Buying Voice AI in 2025: A Field Guide for Operators

I spent last Tuesday sitting in a Hyundai in a drive-thru lane in suburban Dallas, parked three cars back from the speaker post, watching an operator client toggle between two phones. One phone had a Loom recording of a SoundHound Polaris demo his regional VP forwarded over the weekend. The other had a calendar invite for a Presto pitch the next morning and a Kea email asking if he wanted to be on the early-access list for a self-service tier. He looked at me, looked at the speaker post, looked back at me, and said: “Sofia, I just want to know which of these I’m allowed to buy without it becoming the next handheld disaster.”

That is the question. So this is the answer — or at least, the field guide I wish I’d been able to hand him before he got in the car.

The category got crowded fast. Restaurant Business counted seven serious voice-AI vendors actively pitching multi-unit operators heading into 2025, and that count was already stale by the time the piece hit. SoundHound’s next-generation platform launch on February 25 reset the anchor on accuracy claims. GTC is two weeks out and at least three vendors have told me they’re saving announcements for the floor. The Wienerschnitzel-Presto rollout is rumored to drop any day. Kea’s anticipated next-tier launch is on operators’ calendars. By the time you finish reading this, somebody on this list will have shipped something I haven’t seen.

What I can do is tell you how to think about the seven names that are actually showing up in operator RFPs right now — and the five newer names you should put on a watch-list but should not, under any circumstances, be writing a check to in March.

Vibe Check verdict at the top, because that’s what you’re here for: the anchor brands win on the math. SoundHound is the only credible answer at chain scale. Presto is the right answer if your concept is in the middle of restructuring its tech stack and you need a vendor with margin discipline. Kea is the indie and small-regional unlock. Everyone else is watch-list. I’ll show my work.

The category in three buckets

Voice AI for restaurants is not one market. It is three markets that share a microphone, and they sort by where the voice runs and what it’s allowed to do.

Bucket one: drive-thru order-taking. The headline category. Voice replaces (or augments) the human at the speaker post. SoundHound, Presto, and Hi Auto play here. This is the hardest technical problem in the category — open menu, ambient noise, accents, modifiers, upsell logic, POS handoff — and it’s where the accuracy claims get loudest and the receipts get hardest to find.

Bucket two: phone ordering and reservations. Voice replaces the host or the order-line phone. Kea is the clear leader. ConverseNow plays here too, alongside its drive-thru pitch. This is the easier technical problem — quieter audio, slower cadence, single caller — but the operations problem (call routing, hours, integration with the POS or reservation system) is the actual moat.

Bucket three: in-store and curbside voice. Smallest bucket, mostly experimental. SoundHound and a couple of the newcomers are circling. I’m not aware of a single operator running this at production scale today, so I’m going to mostly skip it.

The buckets matter because the vendor pitches blur them on purpose. Every voice-AI deck I’ve seen in Q1 2025 has a slide that says “omnichannel” and shows drive-thru, phone, and kiosk icons next to each other. In practice, every vendor on this list is meaningfully stronger in exactly one bucket, and the others are roadmap. Make them tell you which bucket has paying production customers and which bucket is the deck.

The four numbers I care about per vendor, and the ones I’d put on every RFP:

• Accuracy — order intent accuracy on the live menu, not the lab menu. Anything north of 95% is competitive; anything reported without a denominator is marketing.
• Human-in-the-loop % — the share of orders that get escalated to a remote agent or the in-store team. Lower is better; “zero” is a lie. The honest vendors will tell you their HITL rate by daypart.
• Integrations — POS, loyalty, kitchen display, drive-thru hardware, and the SMS / phone stack. The right answer is “we have a production integration with the POS you’re already running”; the wrong answer is “we can build that.”
• Pricing transparency — per-store per-month, per-order, or revenue share? Is there a setup fee? What’s the integration cost? Operators who get burned in this category get burned on pricing structure, not headline rate.

The rest of this piece is per-vendor on those four numbers, plus a Vibe Check verdict.

