Against the AI Premium: What Operators Should Pay

It is a little past four on a Thursday afternoon, the kind of late January light that turns office windows into mirrors, and I am staring at a comp sheet that I have rebuilt twice in the last hour. The reason is the headline that came across the wire this morning: ElevenLabs closed a Series C, and the multiple in the deck I keep getting forwarded by founders and operators alike is the kind of number that ends arguments before they start. I have a coffee on my left, a printout of a build-vs-buy model from a thirty-location group on my right, and a feeling I have had two or three times in the last fifteen years — that the conversation operators are about to have with their boards in February is going to be very different from the one they had in November.

So I am going to write it down before the price gets baked in.

The thesis of this column is simple, and I want it up top so nobody has to mine for it. At current voice-AI vendor valuations, every dollar of operator margin you hand over in a multi-year contract is being capitalized into vendor equity, not into your P&L. The build-vs-buy framework most restaurant CFOs are using to make 2025 decisions was written for SaaS at five to eight times ARR. It is being applied to a category trading at four to seven times that. The denominator is wrong, and if you sign the wrong paper this quarter you will be the one explaining it to your board in Q3.

This is not an argument against voice AI. I think the technology is real, the use cases are right, and the operators who get it working will pull labor cost out of their business in ways they cannot pull it out any other way. It is an argument about what you should pay for it, who captures the surplus, and how to think about the next ninety days.

What 37x ARR actually means at the operator’s check book

Let me start with the number that prompted the rewrite. TechCrunch’s piece on the round today (https://techcrunch.com/2025/01/30/elevenlabs-raises-180-million-in-series-c-funding-at-3-3-billion-valuation) is worth reading in full, but the line that matters is the one I keep coming back to. The reporting puts ElevenLabs at “a $3.3 billion valuation, roughly 37 times its annual recurring revenue.” Thirty-seven times. I want to sit with that for a moment because I do not think it has registered yet with the operators who will see ElevenLabs pricing show up in a voice-agent stack inside the next two quarters.

A 37x ARR multiple is a number you pay for a category that the market believes will compound at least 60-80% annually for the next five years and exit somewhere in the public markets at a software-grade gross margin. To clear that hurdle, the vendor has to extract enough gross profit per logo to justify a discount rate that assumes near-monopoly outcomes. There is one place that gross profit can come from in a restaurant phone-answering use case, and that place is the labor line of the operator. It is not new revenue the AI is creating. It is existing payroll the AI is redirecting.

This is not a moral point. This is an accounting point. If a vendor’s investors are underwriting 37x of $90 million of ARR — call it $3.3 billion of enterprise value — those investors need that ARR to be sticky, expanding, and priced above cost-to-serve by a meaningful margin. They will price accordingly. They have to.

Bland’s Series B yesterday tells the same story from the other end of the maturity curve. The Emergence Capital post on the round (https://www.emcap.com/thoughts/ai-that-speaks-volumes-why-were-backing-bland) walks through the thesis, and the numbers in the public reporting put the Series B at $40 million on top of roughly $65 million of total funding. PolyAI, the most enterprise-leaning of the cohort, has raised on the order of $200 million across its rounds; the company’s own Series D announcement (https://poly.ai/blog/polyai-raises-86-million-series-d) is the cleanest primary source. Three different vintages, three different go-to-market models, one shared reality: the capital stack behind every voice-AI vendor an operator might sign with this quarter is priced for a future in which the vendor takes a large, durable share of what is currently restaurant payroll.

That is the number that should be in the denominator of your build-vs-buy model. Not the per-minute price. Not the per-seat price. The vendor’s ARR multiple.

Why every dollar of margin is now capitalized into vendor equity

Here is the move I have watched eight or nine times in my career, across point-of-sale, online ordering, delivery aggregation, and reservations. A category gets hot. Vendors raise at multiples that require pricing power. Operators sign three- and five-year contracts to lock in “today’s economics.” Two years later the operator has the technology, has improved the customer experience modestly, and has handed the vendor a recurring stream that the public market then capitalizes at 10-30x. The operator’s improved unit economics get reflected in the vendor’s equity, not the operator’s.

The voice-AI cycle is faster, and the multiple compression on the vendor side has not happened yet. So right now, every dollar of margin that an operator agrees to give up in a take-rate, per-call, or per-minute fee is being multiplied by whatever the vendor’s investors are willing to pay for ARR. If the multiple is 37x, the operator is — in present-value terms — capitalizing 37 dollars of vendor equity for every dollar of margin they sign away. That is not a fair trade unless the operator is getting 37 dollars of present-value benefit back, which I have not seen modeled credibly in any deck that has crossed my desk this month.

CB Insights’ January read on the category is worth quoting because it forces honesty on the maturity question. Their analysis of the voice-AI commercial maturity scale (https://www.cbinsights.com/research/voice-ai-market-opportunities) finds that “85% [of vendors are] in levels 1, 2, or 3 on CB Insights’ Commercial Maturity scale.” Eighty-five percent. The market is paying late-stage multiples for a category that is, by the most rigorous third-party scoring, mostly early-stage on commercial readiness. That gap is the operator’s risk, not the vendor’s.

