The Q3 2025 Hospitality-AI Ledger: Reading the Four Margins, Six Months In

It is the Friday after Thanksgiving and I am at the same back-office desk I write all the Mise essays from, with the Q3 prints spread out in front of me and a yellow legal pad with four columns drawn down it. Toast on November 4. SoundHound on November 6. Sweetgreen on the Spyce sale on November 12. Olo’s close with Thoma Bravo on November 25. Denny’s $620M go-private deal on November 3. Five prints in three weeks. I have been rereading them against the Four Margins framework I have been carrying around since the summer — the one I want to formally put on the table for the December cycle — and the picture that emerges is starker than I expected. The framework holds. The thesis it implies is uncomfortable.

I want to be plain about what this essay is and what it is not. It is the first full statement of the Four Margins framework — the lens I have been using privately on operator pitches for two quarters and that I want the publication to start using publicly heading into the December planning cycle. It is not the final statement. The framework as I write it today has the right shape and the wrong vocabulary in places; I can already see two of the four labels needing work. I will be revisiting and refining the framework in coming months — the Jan 2026 iteration of this framework, the Feb 2026 refinement, and the forthcoming May refinement of this framework are the dates on the editorial calendar where this conversation will continue. Today is the day I put the first version on paper and read the Q3 tape against it.

The thesis

There are four margins that hospitality AI can move. I am calling them, today, labor, throughput, CAC, and shrink. Each has a different lever, a different evidence standard, and a different failure mode. The four margins are the operator’s lens. The argument I want to make in this essay is that when you read the Q3 2025 hospitality-AI tape against the four-margin frame, the margin gains accrue to the platform layer faster than they accrue to the operator layer. That is interpretation, not fact. But the prints are stacking in one direction. Toast is the loudest, SoundHound is the cleanest, DoorDash is the most strategic, and the operator-side prints — Sweetgreen divesting its kitchen robotics, Olo going private at a price that implies the public market lost faith in the operator-software thesis — are reading like the other side of the same trade.

I will work through each of the four margins in turn, anchored to a Q3 print, and I will mark interpretation where I am offering one. Then I will say what the platform-versus-operator pattern means for the December cycle and where I think the framework needs to evolve before the next earnings season.

Margin one: labor — the SoundHound print as the cleanest signal

Labor is the margin most operators talk about when they say AI. It is also, structurally, the margin most easily measured: hours per shift, hours per cover, hours per occasion. The Q3 print where the labor lever rang the bell most clearly was SoundHound’s.

SoundHound reported record Q3 revenue of $42 million, up 68% year-over-year, and raised full-year guidance. The release named more than 14,000 live restaurant locations on the voice-AI stack, against a year-ago base that the company has previously characterised as a fraction of that. The voice-agent category is the cleanest live labor-margin lever in hospitality AI today. A Stage 2 reservation-and-ordering voice agent is, structurally, a labor contract dressed as a technology contract. The vendor is telling you that the host stand can cover the same volume of calls with fewer human-hours, or alternatively that an existing host can be redeployed to floor service while the agent handles the phone. That is a labor contract. Read it as one.

What the SoundHound print told me about labor — and this is interpretation — is that the maturity curve has moved faster in 2025 than I expected. Fourteen thousand live locations is the kind of number that, six months ago, would have been a 2026 milestone. The category is shipping. I will come back to the voice-agent stage curve in a forthcoming May framework piece that walks through the five stages of voice-AI maturity and which operators should buy at which stage; for the four-margin frame today, the simple version is that voice-AI labor contracts are now real enough to be evaluated on the standard labor-margin terms.

What does a labor contract look like, properly written? Three components. First, a defined cost line on your P&L the vendor commits to reducing — typically front-of-house host hours, drive-thru order-taker hours, or in-store cashier coverage. Second, a measurable baseline — average hours per shift per location before deployment, measured for at least 30 days. Third, a measurable post-deployment number — same metric, same time window, six months in. If the vendor cannot produce all three, the contract is not a labor contract. It is a sentiment contract. Sentiment contracts are how every operator I know has wasted their first $200,000 of AI spend.

The SoundHound revenue print does not, on its own, tell me whether 14,000 operators are seeing the labor-margin compression they are paying for. That is the question the next two quarters of operator-side commentary will have to answer. But the platform print is unambiguous: someone is taking the margin on this category, and right now it is the platform.

Margin two: throughput — Sweetgreen’s Spyce sale as the negative case

Throughput is the most under-named margin in hospitality AI. It is the lever that increases output per unit of input — more covers per labor-hour, more transactions per square foot, more bowls per peak minute. It often produces a labor dividend, but it is structurally a productivity lever, not a labor lever. The cleanest live throughput case in hospitality AI through 2024 and 2025 has been, by general agreement, Sweetgreen’s Infinite Kitchen, the robotic bowl-assembly platform Sweetgreen acquired through its 2021 Spyce purchase.

