The Four Margins of Hospitality AI: A September 2025 Scoreboard

It is the last Friday of September and my desk is the kind of mess that only accumulates when a news cycle refuses to slow down. Two empty coffee cups, a printout of the Darden Q1 FY26 release marked up in three different colors, a yellow legal pad covered in arrows, and a single line at the top of the pad that I have been circling all afternoon: four margins, not one.

I have been writing about hospitality AI since the cycle was still mostly slideware, and the question I get most often from operators — and, increasingly, from investors who have decided that the restaurant tape is more interesting than the SaaS tape — is some version of: which deployments will actually compound? The honest answer through most of 2024 was that we did not yet have enough data to mark anything to market. We had pilots, we had press releases, and we had a small number of operator anecdotes that lived somewhere between sales theater and field evidence. We did not have a scoreboard.

September gave us a scoreboard. Not a complete one — the next two quarters will fill in the gaps the Q1 prints could not — but enough of one that I am willing, for the first time, to commit a framework to the page rather than to a notebook. The frame I keep coming back to has four columns. Every defensible AI deployment I have seen in hospitality this year sits in exactly one of them. Most of the deployments that have failed to compound have either chased two columns at once or, more commonly, chased none of the four and tried to invent a fifth.

This essay is the first draft of that framework. I want to mark it as a first draft because I expect to revise it monthly through year-end and into 2026 — the May refinement of this framework will look different in important ways, and the voice-agent maturity model that I have been sketching in parallel will absorb some of what is here. The honest version of any framework essay is: here is the cleanest cut I can make today, and here is where I expect to be wrong. So let me give you both.

The contrarian thesis

The contrarian claim — and I want to be precise about why it is contrarian — is that there are exactly four margin pools where a hospitality AI deployment can defensibly live. Not three. Not five. Not “it depends.” Four.

The conventional taxonomy that you will hear at every conference this year sorts AI deployments by function: front-of-house versus back-of-house, customer-facing versus operator-facing, generative versus predictive, agentic versus assistive. Those taxonomies are not wrong, exactly, but they are taxonomies for vendors, not for operators. An operator does not care whether their drive-thru voice agent is “agentic”; an operator cares whether the agent moves a P&L line. A taxonomy that does not sort by P&L impact is a taxonomy that flatters builders and confuses buyers.

So here is the cut. The four margin pools are:

1. Labor — deployments that reduce hours per cover, per check, per room, or per shift. Wendy’s FreshAI is the canonical case; Kirk Tanner’s 80-basis-point claim on the February call is the canonical mark.
2. Traffic — deployments that move incremental covers, room-nights, or sessions to the operator’s owned channel rather than to a third party. Marriott’s HQ event this week is the canonical case; the direct-loyalty test is the canonical mark.
3. Attach — deployments that lift average check, average daily rate, or average basket through better recommendation, better merchandising, or better timing of the upsell. The Toast/Bon Appétit experiments are the canonical case; the attach-rate delta is the canonical mark.
4. Working capital — deployments that move cash conversion, inventory turns, or payment terms. The Sysco/PFG/USFD merger math, surfaced this month by the Sachem Head letter, is the canonical case; the working-capital arbitrage is the canonical mark.

Everything else — and I do mean everything — is either a feature inside one of these pools, a vendor pitch dressed up as a category, or a deployment that has not yet figured out which pool it lives in. The third case is by far the most common, and it is the single best predictor that a deployment will not compound.

Let me walk each pool in turn, mark it to market with September’s data, and then explain why I think Darden’s +10.4% revenue print — without a flagship AI deployment — is the negative-case test that makes the whole framework falsifiable.

Margin one: Labor

The labor pool is the one I am most confident we can mark, because it is the one where the operator’s incentive to disclose the number aligns with the analyst’s incentive to verify it. If you have moved 80 basis points of labor out of a four-wall P&L, you tell your shareholders. You tell them on the call, you tell them in the deck, and you tell them in the language of operating leverage because operating leverage is what restaurant equity analysts have been starved for since 2022.

Kirk Tanner’s February call is the cleanest mark we have. The FreshAI deployment — Wendy’s voice-AI drive-thru rollout, in partnership with Google Cloud — was cited as contributing roughly 80 basis points of restaurant-level margin expansion. I want to mark interpretation explicitly here: 80 bps is a contribution claim, not an attribution claim. The CFO did not say “FreshAI alone produced 80 bps”; he said FreshAI was a contributor to a bundle that, in aggregate, moved the line. The honest read is somewhere between 30 and 80 bps of pure-AI lift, depending on how generous you want to be about the counterfactual. That is still — and I want to be careful with this word — enormous by the standards of any operating initiative that QSR has tried in the last decade.

