Restaurant phone conversion isn’t one leak. It’s seven, and most operators only see two of them.

The phrase “missed calls” quietly hides the size of the problem. It treats phone-order conversion as a coin flip: the phone rang and either a ticket came out the other side, or it didn’t. The real funnel is seven stages long. Five of those stages live inside the call itself, and your carrier log can’t see any of them.

Matthew Diakonov, Written with AI

Published May 21, 20266 min read

Direct answer (verified 2026-05-21)

Where do restaurants leak phone-order conversion? At seven stages, not one. Ring-pickup accounts for roughly 35 percent of the visible leak at peak. The other 15 to 25 percent is split across answered-then-parked, mishearing during ticket entry, no readback, paper tickets that never get keyed into the POS, and no spoken pickup time. The seven stages are individually instrumented in PieLine’s open-source Denny’s call sample at src/components/voice-activity-data.ts, with timestamps you can verify against the audio at public/audio/dennys-order.mp3.

The two stages your carrier log shows you

Stage 1 is the ring. Stage 7 is the ticket landing in the POS. Everything you usually call “missed calls reporting” is really a comparison of those two numbers: how many rings showed up, how many tickets came out. The carrier log shows stage 1. The POS shows stage 7. The space between them is opaque.

The PieLine homepage ROI calculator (src/app/page.tsx, the RoiCalculator block) uses the conservative 35 percent peak miss rate published by Toast and Upserve. That number is real and a good starting point, but it only describes stage 1. By the time you read this paragraph, you may already be thinking the 35 percent is overstated for your shop because your tickets-per-call ratio looks fine. The next five stages are why that intuition is misleading.

The five stages between ring and ticket

Stage 2. Pickup vs. parked-hold hangup.

Host picks up, says “hold on one sec,” sets the handset down. The caller waits 25 seconds and hangs up. The carrier log records this as an answered call. The POS sees no ticket. From a reporting standpoint, this call is invisible: it isn’t in the missed-calls bucket and it isn’t in the tickets bucket. The PieLine concurrency ceiling of 20 simultaneous calls exists because this is the most common in-call leak in single-line shops at peak.

Stage 3. Heard vs. misheard.

Hood fans, fryers, expo calls, walk-in customers at the counter. A human listening to the phone with one ear is going to miss a modifier. Caller says “half pepperoni, no onions,” the host writes “pepperoni.” The order goes into the POS. The customer arrives and the order is wrong. This shows up in the remake count, not in the missed-call count. Most operators never connect the two.

Stage 4. Capture, with structure.

The Denny’s sample call in the PieLine repo shows what a clean capture looks like. Caption 4 is “Hi. Yeah. Can I get one Lumberjack Slam and one Coke.” Captions 9 through 12 are the AI asking “how would you like your eggs cooked, and what kind of bread? White, brown, multigrain, or sourdough.” Captions 13 through 15 are the customer answering “sourdough bread, scrambled eggs.” The structured cart has those modifiers attached to the parent item the moment the words land. The paper-ticket equivalent of this is a host writing “LJ Slam” on a pad and intending to ask about the eggs “when it’s quieter.”

Stage 5. Readback vs. no readback.

In the Denny’s call, the AI reads the entire order back between 75.42 and 83.34 seconds, then asks for confirmation. The customer confirms at 85.78 seconds. Most phone orders at most restaurants skip this step entirely because the host is in a hurry and the caller doesn’t ask. The cost shows up later, at pickup, when the customer says “I asked for no onions.” If you have never made a habit of reading orders back, this is the highest-leverage stage to fix without changing anything else in your operation.

Stage 6. Post to POS, immediately or eventually.

The POS round-trip in the Denny’s sample takes 2.40 seconds. Caption 41 at 89.12 seconds: “Placing your order now.” Caption 42 at 91.52 seconds: “Done.” That window is the entire post-and-ack against the integrated POS (Clover, Square, Toast, NCR Aloha, Revel). In a human-handled flow this stage takes anywhere from 90 seconds to never. Tickets get scratched on a paper pad during the rush and entered after the rush ends, by which point a few of them have been lost, mis-keyed, or assigned to the wrong customer name. This is the second-largest source of invisible loss after stage 2.

The 7-stage phone-order funnel

1. Ring

Inbound call lands on the main line

2. Pickup

Handset answered before caller bails

3. Listen

Order heard over kitchen noise

4. Capture

Items + mods + name written down

5. Readback

Order repeated for confirmation

6. Post

Cart sent to POS, ticket cut

7. Confirm

Total + pickup time spoken back

Stage 7, which actually closes the loop

The final stage is the one that prevents a stage-3 or stage-4 leak from showing up at pickup. The AI in the Denny’s sample announces the total ($34.11) at 92.00 seconds and the pickup window (12:45 AM) at 95.12 seconds, both read straight from the POS response. The caller hangs up at 101.81 seconds knowing the price they will be charged and when their food will be ready. The human equivalent in most shops is “it’ll be like twenty minutes” followed by the customer showing up in fifteen and standing in the lobby. Or showing up in thirty and finding the food cold. Without stage 7, every upstream success can still feel like a bad experience.

102.36s

“The seven stages are visible inside one 102.36-second captured call in the PieLine repo. POS round-trip between caption 41 and caption 42 is 2.40 seconds. Full transcript with start and end times: src/components/voice-activity-data.ts.”

PieLine open-source Denny's sample, verified 2026-05-21

Audit your own funnel without buying anything

Three data sources you already have, mapped to the seven stages. Run this against last week and you will know which stage is the heaviest leak in your shop. Knowing where the leak is matters more than picking a tool to fix it.

