A restaurant missed phone call is 14 seconds of dead air. Here is the 102 second call that replaces it.

Any inbound call to the restaurant’s main line that did not end in a human conversation or a completed order. That covers the obvious (ringing out, dropped to voicemail, hung up on hold) and the less obvious (answered but parked for more than a minute, answered and transferred to a line that also did not pick up, answered and the caller hung up because they could not hear over kitchen noise). Industry capture-rate studies treat anything that did not become a booking, an order, or a resolved question as a miss. The PieLine homepage ROI calculator at src/app/page.tsx lines 214 to 397 assumes a 35 percent miss rate on the primary inbound line during peak hours, which matches the Upserve and Toast-published benchmarks.

Most research converges on 12 to 18 seconds for a first-time caller at a quick-serve or casual concept, and 20 to 25 seconds at a destination restaurant where the caller is committed. That is roughly three to five rings before the caller decides the restaurant does not want their money. The exact threshold depends on the caller’s intent (browsing vs. placing an order they have already mentally committed to), whether they are on speakerphone in a car, and whether the restaurant has a voicemail greeting or rings out. Voicemail is only marginally better than silence: 85 percent of missed callers do not leave a message and do not call back, per the homepage StatsStrip at src/app/page.tsx lines 144 to 169.

Because the same rush that is killing the pickup rate is also filling whatever voicemail queue exists. A small restaurant phone system often gets configured with a very short hold window and no after-hours redirect, so during peak the inbound call goes straight to a busy tone rather than to voicemail. The owner sees zero missed-call entries in the log and assumes the phone line is quiet. It is not. It is blocked. The StatsStrip on the PieLine homepage cites 43 percent of calls going unanswered during peak, and 85 percent of those callers dialing a competitor rather than waiting.

PieLine ships a real captured example in its public repo. The audio file lives at public/audio/dennys-order.mp3 and is 1,229,949 bytes. The transcript and per-channel RMS envelope live at src/components/voice-activity-data.ts, which is 76,438 bytes. The file declares duration 102.36 seconds at 60 Hz sample rate, which produces 6,157 amplitude samples per channel, 12,314 total across the two channels. The caption stream has 46 segments. The AI greets at 0.00 to 3.44 seconds. The customer orders a Lumberjack Slam and a Coke between 5.36 and 9.36 seconds. The AI clarifies eggs, bread, and confirms the combo between 15.98 and 25.66 seconds. The customer answers with sourdough and scrambled eggs at 29.39 to 35.15 seconds. The AI upsells a New York cheesecake with a joke (“it might make your Coke jealous”) at 58.16 to 59.44 seconds. The customer accepts and requests strawberries at 62.47 to 68.31 seconds. The AI reads back the full order for confirmation at 75.42 to 83.34 seconds. The customer confirms at 85.78 to 86.81 seconds. The AI says “Placing your order now” at 89.12 seconds and “Done” at 91.52 seconds. That 2.4 second gap is the POS round-trip. The AI announces the $34.11 total and a 12:45 AM pickup window at 92.00 to 97.84 seconds, thanks the caller by name at 97.84 to 99.28 seconds, and the caller hangs up at 101.81 seconds.

The regeneration pipeline is a 4,709 byte Python script at scripts/build-voice-activity-data.py. It posts the stereo WAV to Deepgram’s v1/listen endpoint with model=nova-3, multichannel=true, punctuate=true, smart_format=true. Channel 0 is the customer, channel 1 is the AI. Words are grouped into caption segments on punctuation boundaries with a 0.55 second pause gap, a 3.2 second max duration, and a 75 character line limit. The amplitude envelope is computed per channel at 60 Hz using RMS over (sample_rate / 60) sample windows, then smoothed with an attack coefficient of 0.5 and a release coefficient of 0.12. Verify it yourself by cloning the repo, setting DEEPGRAM_API_KEY, and running python scripts/build-voice-activity-data.py public/audio/dennys-order.mp3.

That window is the POS round-trip. PieLine posts the confirmed cart to whichever of the integrated POSes is connected (Clover, Square, Toast, NCR Aloha, Revel, per the POSSection at src/app/page.tsx lines 876 to 935), waits for the ticket acknowledgement, and only then reads back the total to the caller. In a missed call that window never exists at all: nothing is posted, no ticket is cut, no kitchen display fires. In a poorly answered call that window exists but is unstructured: a human takes the order on paper, walks it to the POS after hanging up, and introduces an error rate that the StatsStrip benchmarks at 5 to 15 percent. PieLine’s published order accuracy is 95 percent or better, measured against the same set of complex modification orders.

Because a missed call produces no artifact. A missed text leaves an unread badge. A missed email sits in an inbox. A missed phone call, if it never reaches voicemail, leaves nothing at all: no KDS ticket, no POS void, no dashboard entry, no inbox row. Operators rely on proxies: the weekly revenue number being soft, a POS lookup showing fewer phone-source tickets on Friday, a Google call history that does not match the number of times staff remember hearing the phone. None of those proxies resolve to a specific missed call. That opacity is why the phone leak tends to outlast every other operational fix: you cannot sprint on a number you cannot see.

Only in a narrow case. Voicemail-to-text picks up the 15 percent of missed callers who actually leave a message. The other 85 percent hang up on silence and dial another restaurant. For a voicemail-to-text product to convert those 85 percent you would need a callback workflow, a staff member to hit the callbacks, and an average response time under roughly four minutes to preserve intent. In practice those callbacks happen an hour later when the rush is over and the customer has already ordered somewhere else. The only system that recovers the 85 percent is one that answers the call inside the first ring, and the only economic way to answer every inbound call at peak is an AI agent that can hold 20 concurrent conversations, which is the published PieLine concurrency ceiling.

The homepage at aiphoneordering.com has an ROI calculator whose math is visible at src/app/page.tsx lines 214 to 397. The model: daily missed revenue equals (phone calls per day) times 0.35 (miss rate) times (average ticket). Monthly loss is that times 30. Payback against a $350 monthly PieLine fee is (PieLine fee) divided by daily missed revenue. At 80 calls per day and a $35 ticket that pencils to a $29,400 monthly loss with a payback under one day. The numbers are deliberately conservative on miss rate (35 percent is the low end of published peak-hour figures) and on ticket (most restaurants underreport average phone ticket by 10 to 15 percent because their phone orders skew higher than dine-in).

Four things. First, pull your carrier call log for the past seven days and filter on inbound calls under 10 seconds. Those are the rings that were not answered and bailed on voicemail. Second, cross-reference those timestamps against your POS phone-source tickets. Any hour with more than a handful of short inbound calls and zero phone tickets is the hour the miss rate is spiking. Third, check your voicemail inbox count against the count of short inbound calls; the ratio is your voicemail capture rate (expect 10 to 20 percent). Fourth, if you are on a service with an answered-within-N-seconds metric, look for peak hours where that number sags below 85 percent. Those are the hours PieLine would be catching the calls that ring out.

The public repo ships the actual audio and transcript. The audio file is public/audio/dennys-order.mp3 (1.2 MB, 102.36 seconds stereo). The word-level caption data is at src/components/voice-activity-data.ts. The component that renders it live is src/components/voice-activity-clip.tsx (embedded on the homepage hero at src/app/page.tsx line 94). None of this is marketing footage; it is the same recording the homepage plays, with the stereo channels split so you can see the customer waveform and the AI waveform as independent envelopes.