The Q3 print on Sweetgreen this month was the one I read most carefully, and not for the headline number. Sweetgreen’s Q3 2025 10-Q disclosed the divestiture of Spyce Food Co. for $186.4 million, with the company retaining a license to the Infinite Kitchen technology. The structure of the transaction matters: Sweetgreen is not abandoning the kitchen-automation thesis; it is moving the capex off its own balance sheet and converting its operating exposure to a licensing relationship.

That is a meaningful read on the operator-side throughput thesis, and I want to be careful here because it is interpretation, not fact. The fact is that Sweetgreen sold the entity and licensed the technology. The interpretation is that for an operator running a public-market P&L, the capital structure of a throughput bet matters at least as much as the throughput math. The Infinite Kitchen throughput print has, by all available reporting, been good. The capital structure of building, maintaining, and refreshing the robotics fleet against the public-market expectations for a roughly billion-dollar revenue operator was, at the very least, not the cleanest match. Mark this as a read: the Sweetgreen divestiture tells me operators are starting to internalise that throughput AI is a capex lever first and a labor dividend lever second, and the public-market discipline on operator capex is now strict enough that even working deployments are being moved off the balance sheet.

The throughput contract, properly written, looks like this. Output per unit of input at peak — bowls per minute, tickets per hour, covers per labor-hour — measured for at least 30 days against the pre-deployment baseline, then again at 90 days. The trap on throughput is that the savings appear in peak-hour operations and disappear in off-peak. A KDS-AI tool that adds twenty seconds to a slow Tuesday lunch ticket is not net negative — slow Tuesdays are not where the contract lives. But the same tool that saves forty seconds at the Saturday-night push is net hugely positive. Measure throughput on the peak-hour ninety-fifth percentile, not the trailing-thirty average. Vendors who do not understand that distinction are not yet ready to sell you a throughput product.

The other throughput frontier is the KDS layer — kitchen display systems with intelligent prep-time prediction, dynamic ticket reordering, and cross-station coordination. Several Toast IQ surfaces fall in this category; the throughput contract on a KDS-AI deployment is straightforward: tickets per hour at peak, average ticket-completion time, mis-fire rate. If the vendor cannot produce those three numbers in a 90-day pilot, the deployment is not a throughput contract.

Margin three: CAC — DoorDash’s $1B ad business as the platform read

CAC — customer acquisition cost — is the third margin, and it is the one where the platform-versus-operator gap is widest right now. The vendor category is crowded. The operator confusion is high. And the platform layer is taking the lion’s share of the economic value.

The Q3 print that anchored this for me was DoorDash. DoorDash’s Q3 2025 earnings release was the standard top-line beat — strong order growth, strong gross order value — but the more important disclosure came on the call, where CEO Tony Xu characterised the company’s advertising business as running at approximately a $1 billion annualised revenue run-rate. That is, by any reasonable reading, the largest CAC-adjacent revenue line in hospitality-AI software right now, and it is sitting on a platform that intermediates the operator-guest relationship.

The CAC margin contract is, in principle, the simplest of the four to write: cost per booked guest before deployment, cost per booked guest after. Or, for a loyalty programme, cost per repeat visit before and after. Or, for a marketing-AI surface, cost per first-party email opt-in or cost per re-engagement. The numerator is dollars; the denominator is guests; the math is the math.

The reason CAC is the most-frequently-blurred margin is that AI-driven CAC tools often produce attribution improvements rather than acquisition improvements. The tool tells you, with higher confidence, which channel actually brought the guest in. That is genuinely useful — it should result in reallocated spend. But it is not the same as the tool lowering CAC. It is the tool measuring CAC more honestly. Operators who confuse the two will find their CAC number unchanged six months in and conclude the AI did not work, when in fact the AI did work — it taught them their real CAC was always higher than they thought.

What the DoorDash print told me about CAC — and this is the interpretation that I think is starkest — is that the platform layer has built a $1B annualised CAC-side revenue line by selling operators access to the operators’ own guests. The structural critique is not new; the magnitude in Q3 is. A billion-dollar ad business sitting on top of the third-party delivery platform layer is, in CAC terms, a billion-dollar tax on operator-side guest economics. Whether it is worth paying — whether the incremental order volume DoorDash drives at that ad price is net-accretive to the operator — is the question operators have to answer for themselves. But the platform print is unambiguous: this category is scaling fast and the economics flow to the platform first. I will return to the reservation-aggregator and delivery-platform CAC theatre in a forthcoming May piece on the DoorDash–SevenRooms transaction, which is the M&A expression of the same trade.

