Restaurant management and operations has six measurable channels. The phone makes seven.

FOH (covers, turns, wait time, hospitality moments), BOH (ticket times, station load, waste, temp logs), inventory (counts, pars, depletion, shrinkage), labor (schedule, hours, labor percent, overtime), POS and cost of sales (voids, discounts, COGS, modifier mix), and marketing and loyalty (repeat rate, campaigns, attribution). Every top guide you will find is some reshuffle of these six. They all assume the phone is either (a) a low-volume auxiliary to walk-ins or (b) noise that a cashier handles on the side. That assumption is load-bearing, and it breaks the moment the restaurant discovers that phone orders are 20 to 40 percent of a typical dinner rush.

Because there was no artifact. BOH leaves tickets. Inventory leaves counts. POS leaves rows in a database. The phone left a customer who either hung up, complained, or placed an order that a human cashier re-entered by hand. A manager could not open the phone channel on Monday morning and read it. You could only ask the staff how it went, which is the same as not managing it. PieLine's voice-activity-data.ts changes that. Every call becomes a 60 Hz amplitude envelope plus speaker-tagged captions, which is readable the same way a ticket screen is readable.

The file is src/components/voice-activity-data.ts in the pieline-phones repo, regenerated by scripts/build-voice-activity-data.py against public/audio/dennys-order.mp3. Duration is 102.36 seconds. Sample rate is 60 Hz. Each channel (customer, AI) is an array of 6,157 floating-point RMS values. The file also contains 46 speaker-tagged captions with start and end timestamps. Peak amplitude for the customer channel is 0.8857; peak for the AI channel is 0.9067. If you run the build script locally with your own stereo WAV, the resulting TypeScript file has the same shape, which is why this works as a management artifact and not a one-off marketing asset.

A cross-talk frame is a 1/60th-of-a-second slice of the call where both the customer envelope and the AI envelope are above 0.1 amplitude at the same time. In plain terms, both parties are speaking. In human-staffed phone ordering, cross-talk frames go up under pressure because the cashier starts interrupting to keep the line moving, which is where modifier misses originate. In the shipped reference call the cross-talk count is exactly 0 across all 6,157 frames. A manager can pull that number from any call the product processes. It is the first numeric KPI the phone channel has ever had.

A silent frame is a 1/60th-of-a-second slice where both envelopes are under 0.02. The reference call has 1,922 silent frames out of 6,157, which is 31.2 percent. That silence is think time: the AI working through menu mapping, the customer deciding, the pause before an upsell. Managers reading the phone channel want silence to stay in a band. Too little means the agent is rushing and skipping clarifiers; too much means the agent is stalling on a menu it does not have mapped. The number is only useful because it is comparable across calls at 60 Hz resolution.

AI speech totals 49.04 seconds; customer speech totals 19.76 seconds. The remainder is silence and hand-off framing. That ratio (2.48 to 1, AI to customer) is high on purpose: the AI is reading back the full order twice and narrating the POS placement, both of which exist so the manager can audit the call against the seven structural segments. A cashier who interrupted would compress both sides and lose the audit trail. The manager treats the ratio itself as a metric, not just the total duration.

Add one line to the review agenda called Phone (with KPIs: answer rate, cross-talk frames, silent frame percent, segment coverage). On Monday, open five sampled calls from the prior week in the dashboard. For each, tick the seven structural segments (greet, intent, clarifier, resolve, upsell, readback, POS placement) and check whether the cross-talk count is still 0 and the silent percent is in its band. The whole exercise is ten minutes. It replaces the old agenda line that said Phone etiquette, which had no measurable output.

Yes. The envelope-and-caption artifact is per-call and per-location. Multi-unit operators aggregate by location and by daypart, the same way they already aggregate FOH covers. The data is small (roughly 6,000 floats per channel per 100 seconds of audio, plus the captions), so rolling it up across 10 or 100 locations is not a data-engineering problem. The management practice is new, but the infrastructure is not.

The AI agent that answers the phone 24/7 across up to 20 simultaneous calls, POS integration with Clover, Square, Toast, NCR Aloha and Revel, menu scraping and modifier mapping, same-day onboarding, and a dashboard that exposes the envelope, captions, and KPIs per call. Pricing is $350 per month for the first 1,000 calls and $0.50 per call after. The data artifact is a byproduct of the product doing its job, not a separate service.

A raw recording is not management infrastructure. It is a 102-second MP3 that a manager would have to listen to end to end. The envelope and the captions compress the call into a shape a human can read in seconds: where the speakers overlapped, where there was silence, which of the seven segments the agent actually covered. That is the difference between a video file and a dashboard, and it is why the phone can now be the seventh channel in the weekly review rather than the one everybody ignores.