Voice AI's Pre-Money Math: What Operators Should Underwrite

It is the Friday afternoon that CES closes, and I am three hours into a comp model I do not entirely want to build. The tabs are familiar: PolyAI’s most recent raise, the Hume and SoundHound prints from last year, a half-finished revenue-per-location grid, and a CB Insights tab I keep returning to because the headline number is doing real work in my head. Voice AI equity funding grabbed $2.1B in 2024, and the same dataset previews companies raising nearly $500M in Q1’25 — a figure we will almost certainly clear before the end of this month given the rumored rounds I keep hearing about from late-January calendar invites. NRF opens Sunday. ICR runs into the middle of next week. There is an inauguration on the 20th. And somewhere in the back half of January, at least two voice-AI rounds I am tracking are expected to print at numbers that will make every restaurant-tech operator I know either celebrate or wince, depending on whether they are a buyer or a seller of the technology.

The reason I am building the comp model on a Friday instead of Monday is that the underwriting question for 2025 is not “is voice AI a category.” It is. The question is whether the per-location economics that operators face — the actual dollars-per-restaurant-per-month math — supports the vendor valuations that the late-January fundraise wave is about to set. My base case going in: they do not, not at the multiples being whispered, and the gap between vendor ARR pricing power and operator willingness-to-pay is the single most important thing to underwrite this year.

The Number That Anchors Everything

Let me start with the demand-side number, because it is the one every voice-AI deck I have read in the last six months opens with. PolyAI’s own customer research, which they cite in the announcement of their OpenTable partnership last September, shows that restaurants miss “between 30-60% of phone calls.” That is the wedge. Combine it with the National Restaurant Association’s most recent Operators Survey finding — that 62% of operators said their restaurants did not have enough employees to support existing demand — and the topline TAM story writes itself. Missed calls are missed covers. Missed covers are lost revenue. Voice AI promises to convert those missed calls into bookings, takeout orders, or at minimum a callback queue that does not depend on a host having a free hand.

That story is real. The 30-60% range is, in my experience modeling this stuff, directionally correct for full-service independents and mid-market groups, though the tail is uglier for concepts with bar-heavy phone traffic at peak. So the demand-side wedge is genuine, and that is exactly why the category attracted $2.1B last year and is on pace to materially exceed that in 2025.

What the decks tend to skip is the next slide — the one I am building right now — which is what a restaurant actually pays per location and what that implies for vendor unit economics at scale.

My Base Case on Per-Location Pricing

Here is where I land after talking to a half-dozen operators in the last month and triangulating against published partnership deals. Enterprise voice-AI for restaurants is currently being priced in a band of roughly $150 to $600 per location per month, depending on call volume, channel mix (reservations only versus reservations plus takeout plus FAQ), and whether the deal is bundled with an existing POS or reservations platform. For a 200-location enterprise group, that is somewhere between $360K and $1.4M in annual contract value. For a 50-unit emerging brand, it is $90K to $360K.

Now run the vendor side. If voice-AI vendors collectively raised $2.1B in 2024 and are pacing to clear another $500M in Q1 alone — and if the late-month fundraise wave I expect prints at the multiples I am hearing — the implied ARR these companies need to grow into is meaningful. PolyAI’s $86M Series D last year set one anchor. The anticipated rounds in the next three weeks will set another, and at least one of them is expected to clear nine-figure new capital at a valuation that, if the rumors are right, implies the vendor needs to be at substantial ARR within twelve to eighteen months to be in-the-money for the round.

My base case: at $300-500 ARPU per restaurant location, you need somewhere between 8,000 and 25,000 paying restaurant locations to support the ARR implied by the valuations being whispered for the late-January rounds — and that is before we discuss gross margin, churn, or the fact that the voice-AI category is going to be one of the most contested vendor markets in restaurant tech this year. There are roughly 750,000 commercial foodservice locations in the United States. So the addressable count exists. The question is whether any single vendor captures enough of it, fast enough, at pricing that holds, to justify the round.

Why The Vendor ARPU Story Compresses

This is the part I find under-discussed in the analyst write-ups I have seen on the category. Voice-AI pricing for restaurants is going to compress in 2025, and probably faster than the current models assume. Three reasons.

First, the underlying model costs are falling. The marginal cost of a voice-agent turn has dropped substantially in the last twelve months, and the late-January rounds I am tracking include at least one infrastructure-layer raise that, if it prints, will accelerate that trend. As model and inference costs fall, the value capture argument shifts from vendors to operators, and operators will renegotiate. They always do.

