The Voice-AI Bubble Starts Deflating

I spent the last week of June with a notebook full of voice-AI demo recordings, three cold coffees, and the growing suspicion that something has snapped in this category. Not the technology — the technology is real and getting better every quarter. What’s snapped is the gap between what voice-AI vendors are pitching to operators and what their systems can actually do in a Tuesday lunch rush with a contractor truck idling at the speaker post.

I want to say this directly in the lead, because I think the rest of the summer’s coverage is going to dance around it: the voice-AI bubble within the broader voice-AI category is bursting. The category is real. The bubble is the breathless “fully autonomous drive-thru by Q4” pitch deck that operators have been handed for two straight years. That deck doesn’t survive contact with a customer ordering eighteen thousand waters as a joke, and we are now far enough into the deployment cycle that we have receipts.

This is a Vibe Check desk review, not a benchmark. I’m pulling what’s public, what vendors are telling me off the record, and what operators say when the camera is off. Mark interpretation — I’ll flag where I’m reading tea leaves versus press releases.

The summer the marketing finally cracked

If you’d asked me in January, I would have said the voice-AI narrative was on a one-way escalator. Funding numbers were up, vendor logos were multiplying on every restaurant tech conference slide, and at least three different CEOs were quoted saying their drive-thru AI was “indistinguishable from a human” — a phrase that should require regulatory disclosure at this point.

Then McDonald’s quietly wound down its IBM partnership in 2024, the clips of misorders went viral, and the floor of the conversation shifted. Restaurant Business’s survey of the voice-AI vendor landscape — published as the field crowded past a dozen named players — reads less like a starting gun and more like a triage chart. There are now too many vendors chasing the same handful of enterprise pilots, and the operators I talk to are starting to compare notes on which ones are real and which ones are PowerPoint.

That’s the bubble. The escalator stopped, and the people who were standing closest to the top edge are the ones with the loudest landing.

The category, meanwhile, keeps going. PYMNTS reported voice-AI funding surged 8x year-over-year to roughly $2.1 billion on the back of CB Insights data, and that capital isn’t going to evaporate. It will consolidate, route around the misfires, and find the niches where voice-AI moves the needle — on terms that look much less like “replace the drive-thru employee” and much more like “augment the order-taker and stop promising autonomy you can’t deliver.”

If you’re an operator reading this, the implication is straightforward. The vendor pitches you’re getting in July are the last ones that will sound this confident. By the fall, the surviving vendors will be talking about hybrid deployments, narrow domains, and accuracy thresholds with error bars — because the operators have started asking for error bars.

What McDonald’s actually told the industry

Let’s deal with the McDonald’s-IBM ending first, because it’s the load-bearing event in this whole story and I keep seeing it misread.

McDonald’s ended its IBM partnership on automated order taking in 2024. The reporting framed it as a setback, the company framed it as a pause-and-reset, and Restaurant Business’s follow-up assessment of whether AI drive-thru is inevitable was — to my read — the most honest piece of trade press in this category in two years. The headline says “only a matter of time.” The body says, between the lines, “but the time is much longer than the vendors are telling you, and the first generation of partnerships is not going to be the one that wins.”

Here’s what I think McDonald’s actually told the industry by walking away. They told everyone that the world’s most operationally sophisticated QSR, working with a tier-one enterprise AI vendor, could not get the accuracy curve to bend fast enough to justify network-wide deployment. That’s not a vendor problem. That’s a problem with the implicit assumption that voice-AI drive-thru is a one-vendor, one-stack, one-rollout deal.

Which is exactly the assumption Wendy’s seems to have abandoned, and we’ll get there. But the McDonald’s lesson is the first thing every operator I talk to brings up unprompted, and the smartest ones are using it as cover to slow their own pilots down and ask harder questions. Good. That’s the right response. I’d rather see twenty cautious pilots in 2025 than five reckless rollouts.

Taco Bell’s throttle, and what it means

Taco Bell has been the second most-watched program in the category, and the signals out of Yum! this spring and summer have been notably more measured than they were a year ago. The team there isn’t pulling the system — they’re throttling the rollout, scoping deployments more carefully by daypart and store volume, and being noticeably less effusive about “fully autonomous” in earnings commentary.