SoundHound Polaris deep-dive

SoundHound Polaris / NextGen Platform Accuracy: 9/10 — best-in-class on the demos I’ve sat through, and the Feb 25 platform launch raised the floor again. HITL %: 7/10 — they don’t publish a number, but the chains running it report low single digits at peak. Integrations: 8/10 — Synq3 acquisition pulled in the chain integrations that mattered. Pricing transparency: 5/10 — chain-tier deals, not a rate card. Vibe Check verdict: The anchor brand. If you have more than 100 units, this is the call you have to take.

SoundHound unveiled its next-generation AI platform for restaurants on February 25, and the headline was Polaris — their in-house ASR model that they say is now live across the former Synq3 customer base. That last clause is the one operators should care about. Synq3 was the company SoundHound bought in early 2024 to instantly inherit a chain footprint. Twelve months later, the Polaris migration across that base is the proof point that the integration didn’t stall.

What you’re buying with Polaris, in plain English, is a voice model trained on restaurant audio specifically — drive-thru recordings, phone-order recordings, multi-speaker scenarios — rather than a general-purpose LLM that someone stapled a menu onto. That distinction is the whole game. The first generation of voice AI in this category (call it 2021-2023) leaned on Google or Amazon ASR plus rules engines, and the failure mode was always the same: accuracy collapses the moment you leave the demo menu. Polaris is a swing at the structural problem.

SoundHound has been careful in its disclosures about how many chains are running this in production. The Feb 25 release talks about the platform; the chain count is the metric they’ll publish when they’re ready to publish it — likely at the next investor checkpoint. For your purposes, the honest read is: more than a handful, fewer than the marketing would have you assume, and the operator references are real but gated. Ask for two, not one.

The pricing question is where the buy-side conversation gets thorny. SoundHound does chain-tier deals that look more like enterprise software than per-store SaaS. You will not find a rate card. What you’ll find is a per-order or per-store-per-month bundle with implementation services attached and a multi-year term. Negotiate the term down; negotiate the SLA up; insist on an accuracy floor with credits attached. Every chain that has signed in the last six months has, to my knowledge, gotten some version of those three concessions.

Where Polaris is weakest: phone ordering. The Synq3 base was drive-thru-heavy, and the Polaris demos I’ve sat through have been drive-thru-heavy. If your problem is the phone, Kea will out-execute SoundHound on a six-month horizon. If your problem is the drive-thru, this is the call.

ConverseNow: still in the conversation, barely

ConverseNow Accuracy: 7/10 — solid on phone, the drive-thru pitch is thinner. HITL %: 6/10 — opaque; ask hard questions. Integrations: 7/10 — broad POS coverage, fewer drive-thru hardware partners than SoundHound. Pricing transparency: 6/10 — per-store-per-month with a setup fee, generally negotiable. Vibe Check verdict: A defensible second choice for mid-market chains with phone-heavy off-premise.

ConverseNow was, eighteen months ago, the company I would have bet was about to break out. They have the phone-order pedigree (pizza chains, mostly), the POS integrations, and a reasonable drive-thru story. What’s happened since is that SoundHound got bigger and Kea got better at the phone game specifically, and ConverseNow has ended up squeezed between them.

I still recommend taking the meeting. Their phone product is genuinely good, the pricing is the most operator-friendly of the chain-targeted vendors, and the implementation team is the one I’d trust most for a 30-to-150-unit concept that wants voice on both phone and drive-thru without writing an enterprise check. If you’re a mid-market operator with significant phone-order volume — pizza, wings, anything where the phone is still 20%+ of the order mix — they should be on the shortlist.

What I’d push them on: the drive-thru references. The phone references are easy to get. The drive-thru references are scarcer, and that’s the half of the pitch where they’re trying to compete with SoundHound on accuracy. Make them prove it on your menu, not theirs.

Hi Auto: the focused player

Hi Auto Accuracy: 7/10 — narrowly-scoped and good at what they do. HITL %: 7/10 — they’ll tell you the number; that alone is worth points. Integrations: 6/10 — fewer than SoundHound, deeper where they exist. Pricing transparency: 7/10 — clear per-store pricing, no surprises. Vibe Check verdict: Underrated. The right answer for a single-concept chain that wants to move now.