Second, the bundling dynamic is going to dominate. Reservations platforms, POS vendors, and host-stand software incumbents are all going to either build, buy, or partner their way into voice-AI features in 2025. The OpenTable-PolyAI partnership from September is a preview of that. When voice AI becomes a feature of a platform an operator already pays for, the standalone-vendor pricing premium evaporates. As our framework piece on the voice-agent maturity curve we later publish argues, the standalone phase of a category typically lasts twelve to twenty-four months before bundling pressure resets vendor unit economics.

Third — and this is the one I keep flagging in conversations with LPs — the deployment math at scale is harder than the demo suggests. A voice agent that handles reservations beautifully in a controlled pilot at three flagship locations is not the same product as a voice agent that handles 200 locations across six concepts with different menu schemas, different reservations rules, and different “do we accept walk-ins after 9pm” policies. The cost to deploy and maintain across an enterprise footprint is real, and it eats into the gross margin story that the pre-money math depends on.

What I Would Underwrite

If I am writing a check into the category at the multiples I expect to see in the next three weeks, here is the diligence I want to do, and the diligence I think operators evaluating vendors should do.

Net revenue retention by cohort, ideally by location count tier. Not aggregate NRR — the cohort math. A vendor whose 50-location enterprise cohort is at 120% NRR while their sub-10-location SMB cohort is at 75% is telling you something very specific about where the durable business is.

Gross margin by deal structure. Bundled deals through a reservations or POS partner will look very different from direct-sold enterprise deals, and the rev-share economics on the bundled deals are going to determine whether the vendor’s path to default-alive is real or aspirational.

Call-handling SLA performance, audited. Not the deck number. The actual operator-reported number. Voice AI lives or dies on whether the agent handles the awkward edge cases — the bar-on-hold-while-finding-the-host call, the customer who wants to modify a reservation made by someone else, the call where the menu has changed and the model has not been updated. If the SLA performance does not hold up across a representative deployment, the renewal math falls apart.

Operator-side switching cost. This is the underrated diligence question. If the vendor has built deep integration into reservations, POS, and host workflow, the switching cost is meaningful and pricing power is real. If the integration is shallow — a SIP trunk and a webhook — the operator can switch with a quarter’s notice, and pricing will compress in the next renewal cycle.

For operators evaluating vendors, the analogue is straightforward. Demand a cohort-level reference call list. Ask for the redacted SLA performance numbers from a peer group of similar size and concept type. Negotiate the renewal pricing schedule into the initial contract, because the vendor’s model assumes you renew at flat or higher, and the market is going to give you leverage to push that down. As in a piece we later publish on restaurant-tech valuations, the leverage in 2025 belongs to the operators who do the unit-economics homework before the vendor sets the anchor price.

What This Means for The Late-January Wave

So back to the comp model on my screen. The way I am framing the next three weeks for the LPs I talk to is this: the anticipated voice-AI rounds in late January will set the new ceiling for the category, and the CB Insights $500M Q1 print will be a leading indicator that the category is going to clear $3B or more in 2025. That is a bullish topline. It is also a setup for the kind of multiple compression you see in any infrastructure-and-application category where the underlying model layer is commoditizing faster than the application layer can build moats.

My base case is not “voice AI is overhyped.” My base case is “voice AI is a real category whose vendor valuations in early 2025 will price in a TAM capture rate that no single vendor is likely to achieve, and the operators paying $300-500 per location per month today are going to be paying meaningfully less in eighteen months as bundling pressure and model-cost compression flow through.” If you are an operator, the implication is to negotiate hard now and keep renewal terms short. If you are a vendor, the implication is to use the late-January round to build distribution moats — POS, reservations, and host-stand integrations — before the bundled competition reprices the category.

The number I keep coming back to is the 30-60% missed-call figure. That is the demand-side anchor, and it is what makes the category investable at any reasonable multiple. The underwriting question is not whether that demand is real. It is whether the value capture sits with the vendor or with the platform that the vendor sells through, and right now the pre-money math being whispered into the late-January wave assumes the answer is the former. I am not sure it is.

I will be watching the prints closely over the next three weeks. NRF and ICR will give us the first reads on operator sentiment. The rounds themselves will set the ceiling. And by the end of January, the spread between vendor valuation and per-location economics will either narrow, in which case the category clears, or widen, in which case the back half of 2025 gets very interesting for anyone holding the bag at peak multiple.

For now, the comp model goes back in the drawer and I go open a beer. The math will still be there Monday.

— Oliver writes The Bottom Line on M&A and valuations. Tips: tips@tabletransfers.com.