The CAC frontier on the operator-friendly side is the Toast loyalty-and-marketing surface, which I will get to in the platform-pattern section below. For now, the CAC trap is the standard one: a loyalty programme that lowers CAC because the guest already loves the concept is not really lowering CAC — it is expressing the brand margin in dollars. A loyalty programme that lowers CAC because the AI is sending the right offer to the right guest at the right time is doing actual work. Telling the two apart requires running the AI surface against a hold-out cohort. If a vendor refuses a hold-out cohort, the vendor is selling you sentiment again.

Margin four: shrink — the under-discussed margin in the Q3 tape

Shrink is the fourth margin and the one Q3 said the least about, which I think is itself the signal. Shrink is the margin of what you bought but did not sell — over-prep, waste, theft, mis-portion, over-pour, expired inventory, mis-counted receipts. In a full-service operator running at a 31% food-cost ratio, the shrink line inside the cost-of-goods number is typically two to four points of sales. That is meaningful money. Operators almost never see it broken out cleanly; the daily report rolls it into food cost. The chef carries it as variance against theoretical food cost. The owner carries it as a vague sense that the produce invoice is “high.”

Shrink is the margin where AI ought to be doing the most evidence-rich work — counting problems are the problems AI is best at — and where the Q3 print was, in my read, the most under-served. The inventory-AI vendor landscape is fragmented, the language is messy, and the proof-points are unevenly distributed across the public tape. There were no Q3 prints in the hospitality-AI category that landed on shrink as their headline number. That is not because the work is not happening; it is because the work, when it lands, lands inside food-cost variance and gets absorbed before it becomes a public number.

The shrink contract is the cleanest of the four to write. Before deployment: theoretical food cost minus actual food cost, expressed as percent of sales, averaged across a 90-day baseline. After deployment: same calculation, 90 days post-go-live, against the same SKU mix. The vendor who delivers shrink reduction shows you a variance line that compresses. The vendor who does not, does not. There is essentially no room for sentiment in the contract — the math is the math.

The trap on shrink is that the savings hide inside food-cost variance and look like seasonality. A produce price that drops three percent on a quarterly basis can mask a one-percent shrink reduction the AI actually delivered, or it can flatter a one-percent shrink increase the AI is hiding. Run the variance against the theoretical-cost line, not the actual-cost line. The theoretical line tells you whether the AI is doing work. The actual line tells you whether the supplier is doing work. Two different questions.

I am going to flag, looking ahead, that shrink is the label I am least confident in. By the time I write the next iteration of this framework I suspect this margin will get a different name — accuracy, or variance, or something else that captures the broader category of errors that compound silently into lost margin. The shape of the lever is right. The label has work to do.

The platform-versus-operator pattern: where the margin is actually accruing

Now to the pattern that emerged when I laid the prints next to each other and read them as a single picture.

The headline Q3 print on the platform side was Toast. Toast announced Q3 2025 results on November 4, and the disclosure that matters for the framework is the Toast IQ adoption number: the AI assistant is now in roughly 25,000 restaurants, with approximately 235,000 distinct uses of the assistant on the platform. That adoption curve — from a launch announcement on October 29 to a quarter-end disclosure four weeks later — is faster than any previous AI-feature launch on the Toast surface. I covered the Toast IQ launch in detail in a forthcoming May piece on what the Sous Chef pilot taught Toast; the short version is that the product is built around action rather than insight, and the early adoption print is consistent with that being the right design choice.

The Toast IQ adoption number sits inside the platform’s broader Q3 strength. Toast is now serving roughly 148,000 customer locations on the underlying platform; the AI surface is reaching ~17% of the installed base in the first month. That is platform-margin work. The platform is monetising the AI feature across an installed base it took a decade to build. The operator-side question — is the operator using Toast IQ seeing a measurable margin lift in their P&L? — is the question the next two quarters of operator commentary will have to answer.

Pair that with SoundHound’s 14,000 live locations and DoorDash’s $1B ad business and the platform side of the Q3 ledger reads like this: three large, well-capitalised platform companies, each scaling an AI-adjacent revenue line into the tens or hundreds of thousands of operator endpoints, each with the platform’s name on the contract.

Now read the operator side of the same Q3. Sweetgreen sold Spyce. The cleanest operator-side throughput deployment in the public tape moved off the operator’s balance sheet. Olo, the operator-software platform that intermediates digital ordering for tens of thousands of restaurant locations, completed its take-private with Thoma Bravo on November 25 at the previously disclosed roughly $2 billion enterprise value. Denny’s announced its $620M go-private earlier in the month. The PitchBook tape on H1 2025 tech M&A had already shown deal volume up roughly 45% year-over-year; Q3 and the November tape extend the pattern.

