The Drive-Thru AI Scoreboard at Midyear

I spent Saturday ordering at three drive-thrus inside a forty-mile radius — Wendy’s, White Castle, Bojangles — just to hear who picked up. All three were machines. None stumbled on a basic combo. One nudged me toward a Frosty I had not asked about. Six months after the IBM-McDonald’s partnership unwound on every trade-press front page, voice AI at the speaker post has stopped being a pilot category and started being a deployment one — and the operators who got the architecture right last year are putting real distance between themselves and the rest.

The contrarian thing worth saying at midyear: the post-IBM era is producing winners and losers, and the gap is becoming structural.

The scoreboard, as of this week

Four numbers are doing most of the work. Wendy’s FreshAI is on track to reach more than 500 restaurants by year-end, up from roughly 160 in January — the largest single-brand voice-AI rollout in QSR. White Castle’s “Julia,” a SoundHound partnership running since 2020, is now in about 100 lanes, roughly a third of the chain. Bojangles’ “Bo-Linda,” built by Hi Auto, is live in approximately 200 of the brand’s 800-plus locations. And CKE Restaurants has told the trade press Presto Voice is producing a 6% sales lift in the lanes where it runs — an attribution number unusually clean for this category.

Read those four data points together and the takeaway is obvious. The brands that bet on a single voice-AI architecture in 2023 — and stuck with it through the embarrassing accuracy stories of 2024 — are now compounding. The brands still in vendor selection are looking at a market where the leaders have a year of in-lane recordings, menu-mix retraining, and crew muscle memory their challengers do not have.

What the leaders share

The leaders look more alike than the headlines suggest. Three things keep showing up:

• Single-architecture commitment. Wendy’s runs Google Cloud under FreshAI; White Castle runs SoundHound under Julia; Bojangles runs Hi Auto under Bo-Linda. None is multi-sourcing the voice layer. The brands stuck below 5% deployment are, almost without exception, still piloting two or three vendors against each other.
• A clear escalation path to a human. Every leader I have visited routes ambiguous orders to crew in seconds, not interactions. Mark this as interpretation: the brands that treat the AI as a first responder rather than a final answer are the ones whose accuracy stories have stopped going viral.
• Menu discipline at the speaker post. Wendy’s pruned modifier sprawl before scaling FreshAI. Bojangles standardized combo numbering before turning Bo-Linda on. The AI does not have to be smarter if the menu does less.

The laggards, by contrast, are running into the failure modes Canopy catalogued in its midyear piece on AI drive-thru problems — accent miscues, modifier collisions, the long tail of off-menu requests. None unsolvable. All easier with a year of in-lane data than with three months.

Why the gap is structural, not temporary

Voice AI in the drive-thru is a data-advantaged category. The leaders are training on lane recordings the laggards do not have, tuning against menu-mix shifts that have not hit the challengers’ P&L, and negotiating the next contracts from a position where the vendor can point to a working deployment at scale.

The thesis I keep landing on: by the time a five-store regional operator gets serious in Q4, the per-lane economics will look meaningfully worse than they did for the Wendy’s franchisees who signed in 2024. Not because the technology costs more — because the operator is buying a deployment that has to catch up rather than one compounding.

A forthcoming May piece on Sweetgreen’s Infinite Kitchen makes the point that automated make-lines hold their unit economics across wage floors because the labor schedule is fixed. The drive-thru voice layer is the same shape of bet. An upcoming May piece on Chipotle’s AI stack sharpens it: vertical specificity beats horizontal cleverness.

What to do this week

If you operate a QSR with more than 50 lanes:

• Pick the architecture before you pick the pilot. A six-month bake-off costs you a year of training data. Pick one, run it hard, renegotiate at scale.
• Audit your menu before you audit the AI. Modifier sprawl is the input the leaders eliminated. The accuracy gain is downstream.
• Ask your vendor for escalation latency, not recognition rate. Seconds-to-human is the metric the leaders manage to. Accuracy headlines lag it.

The IBM-McDonald’s unwinding made it easy to write the obituary for restaurant voice AI in 2024. The midyear scoreboard says the obituary was for one vendor pairing, not the category. The winners are pulling away. The question is whether the boardrooms still deliberating notice in time.

— Maya covers restaurant tech for TableTransfers. Tips: [email protected].