The 18,000 water cups problem, explained by the medium it happened in.

Omilia, a conversational AI vendor that won a competitive RFP with Yum! Brands. The first live store was in Irvine, California in 2023; four more pilot locations followed that year, and by early-to-mid 2025 the Omilia Voice AI Solution was live at 650+ Taco Bell U.S. stores. Omilia's own case study quotes 890+ drive-thrus on the platform in its broader Yum! footprint, with 100% menu coverage and a 5-10% lift in check size. Yum! Brands also announced a separate partnership with NVIDIA in March 2025 for future AI work across its brands.

In August 2025 a customer recorded themselves ordering 18,000 water cups at a Taco Bell drive-thru running Omilia. The AI accepted the order with no sanity check, no rate limit, and no human handoff; the video went viral. A separate, equally viral clip showed the assistant stuck in a loop responding 'And what will you drink with that?' after a soda had already been ordered. Dane Mathews, Taco Bell's Chief Digital and Technology Officer, told the Wall Street Journal 'we're learning a lot' and announced a hybrid strategy: voice AI at some locations during some parts of the day, with staff advised to 'jump in as necessary'.

Three structural reasons. First, the lane has no durable identity for the caller: the only signal is the audio stream, and once the car pulls away there is nothing to reference. A phone call carries a phone number from the first millisecond, so sanity limits and blocklists can be applied before the AI even speaks. Second, the lane cannot tear down the session: you cannot 'hang up' on someone at a speaker, and they can keep talking over the AI indefinitely. Phones hang up. Third, the lane has no warm staff handoff with conversation context: staff inside the store are looking at a ticket, not listening to the call, so they cannot seamlessly take over when the AI loops or stalls. PieLine's Smart Call Transfer keeps the live audio, routes to a manager, and includes the full conversation transcript in the handoff. All three gaps are eliminated by the medium.

No. Mathews told the Wall Street Journal Taco Bell is reassessing where and when to run voice AI, not shutting it down. The public position is a hybrid: keep AI at locations and dayparts where it performs, monitor it more closely elsewhere, and have staff 'jump in as necessary'. The McDonald's comparison is often cited alongside this, because McDonald's pulled IBM's drive-thru AI after its own bacon-on-ice-cream class of incidents in 2024.

Omilia's publicly cited numbers are 100% menu coverage and a 5-10% check-size lift from dynamic upsell; order accuracy numbers have not been published at the level Hi Auto publishes for Bojangles (96% accuracy, 93% completion). Independent coverage suggests the real-world failure rate is dominated by three audio-side problems that do not appear on a phone line: engine noise, wind, and passengers speaking over the customer. Those are properties of the lane, not the speech model.

PieLine is a phone-channel voice AI. It answers the restaurant phone line, takes orders with cuisine-specific customization (half-and-half, spice levels, protein subs, combo mods), upsells, and posts the cart directly into Clover, Square, Toast, NCR Aloha, or Revel. It handles up to 20 simultaneous calls with 95%+ accuracy. When the call goes sideways, Smart Call Transfer hands the live call to a human with full conversation context, which is the specific countermeasure a drive-thru lane cannot build.

Three mechanisms that the lane does not have. First, caller identity: the phone number is logged before the AI speaks, so rate limits, blocklists, and return-caller context are trivial. Second, session control: the AI can warn, stall, or end the call when the request is clearly invalid (no one orders 18,000 of anything on a phone line). Third, the Smart Call Transfer path, which keeps the audio alive and lands the live conversation on a manager with the transcript attached. On a QSR drive-thru speaker none of these are possible; the speaker has no 'hang up', the microphone belongs to the vehicle, and no staff member is on the audio.

Yes. Cuisine-specific customization is the core of PieLine's first-month tuning: combos with substitutions, sides, sauces, protein swaps, spice and heat levels, and modifier stacks that look like a Taco Bell cart. The onboarding team scrapes the menu, maps each item to POS item IDs, and trains the AI on the restaurant's modifier schema. Complex modifiers are where generic voice bots fail; that is the capability PieLine is explicitly built around.

Taco Bell's voice AI program took roughly two years to move from one Irvine pilot to 500-plus stores, and even then it is being scaled back into a hybrid model. A single independent restaurant onboarded onto PieLine is typically live the same day on a supported POS (Clover, Square, Toast, NCR Aloha, Revel). The menu scrape, POS mapping, and first-month tuning are handled by PieLine's onboarding team.

PieLine is $350/month flat for up to 1,000 calls, $0.50 per call beyond that, with a money-back guarantee in the first month. Enterprise QSR deals (Taco Bell / Omilia, Bojangles / Hi Auto, White Castle / SoundHound, Wendy's / Google FreshAI) are multi-year platform programs with hardware, per-store install, and pricing that is not published per unit.

PieLine is a phone-channel product, not a drive-thru lane product, so a franchisee could in principle use it on their restaurant phone line for call-ahead and takeout orders. The drive-thru lane side of Taco Bell is governed by the Yum! Brands / Omilia program. The phone line is a separate revenue channel that typically leaks calls (missed 30-40% in peak hours) and is where PieLine is most obviously a fit.

Pick the channel that gives your AI the tools it needs to fail gracefully. The lane is the hardest possible place to run voice AI because it strips the medium of identity, session control, and warm handoff. The phone line is the easiest, because every one of those tools is already built in. The 18,000 water cups clip is a near-perfect demonstration of what 'the medium will not help you recover' looks like in production.