My base case is that the multiple does compress, probably by the back half of 2025 as the second wave of voice-AI vendors raises into a more skeptical market. Operators who sign long-dated contracts at today’s pricing will be locked into a take-rate that reflects a moment, not a category. The build alternative — even an imperfect one — becomes meaningfully more attractive when you put the multiple back into the denominator.

What the build alternative looks like

I want to be careful here because I am not the engineer in the room and I am not going to pretend a build is easy. It is not. But the build-vs-buy framing has changed materially in the last twelve months, and the operators I respect most are the ones updating their priors on it weekly rather than quarterly.

The components an operator needs to assemble a credible voice agent for the most common restaurant use cases — reservations, simple takeout orders, hours-and-location questions, waitlist add — are now available as a set of API calls. The speech-to-text layer is commoditizing fast. The orchestration layer is the genuinely hard part, but it is hard in a way that a competent two-person engineering team inside a thirty-plus-location group can scope. The total cost-to-build for a voice agent that handles eighty percent of inbound call volume on a defined set of intents is, in my reading of the budgets operators have shared with me, in the low-to-mid six figures for the first year and meaningfully less in renewals.

Compare that to the all-in five-year cost of a vendor contract at today’s per-minute pricing, applied to the call volume of a hundred-location group, and the build case is closer than the procurement deck suggests. It is not a slam dunk. The vendor’s product is better today than what most operators could build today. But the vendor’s product is going to get cheaper to replicate every quarter, and the operator’s contract is going to get longer and stickier. The crossover is closer than people think. In a later piece we publish on restaurant-tech valuations (/blog/posts/dont-pay-the-ai-premium-a-buy-side-thesis-on-restaurant-m-a-in-2026), I will lay out the comparable spreads in detail; for the purposes of this column the directional point is what matters.

There is a middle path I think gets too little airtime, which is to buy thin and build deep. License the commodity components — the model layer, the speech layer — at usage-based pricing, and build the restaurant-specific orchestration in-house. The vendors selling end-to-end voice agents do not love this configuration, for obvious reasons. It is, in my view, the configuration that captures the most operator margin in 2025.

The CFO’s checklist for Q1

If you are an operator CFO reading this with a vendor contract in front of you, the checklist I would run before signing anything this quarter is short. I am going to be specific because vague advice does not help anyone.

First, model the vendor’s ARR multiple into your decision. Ask the vendor what their last round’s valuation was, divide by their stated ARR, and treat any contract that gives the vendor more than two percent of your annualized labor savings as a transfer of equity value out of your business. If the multiple is north of 20x, that transfer is being capitalized aggressively on the other side of the table.

Second, refuse any term longer than twelve months at the current pricing tier. I would put this in writing. The category is moving too fast and the multiple is going to compress. A three-year contract signed in February 2025 will look, in February 2026, like a five-year POS contract signed in 2014 looked in 2019: expensive and hard to exit.

Third, separate the commodity layers from the orchestration layer in your procurement process. If you let the vendor sell you the whole stack as a single SKU, you are paying their multiple on the commodity components. Make them quote you per-component. The ones who refuse are telling you something about where their margin actually comes from.

Fourth, build a real teardown of one competitor product before you sign. Not a demo. A teardown. Twenty hours of an engineer’s time is the cheapest insurance you can buy against signing the wrong paper in February. If your team cannot do the teardown, the procurement decision is premature.

Fifth, and most importantly, decide who in the org owns the renegotiation in twelve months. Vendor contracts in hot categories get repriced at renewal; the operators who win are the ones who treat the first contract as the opening bid and staff accordingly. As our framework on the voice-agent maturity curve we later publish argues (/blog/posts/the-voice-agent-maturity-curve), the technology will mature on a curve operators can predict; the pricing will not.

What I’d pay

I have been holding this section because I want to put a number on it, and I have rewritten it three times.

For a vended voice-AI solution that genuinely handles eighty percent of inbound call volume on the four use cases that matter — reservations, takeout, hours, waitlist — at a thirty-plus location group, my willingness-to-pay is in the range of fifteen to twenty-five cents of every dollar of net labor saved, with the contract structured month-to-month for the first year and a hard cap on take-rate at renewal. Below twenty-five cents I think the vendor is being squeezed and may not survive; above forty cents I think the operator is, in present value, transferring equity to the vendor’s cap table.

The pricing I am seeing in market today, when you back out the per-minute charges and the implementation fees and the platform fees into an effective take-rate on labor saved, is consistently north of fifty percent and in a few cases approaches seventy. That is too much. That is the operator subsidizing vendor multiple expansion, and it is the trade I do not want any restaurant CFO making this quarter without doing the math first.

The vendors will tell you this is fair because their product is hard to build and the alternative is worse. The first part is true. The second part is becoming less true every month. The question is whether the operator’s procurement cycle moves faster than the vendor’s pricing power does. Right now, the vendor is winning that race. My base case is that by the back half of 2025 the operator will be back in front, and the contracts written in January and February will look, in retrospect, like the ones that locked in the highest take-rate of the cycle.

So my advice, for whatever it is worth from someone who has watched this movie before: do not be the operator who funds the next ElevenLabs round. Be the operator who gets the technology cheap, on the right terms, and keeps the margin where it belongs — on your own P&L.

I will be wrong about some of this. I am usually wrong about timing. But I would rather be early on the multiple-compression call than late on a five-year contract.

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