Hi Auto is the most underrated name on this list. They started narrower than the competition — drive-thru only, fewer cuisine verticals, more conservative on the AI marketing — and the discipline shows up in the product. They run, in production, in several quick-service chains, and the operators I’ve spoken to have the calmest things to say about deployment and ongoing support of any vendor in the category.

The catch is scope. Hi Auto does drive-thru order-taking and does it well. They don’t have the phone product, don’t have the broader platform play, and don’t have the Synq3-style chain Rolodex. If you’re a single-concept multi-unit operator who wants to put voice on the speaker post and not buy a platform, Hi Auto is the most boring, most defensible call.

Push them on: the integration with your specific POS, and the support model for non-English speakers. Both are areas where they’ve improved measurably in the last twelve months, but neither is automatic. Get the references on both.

Presto: post-restructuring, pre-Wienerschnitzel

Presto Accuracy: 6/10 — they’ve improved, but the receipts are still uneven. HITL %: 5/10 — historically high; the new architecture is supposed to fix this. Integrations: 6/10 — narrower than competitors, deeper in concepts where they ship. Pricing transparency: 7/10 — public-company pressure has made them more legible. Vibe Check verdict: The restructuring story is real. Right answer if your concept needs margin discipline more than accuracy theater.

Presto is the company I would have written off two years ago and now have to explicitly recommend you not write off. They pivoted hard to voice, restructured the business, restructured the engineering team, and have spent the last several quarters trying to prove that the company that lost the early-stage drive-thru race deserves a second look.

The argument for taking the meeting goes like this: the rumored nationwide expansion at Wienerschnitzel is, if it lands the way the Galardi side has been previewing it, the biggest single deployment proof point Presto could ask for. Wienerschnitzel is the kind of concept — limited menu, drive-thru dominant, regional density — where a voice deployment can actually move the operations math. If that rollout ships and holds at scale, Presto goes from “restructuring story” to “category contender” in a quarter.

The argument against: the receipts are still uneven. Presto’s accuracy claims have historically come paired with HITL rates that, when operators get the real numbers, were higher than the marketing suggested. The new architecture is supposed to fix that. The Wienerschnitzel data will tell us. Until then, the smart move is to be on Presto’s call list, not their order book.

If you do sign with them in 2025, here’s what I’d insist on: monthly HITL reporting with daypart breakouts, accuracy guarantees with credits, and a six-month opt-out. Presto, to their credit, has been willing to negotiate on all three in recent conversations I’ve had operators forward me.

Kea: the indie unlock

Kea Accuracy: 8/10 — phone product is best-in-category. HITL %: 8/10 — published, with daypart detail. Integrations: 7/10 — POS-light, phone-stack-heavy; the trade-off works for the target operator. Pricing transparency: 9/10 — per-store-per-month, public-ish, no setup-fee shenanigans. Vibe Check verdict: The clearest “buy now” on the list for the right operator size.

Kea has done the thing in this category that I keep telling other vendors to do: they picked a lane, they shipped, and they didn’t try to be SoundHound. The lane is phone-order voice AI for independents and small regionals. The execution is, by a margin, the best on the list. The operators running Kea — and there are a lot of them — talk about it the way operators used to talk about Toast in 2018. Boring, in the best possible way.

The deployment claim — sub-one-hour to get a store on the platform — is the kind of marketing line I’d normally roll my eyes at. With Kea, I’ve watched it happen. Three of the operators I work with have gone from contract signature to live calls in less than a day. That’s a structural moat. It’s also why the anticipated next-tier launch matters: Kea has signaled a self-service tier for smaller operators, and if that ships the way it’s been previewed, the bottom of the market — the 1-to-10-unit independents — gets voice AI in a way that nothing else on this list can match.

Where Kea is weakest: the drive-thru. They don’t compete there and they don’t pretend to. If your problem is the speaker post, Kea is not the answer. If your problem is that your phone is going to voicemail at lunch and you’re losing $40,000 a year per store on missed calls, this is the easiest yes on the list.