I’m reading this as a sign of organizational maturity, not failure. Taco Bell built voice-AI into its tech narrative early, but the people running operations there are the people who have to make speed-of-service numbers in a real store with a real labor model. If they’re throttling, it’s because they ran the math on the next 500 stores and the math didn’t pencil at the speed the deck implied.

That’s the bubble deflating in real time. The category survives. The “fully autonomous, every store, by Christmas” line does not.

I’ll have more on the Yum! AI stack in a forthcoming desk review (and on how this compares to Chipotle — see the Chipotle AI stack writeup, a forthcoming May piece). The takeaway: when the most aggressive program in the category goes from gas to brake, every operator behind them gets permission to ask harder questions.

Wendy’s is the methodical case study

Wendy’s is the other side of the coin, and I think it’s the program operators should actually be studying. Restaurant Dive reported Wendy’s plans to deploy FreshAI to 500+ locations by year-end 2025, paired with digital menu boards as part of an integrated drive-thru refresh.

Two things stand out about the Wendy’s approach that I haven’t seen as clearly elsewhere.

First, the program is bundled with menu board hardware. That’s not a coincidence. It tells me Wendy’s leadership understands that voice-AI accuracy is partly a UI problem, not just an NLU problem. If the customer can see the order being built on a screen, the system can catch its own errors and the customer can correct them without escalating to a human. That’s a hybrid design, not a replacement design.

Second, 500 stores is a real number but it’s not a network. Wendy’s has roughly 6,000 US units. 500 is meaningful penetration but it’s still a “we are running this as a managed program, not a rollout” number. Compare that to the earlier marketing in this category where vendors were quoting full-system numbers and you can see the discipline.

I’m bullish on Wendy’s program in a way I’m not bullish on the category as a whole. Mark interpretation: they appear to be measuring, iterating, and not over-promising. That’s a much better signal than a press release with the word “autonomous” in it.

White Castle: the quiet leader nobody’s writing about

White Castle is the most interesting story in voice-AI that almost no one outside trade press is covering. By public reporting they are at roughly 30% of footprint deployed with SoundHound’s voice-AI. Thirty percent of a chain — even a small one by QSR standards — is the highest deployment penetration I’m aware of in US fast food.

A few things worth noting. White Castle has been on this for a long time — their pilot dates back several years, which means their operations team has had time to actually integrate voice-AI into shift workflow, exception handling, and labor model. The maturity shows in how they talk about it: their executives don’t pitch autonomy. They pitch accuracy improvements, reduced cognitive load on the order-taker, and consistency.

That framing is the one that survives the bubble. If your voice-AI vendor is talking to you about reducing cognitive load on a human order-taker, you are having a 2026 conversation. If your vendor is still talking about removing the order-taker entirely, you are buying 2023 marketing in 2025.

The Canopy critique is the one to read

If you only have time to read one third-party analysis of where voice-AI is right now, make it Canopy’s piece on AI drive-thru problems. It’s a vendor-adjacent perspective, so adjust your priors, but the issue tree is the cleanest summary of operator pain points I’ve seen this year.

The Canopy framing breaks down into three buckets that I think are roughly right:

• Recognition accuracy in adverse conditions. Truck idling, multiple voices, weather, kids in the back seat. The accuracy delta between a quiet booth demo and a real drive-thru in July is enormous and the marketing rarely concedes this.
• Menu complexity and modification handling. “No onions, extra pickle, light cheese, can you sub the fries for a salad” is where these systems break, and customization is exactly what differentiates a chain from a vending machine.
• Exception escalation UX. When the system can’t handle a request, what does the handoff to a human look like? Most early deployments make the customer repeat the entire order, which is worse than having had no AI in the first place.

The third bucket is the one I’m watching hardest. The vendors that solve exception escalation gracefully — invisible handoff, no order repetition, no audible “let me get a human” — are the ones that will survive this shakeout. The ones that don’t will be acquired or wound down by next summer.

The 45% number that should make every CEO pause

Here’s the consumer data that I think gets underweighted in every voice-AI deck I’ve ever seen. In the 2024 QSR Magazine / Intouch Insight survey of roughly 1,500 consumers, 45% said they did not like the idea of AI-enabled voice in the drive-thru.

