone agent for restaurants is just one real call, repeated. Here is the call, line by line.

It is the literal duration field on the call data file in this repo (src/components/voice-activity-data.ts), which was generated by running scripts/build-voice-activity-data.py against the recorded WAV. Deepgram multichannel was the transcription engine. The file declares duration: 102.36, sampleRate: 60, two amplitude envelopes (customer on the left channel, AI on the right), and 46 caption rows ordered by start time. Nothing about the breakdown on this page is reconstructed; every timestamp is read off that file.

The five gaps between the customer finishing a turn and the AI starting its next one are 1.60s, 2.41s, 2.41s, 3.03s, and 2.31s. The longest single pause inside an AI turn is 4.14s, which sits between "one moment, please" at 11.84s and the next AI word at 15.98s. That is the menu lookup window. We do not pad the gaps; the line of code that prints the captions reads start and end times directly from Deepgram’s word output.

Yes. At 65.98 seconds the customer says "Can you add strawberries, if that’s an option?" Strawberries are not a baseline component of a New York style cheesecake, and the question is phrased tentatively. Three seconds later the AI confirms "One slice of New York style cheesecake with strawberry topping." The full order summary at 75.42s includes "a New York style cheesecake with strawberry topping" before asking for confirmation. That is what handling a non-menu modifier looks like in production.

"Before I finish up, would you like to add a sweet treat like a slice of New York style cheesecake? It’s so good. It might make your Coke jealous." The agent says it after the customer gave their name and confirmed the order was complete. The line is two sentences and 9.5 seconds long. The cheesecake landed on the ticket. PieLine’s baseline is a 15 to 20% lift in average order value from the upsell pass.

Because the order was placed late at night. A live human at most restaurants is not picking up the phone at that hour, and at chains that do, the phone is the lowest-priority queue. PieLine handles the call regardless of clock time. The agent quoted "12:45AM" as the ready-by time at the end of the call. The reason it can quote that is the same reason it can place the order: the POS integration knows the kitchen’s prep time and current load, not because we hard-coded it.

Up to 20 simultaneous calls on a single restaurant line. A human phone employee handles one. That is the entire reason a peak-hour phone bottleneck collapses: every caller hits the first ring, every order goes into the POS as a line item, every modifier including strawberries on the cheesecake is captured. The 20-slot capacity is also why the answer for a Friday at 7pm is the same as the answer for a Tuesday at 12:45 AM.

Clover, Square, Toast, NCR Aloha, and Revel out of the box, with 50+ POS integrations available. The order lands as line items with modifier IDs. In this Denny’s call, that means a lumberjack slam row with two modifiers (eggs scrambled, bread sourdough), a soft drink row, and a cheesecake row with a strawberry-topping modifier. Not a free-text note for the cook to interpret. The agent quoted $34.11 because that is what the POS calculated, not a guess.

$350 per month covers up to 1,000 calls. Beyond that it is $0.50 per call. A dedicated phone employee runs $3,000 to $4,000 per month and handles one call at a time. Onboarding (menu scrape, POS mapping, dish description build) is included. There is a money-back guarantee on the first month if the 95%+ order accuracy or 90%+ AI-handled rate does not hold for your restaurant.