One-week funnel audit

Stage 1 leak: pull the carrier log, count inbound calls under 10 seconds in the 11am to 1pm and 5pm to 8pm windows. Those rang out before a human got there.
Stage 2 leak: count inbound calls with a connect time between 1 and 20 seconds. Picked up, said one word, hung up. The host parked them.
Stage 3 leak: ask anyone at the line what percent of phone orders need the customer to repeat. If they laugh, that is the number you cannot measure in the carrier log but is bleeding tickets.
Stage 4 leak: spot-check yesterday's phone tickets against the kitchen's remake count. Wrong modifiers, missing add-ons, dropped allergies all show up here.
Stage 5 leak: ask three regulars whether anyone read their order back last week. Most no-readback orders never get flagged until pickup.
Stage 6 leak: look at the gap between when phone tickets enter the POS vs. when the call ended. Anything over 90 seconds is staff hand-keying after the rush, which means some tickets never make it in at all.
Stage 7 leak: count phone customers who showed up early or late. No spoken pickup time means they guessed.

Why this framing matters

If you read the operator-facing literature on this topic, almost every article published in the last three years frames the problem as “X percent of restaurant calls go unanswered” and stops there. The X varies by source (Toast 32, Upserve 36, Hostie AI 36, our own 35) but the framing doesn’t. What the framing misses is that an answered call is not the same thing as a captured order. A captured order is not the same thing as a posted ticket. A posted ticket is not the same thing as a satisfied pickup. The unit of analysis matters.

Once you start thinking in stages instead of in a binary, two things become true at once. First, your real leak is bigger than your missed-call number. Second, fixing the leak does not require fixing every stage at the same time. A no-readback shop can recover material upside just by reading orders back, with zero new tools. A shop whose biggest leak is stage 6 (paper tickets getting lost between the rush and the end-of-day POS sweep) can recover most of it by training one host to enter the POS ticket before hanging up. The funnel breakdown is what turns the problem from “buy something” into “here is the specific change to make this week.”

Want help mapping your own 7-stage funnel?

Bring a week of carrier-log exports and a week of POS phone-ticket timestamps. We will sit with you for fifteen minutes and tell you which stage is leaking the most.

Frequently asked questions

Why call it a conversion leak instead of just missed calls?

Because missed calls only describe stage 1 of the funnel. PieLine's homepage StatsStrip (src/app/page.tsx line 152) cites 35 percent of restaurant calls going unanswered at peak. That number is real, but it does not include the answered calls that hang up on hold, the ones where the host wrote 'large cheese' when the caller asked for 'large cheese, half pepperoni, no onions,' or the orders that get scratched on a paper ticket and never make it to the POS. When you add those, the gap between calls placed and tickets the kitchen actually cooks is larger than the headline miss rate suggests.

How big is the leak between the headline miss rate and the actual funnel loss?

Pulled together from operator-published numbers and our own customer benchmarks: of the 65 percent of calls that do get picked up at peak, somewhere between 15 and 25 percent leak further down the funnel. Most of that loss is invisible because it never produces a missed-call entry in the carrier log. It produces a wrong ticket, a long hold that ended in a hangup, or a paper order that was never keyed in. The phrase 'missed calls' actively hides this category of loss, which is why most operators underestimate the size of the leak by a third.

Where in PieLine's public code can I verify the 7-stage funnel?

The whole funnel is visible in one captured call. The audio file is at public/audio/dennys-order.mp3 (1,229,949 bytes, 102.36 seconds stereo). The per-caption transcript with start and end times is at src/components/voice-activity-data.ts (75,855 bytes, 46 segments). Stage 1 happens before second zero. Stage 2 to 3 are captions 1 through 7 (0.00 to 9.36 seconds). Stage 4 is captions 9 through 31 (the eggs, the bread, the cheesecake, the strawberries). Stage 5 is captions 33 through 40 (75.42 to 86.81 seconds, the readback). Stage 6 is the 2.40 second window between caption 41 (89.12 seconds, 'Placing your order now') and caption 42 (91.52 seconds, 'Done'). Stage 7 is captions 43 through 44 (92.00 to 97.84 seconds, the $34.11 total and the 12:45 AM pickup time).

How do I run the audit on my own restaurant without buying new software?

Three sources you already have. Carrier call log gives you stages 1 and 2 (rings, connect duration). POS phone-ticket timestamps cross-referenced against the call log give you stage 6 (the gap between hangup and ticket cut). Direct conversations with your top regulars give you stages 3 to 5 and stage 7 (whether they had to repeat themselves, whether anyone read back the order, whether they got a pickup time). The Yelp and Google reviews mentioning 'rushed,' 'forgot the,' or 'showed up and it wasn't ready' map cleanly onto specific stages. None of this needs a vendor. Most operators have just never sliced the data this way because the carrier log only surfaces stages 1 and 2.

Does putting an AI on the phone close all seven stages?

It closes them by construction. Stage 1 cannot leak because the line never rolls to busy; PieLine's published concurrency ceiling is 20 simultaneous calls. Stage 2 cannot leak because the AI picks up on ring one with no handset to set down. Stage 3 cannot leak because the audio is post-processed (Deepgram nova-3, multichannel, smart_format) and not heard over a hood fan. Stage 4 cannot leak because the cart is structured the moment the words land, with modifiers attached to their parent items. Stage 5 cannot leak because the readback is built into the flow; in the Denny's call it runs from 75.42 to 83.34 seconds. Stage 6 cannot leak because the POS post is a function call, not a paper hand-off, and the 2.40 second round-trip is the latency floor. Stage 7 cannot leak because the spoken total and pickup time are read from the POS response, not invented. Whether the answer to that should be PieLine specifically is a different conversation, but the funnel-level argument is structural.