What I read in those operator-side transactions — and again, I am marking this as interpretation — is that the public market has, at least for a moment, lost patience with the operator-software thesis at the prices that thesis was originally underwritten. Olo at $2B private is a smaller number than the Olo bulls were modelling in 2021. Sweetgreen moving Spyce off the balance sheet at $186.4M is a smaller number than the original Spyce acquisition price implied. Denny’s at $620M is, similarly, a public-market verdict on the operator-side return profile. The platforms are scaling AI revenue lines into eleven and twelve-figure quarterly run-rates. The operators are being repriced.

The interpretation that the framework forces me to write is the one I have been resisting since the summer, and the Q3 tape made it harder to keep resisting: AI margin gains in hospitality are accruing to the platform layer faster than they are accruing to the operator layer. That is the four-margin frame applied to the Q3 ledger. The platforms are taking the labor, throughput, and CAC margins through their software products and selling them back to operators on platform terms. The operators are paying for the surface. The shrink margin — the one most under-discussed in the Q3 tape — is the one operators have the best shot at capturing directly, because counting problems do not yield a platform tax in the same way that loyalty, marketing, and voice-AI do.

I want to be careful about the strength of that claim. The pattern is starker than I expected. It is not airtight. A counter-argument worth taking seriously is that the platform-side adoption is the leading indicator and the operator-side margin compression is the lagging one — that we will see operator P&Ls compress favourably across 2026 as the AI surfaces mature, and the public-market repricing of Olo and Denny’s is unrelated. I take that counter seriously. But the framework’s job is to force a hypothesis that the next two quarters of evidence can falsify. The hypothesis I am putting on the table for the December cycle is the platform-takes-the-margin hypothesis. We will measure it.

What the framework still needs

The framework as I have written it today has the right shape and the wrong vocabulary in places. I want to be explicit about where I think it will need work, because I am committing on the record that I will revisit it.

First, the shrink label is doing work it shouldn’t have to do. The lever is right — reducing the errors and losses that compound silently into lost margin — but the word shrink trains operator attention onto inventory and theft, when the broader category includes order accuracy, prep accuracy, and the entire surface of mis-fired tickets and mis-counted receipts. I suspect by the time I write the next iteration of this framework I will have moved this margin to a different name. Accuracy is the candidate I am holding most loosely. I will see how the next two quarters of evidence read.

Second, CAC is a label that lumps two genuinely different theses together — new-guest acquisition and repeat-guest retention. They are structurally different problems with different evidence standards. The DoorDash ad business is largely an acquisition lever. A Toast loyalty surface is largely a retention lever. The frame would do more work if it broke them apart. I am not breaking them apart today because the four-margin shape is more useful than the five-margin shape for an operator’s first read; but I expect to soften that simplification in the next iteration.

Third, the framework does not yet have a clean way to express the platform-versus-operator split inside each margin. The labor lever exists for Toast and for the Toast customer; they are different contracts, and the platform-side contract is much further along the maturity curve than the operator-side contract. I want the next iteration of this framework to make that explicit — perhaps by adding a column for who captures the margin alongside the which margin column. The Q3 ledger essentially demands it.

Fourth, I have been silent in this essay about the brand and enterprise margins, which is a deliberate omission. Those margins exist; they sit above the operating-margin column the four AI-spend margins live inside. The broader four-margins-of-a-restaurant frame — gross, operating, brand, enterprise — is the lens this AI-spend frame sits inside, and I will write it in the spring.

The December cycle

The four-margin frame is, at minimum, a working vocabulary. Operators in the December planning cycle who use it to read pitch decks will reject perhaps half the meetings on their calendar by the end of the second slide, because half the AI vendors selling into hospitality cannot tell you in a single sentence which of the four margins their product moves. That is the most immediate use of the framework: as a filter.

The second use is harder. It is the use I want operators to take into the December budget conversation. If the Q3 read is right — if AI margin gains are accruing to the platform layer faster than they are accruing to the operator layer — then the operator-side defensive move is not to buy more platform-side AI. It is to invest in the AI surfaces that are not monetisable by the platform: the shrink line, the back-of-house variance, the inventory and accuracy work the platforms are least interested in monetising directly. That is where the operator-captured AI margin is most likely to land in 2026. The platform-captured margins — labor (voice), CAC (loyalty/marketing), throughput (KDS/kitchen automation) — will continue to scale, but the price of the surface will continue to compress operator-side gains.

I will revisit and refine this framework in coming months. The Q1 2026 print cycle, when it lands in early February, will be the next forcing function. I expect the labels to evolve, the platform-versus-operator split to harden into a column, and the shrink-margin definition to broaden. Mise is supposed to be a working frame, not a museum piece. This is the first version.

— Eitan is editor-in-chief of TableTransfers. Tips: eitan@tabletransfers.com.