What I’d ask in the demo: the call-deflection rate by daypart, the SMS upsell take rate (their phone product hands off to SMS in ways the competition mostly doesn’t), and the integration depth with your specific POS. Kea will give you straight answers to all three.

The newcomer watch-list

This is where I have to put on my Vibe Check hat and tell you the thing operators don’t always want to hear. There are five voice-AI vendors that have shown up in restaurant operator inboxes in the last twelve months — Incept, Foreva.AI, Palona, Maple, and Revmo.ai — that you should know exist, should put on a watch-list, and should not buy in Q1 2025.

The pitches are sharp. The demos are clean. The teams, mostly, are credible. What none of them have yet is the receipts: production deployments at scale, multi-unit operator references, integration depth with the POS you’re actually running, and the support infrastructure to handle a Saturday night outage at your busiest store.

Incept and Foreva.AI are the two I’d be most likely to bring back into the conversation by Q3. Incept has been quietly running pilots and their HITL transparency is already better than two of the names higher on this list. Foreva.AI has the strongest “voice agent for the back of house” pitch, which is a category I think actually grows in 2026, but it’s not what most operators are buying for in 2025. Palona, Maple, and Revmo.ai are earlier still — all post-2023 starts, all with single-digit production deployments. Watch them. Don’t sign them.

The Vibe Check rule on newcomers: the right number of voice-AI vendors to pilot at once is two. The right number of voice-AI vendors to sign at once is one. The right number of voice-AI vendors to discover via a cold LinkedIn message and pilot anyway is zero. If you’re getting two cold pitches a week — and several of my operator clients are — let them stack up. The category is not in a “first mover wins” phase anymore. The integrations and the operator playbooks are what win, and those compound on the vendors who’ve been shipping for three-plus years.

That’s not a put-down. It’s a buyer-discipline argument. The thing that kills you in this category is not picking the wrong vendor in a year; it’s picking two vendors in a year and ending up with four months of dual-tracked deployment work and no production wins.

For the broader context on how operators are thinking about consolidating tech stacks in 2025, a later operator case study on Chipotle’s tech stack gets at the same tension from the chain-scale side. And the same buyer-discipline frame is going to apply when we look at our later Vibe Check on Toast’s AI Suite — different category, same question: anchor brand or newcomer?

Vibe Check verdict

Three calls, in order of certainty.

SoundHound Polaris is the right answer for chain-scale operators with drive-thru as the priority. The Feb 25 platform launch consolidated the lead, the Synq3 migration is real, and the operator references — though gated — are credible. Push on pricing structure, push on the accuracy floor, and get the chain references with the actual HITL numbers attached. If you have more than 100 units and a drive-thru, this is the call.

Kea is the right answer for the bottom and middle of the market — independents and small regionals where the phone is the operations problem. The deployment speed is real, the pricing transparency is best-in-category, and the anticipated next-tier launch is going to make this even easier to underwrite for smaller operators. Easiest yes on the list for the right size.

Presto is the right answer for restructuring-mode concepts that need a voice partner with public-company margin discipline and are willing to accept slightly more variance on accuracy in exchange for it. The rumored nationwide expansion at Wienerschnitzel is the proof point I’m watching. If it lands, this story changes by Q2.

Everybody else is watch-list. ConverseNow is the credible second choice. Hi Auto is the underrated focused player. The five newcomers are LinkedIn-message status, not RFP-shortlist status, in March.

The operator in the Hyundai? He took the SoundHound meeting, told Presto he’d circle back after Wienerschnitzel landed, and signed up for the Kea early-access list for the four units in his portfolio that are phone-heavy. That’s the right answer, basically. Anchor brand for the volume, watch the restructuring story, take the easy yes on the side. The category will keep moving. The buyer discipline shouldn’t.

Vibe Check verdict: anchor brands win in Q1, the long tail watches GTC for the next inflection, and the operators who sign two vendors this quarter are the operators who will be writing me apology emails in October.

— Sofia runs Vibe Check. Tips: [email protected].