That’s not a fringe number. That’s nearly half of your customer base telling you up front that they have a negative prior on this technology. And consumer priors matter in a category where speed-of-service and customer satisfaction scores are the two metrics every operator obsesses over.

A 45% negative prior doesn’t mean you can’t deploy voice-AI. It means you have to deploy it in a way that doesn’t make the negative prior worse. Concretely:

• The system has to be at-or-above human accuracy from day one, because customers who don’t like the technology will not extend it the benefit of the doubt the way they would extend it to a new employee.
• The handoff to a human, when it happens, has to be invisible. Any customer-perceived “this AI failed and now I’m starting over” interaction will reinforce every negative prior in your market.
• Branding the deployment matters. White Castle and Wendy’s both lead with operational benefits (“faster, more consistent”) rather than technological novelty (“you’re talking to an AI”). That’s not an accident. It’s a defensive posture against the 45%.

If you’re an operator and your vendor is encouraging you to make a big deal out of the voice-AI deployment in customer-facing marketing, ask them why. The data says quieter is better.

The funding signal is real, but read it carefully

PYMNTS’s reporting on the 8x year-over-year surge in voice-AI funding to roughly $2.1B is real and I’m not going to pretend it isn’t. But I want to be careful about how operators read that signal.

Funding surges in adjacent categories are not always good news for buyers. They are good news for the category in the abstract — more capital means more vendors, more competition on price, and more R&D. They are bad news for buyers in the short term because they incentivize vendors to make bigger claims than their tech supports, in order to capture market share before the next funding round.

We are, by my read, at the peak-marketing-claims point of this cycle. The funding has surged, the vendors are scaling sales orgs, and every deck has gotten more aggressive in the last six months. The next twelve months will sort the real deployments from the demo-ware, and the trough of disillusionment will hit somewhere in early 2026.

If you’re an operator, this is the moment to be cautious, not the moment to commit. The vendors that survive the trough will be there in 2026 with sharper products and more realistic pricing. The vendors that don’t survive — and there will be several, given the count Restaurant Business published — will leave behind orphaned deployments that some other operator will have to clean up.

What the surviving voice-AI category actually looks like

Let me sketch the version of voice-AI that I think is going to be the working model by mid-2026, because I want to be clear that I’m not bearish on the category — I’m bearish on the bubble.

The surviving model is hybrid. Voice-AI handles the high-volume, low-complexity portion of the order — the part that’s roughly the same every time, the modifiers it has been trained on, the menu items that don’t have a long tail of customization. The system runs in shadow mode first, then in assist mode (helping the human order-taker), then in lead mode with a human ready to take over.

The surviving model is bundled with menu board UX. The Wendy’s pattern. The customer sees the order being built on a screen as the system interprets their speech, which catches errors without making the customer feel like they’re correcting a robot.

The surviving model is honest about its accuracy curve. Vendors will publish accuracy thresholds, by daypart, by menu complexity, by ambient noise condition. Operators will measure those numbers themselves and hold vendors to them. The “indistinguishable from a human” line will die, and good riddance.

The surviving model is priced on outcomes, not seats. The vendors that win will charge for measurable accuracy and measurable speed-of-service improvements, not for “AI per drive-thru per month.” This is going to be a painful transition for vendor commercial teams, but it’s the only pricing model that survives a buyer base that has been burned by demo-ware.

And — this is the one I’m most confident about — the surviving model is operator-led, not vendor-led. The chains that win at voice-AI will be the ones that own the deployment playbook, not the ones that outsource the playbook to the vendor. That’s the lesson of every restaurant tech category I’ve covered. The vendor sells the tool. The operator runs the deployment.

The cross-category read

A quick note on how this fits into the broader restaurant tech landscape, because voice-AI doesn’t exist in a vacuum and the operators I talk to are pattern-matching across categories.

The Sweetgreen Infinite Kitchen story — see the Infinite Kitchen piece, an upcoming May piece — is a parallel cautionary tale on the production-automation side. Big promises, real technology, but the rollout pace is being calibrated by what the operations team can actually absorb, not what the marketing team can announce. That’s the right pace. Voice-AI deployments are starting to find the same rhythm, but they’re a step behind.

Toast, meanwhile — Toast desk review, a forthcoming desk review — is the platform layer that most of these voice-AI deployments will eventually have to integrate with at the POS edge. The integration story is going to matter enormously in 2026. The voice-AI vendor that has a clean Toast integration is going to win pilots that the standalone vendor loses, regardless of who has the better speech recognition.

The pattern across all three categories is the same: tech that works, deployed at a pace operators can actually manage, integrated into a stack that operators already trust. The bubble is the part of the marketing that pretends those constraints don’t exist.

What I’d ask vendors right now

If I were sitting in an operator’s chair this month, here are the questions I’d ask every voice-AI vendor that came through the door. Right now the questions in most RFPs are not hard enough.

• What’s your published accuracy threshold by daypart and ambient noise condition, and what happens to my contract if you miss it? If the answer is “we don’t publish thresholds,” the conversation is over.
• Show me a complete exception escalation flow, including what the customer hears when the handoff happens. If the customer hears anything that signals “the AI failed,” the design is wrong.
• How many of your stated deployments are in lead mode versus assist mode versus shadow mode? Vendor counts are routinely conflated across these modes. Pin them down.
• What’s your retention rate among chains over 100 units after 12 months? This is the number that separates real products from demo-ware. If the vendor won’t quote it, they don’t have one.
• What’s your integration story with the major POS platforms? If the answer is a roadmap, the deployment will hurt.
• What’s your roadmap for the modifiers and customization patterns specific to my menu? Generic models do not survive contact with chain-specific menus.

If the vendor can answer five of these with hard numbers and a believable artifact, they’re worth a pilot conversation. If they can’t, you’re looking at a 2024 product wrapped in 2025 marketing.

The forecast — narrow, hedged, and probably right

I’m going to commit to a few predictions for the back half of 2025 and 2026, because I think the category is mature enough now that the forecasts can be specific.

By year-end 2025, at least one of the named voice-AI vendors that’s currently active in QSR will exit the category, either through acquisition, pivot, or wind-down. The math doesn’t support the number of vendors currently chasing the same enterprise pilots, and the funding surge will accelerate consolidation, not delay it.

By mid-2026, the dominant deployment pattern will be assist mode rather than lead mode. The vendors pitching “fully autonomous” today will quietly reposition to “AI-assisted order taking” and the trade press will treat it as a sensible evolution, which it is, but operators should remember which vendors had to reposition and which didn’t.

By end of 2026, at least one tier-one QSR — not McDonald’s, but one of the next five — will publicly walk back a voice-AI deployment in a way that becomes a category event analogous to McDonald’s-IBM. I don’t know which one. I don’t think the vendors know which one. But the base rate on these things, at this point in a hype cycle, is roughly one major walk-back per year, and we are due.

And by end of 2026, the surviving voice-AI category will be larger, healthier, and more valuable than today’s bubble. The bubble bursting is the thing that makes the category real. We’ve seen this movie before with online ordering, with kitchen automation, with predictive labor scheduling. The trough of disillusionment is where the actual product gets built.

Where this leaves operators today

Concretely, in July 2025, here’s what I think operators should be doing.

Run a pilot, not a rollout. Five to ten stores, measured baselines, a clear exit criterion, and vendor compensation tied to accuracy targets. If the vendor won’t structure a pilot this way, they’re not a partner.

Bundle the deployment with menu board UX. Voice-AI without a screen for the customer to see the order being built is a worse product than voice-AI with one.

Hold the line on the 45% number. Don’t make the deployment a marketing event. Brand it on operational outcomes — faster, more consistent — not on technological novelty.

Watch Wendy’s. Watch White Castle. Don’t watch the loudest vendors. The signal is in the operator-led programs publishing real deployment numbers.

And finally — remember that the bubble deflating is good news for the category. The vendors that survive this summer will be the ones building the products you actually want to buy in 2026. The bubble was always going to burst. The question was whether it would burst before or after operators committed to long-term contracts they’d regret. We’re getting lucky on the timing. Use the luck.

I’ll be back with a vendor-by-vendor desk review later this summer. If you’re an operator running a pilot and have data to share — accuracy curves, exception rates, CSAT deltas — my inbox is open.

— Sofia leads Vibe Check vendor reviews for TableTransfers. Tips: vendors@tabletransfers.com.