What makes the labor pool defensible is not the number itself but the structure of the number. Labor cost in QSR is roughly 28-32% of sales depending on geography and daypart. Moving 80 bps off that line is not moving 80 bps off labor — it is moving roughly 250 bps off the labor sub-account, because the denominator is smaller than total revenue. That is the kind of move that, sustained for two or three years, changes the unit economics of an entire chain. It is also the kind of move that, once proven at one chain, becomes a negative-option deployment at every competitor: if Wendy’s has it and you do not, your shareholders will start asking why.

The September data does not give us a new labor mark — we did not get a fresh print from a QSR with a comparable disclosure — but it gives us something arguably more interesting: a fundraise that prices the labor pool. Vox AI’s $8.7M seed, announced late August and discussed widely through September, is a bet that the autonomous-voice category has at least two or three more credible entrants before the market consolidates. The size of the round — modest by AI standards, generous by hospitality-tech seed standards — tells you what a sober investor thinks the option value of being a non-Google player in this category is worth. It is not zero, but it is also not a unicorn-track round, and that gap is the cleanest read we have on how locked-in the incumbents already feel.

The labor pool’s open question through year-end is whether the 80-bps result generalizes. Wendy’s is one chain, one menu structure, one drive-thru geometry, one labor market. The deployments that follow — and there will be many — will have to demonstrate that the result was not idiosyncratic. My working bet is that the lift compresses by roughly half as it generalizes, landing somewhere between 30 and 50 bps for a typical QSR rollout, and that the compression itself becomes the next interesting analyst story. That bet is one I expect to revisit in the October stack roundup once Q3 prints land.

Margin two: Traffic

The traffic pool is the one I have been most reluctant to mark, because the data has been the thinnest. Through most of 2025, the loudest claims in this pool — that AI-driven recommendation engines were redirecting reservations from OTAs back to operator-owned channels, that voice-search optimization was bending the demand curve, that personalized email cadences were lifting direct bookings — were all asserted without disclosure-grade numbers. The traffic pool was where every vendor wanted to live and where no operator wanted to share a P&L line.

Then Marriott hosted an HQ event this week and, in the careful way that Marriott does these things, told us roughly what the test looks like. Loyalty members — Bonvoy’s top tier and a curated set of mid-tier members — were invited to interact with a conversational AI booking surface designed to do the work that, in the legacy stack, was split between a search page, a property-detail page, and (for the highest-tier members) a human concierge.

The deployment is unmistakably traffic-pool. It is not pitched as a cost-out initiative — the human concierge channel is not being eliminated, and the loyalty event itself is, if anything, more expensive to run than a typical pilot. It is pitched as a direct-channel deepening initiative: an attempt to make the Marriott-owned booking surface enough better than the OTA-owned surface that the highest-value guests stop comparison-shopping. The success metric, if we ever see it disclosed, will not be a labor line. It will be a mix-shift line: direct bookings as a percentage of total bookings, with a secondary cut on top-tier loyalty share.

I want to be honest about why I think the Marriott deployment is the right canonical case rather than, say, a more loudly marketed OpenTable or Tock initiative on the restaurant side. The reason is that Marriott has the scale and the loyalty depth to actually move the OTA channel mix, and it has the disclosure habits to eventually tell us whether it worked. A restaurant-side traffic deployment that lifts direct reservations by 200 bps for a 50-unit group is interesting but unprovable; a hotel-side deployment that lifts direct bookings by even 50 bps for Marriott’s footprint is material, disclosable, and — critically — durable enough to become the case study that every other hotel group is measured against. I am sketching the longer Marriott read in a forthcoming May piece and will not litigate the full case here.

The mark to market on the traffic pool is: probable, but not yet proven. The pool exists, the canonical case is now visible, and the disclosure timeline is roughly two to four quarters out. The vendors who are competing for this pool — every voice-search optimization tool, every personalized-email engine, every conversational-booking surface — should be valued on the assumption that Marriott’s number, when it lands, will be the index against which they are measured. If Marriott’s number is good, the pool gets a re-rating. If it is mediocre, the pool gets quiet for a year.

Margin three: Attach

The attach pool is the one I am most worried about overstating, because attach is the pool with the easiest A/B test and the hardest sustained result. Every operator who has run a recommendation experiment has seen a lift in the first two weeks; almost every operator I have spoken to has seen that lift decay over the following two months. The honest version of attach-rate work is that it is real but mean-reverting, and the question is whether AI-driven systems can hold the lift longer than the rules-based systems they replaced.

The September data point I keep coming back to here is not a single deployment but a cluster of them — what I have been calling, in my own notes, the “Toast/Bon Appétit” cohort. These are operators who have been running structured recommendation experiments in partnership with content surfaces (Bon Appétit’s restaurant coverage being the most visible) and POS data (Toast being the most common substrate). The experiments are not new — versions of them have been running since 2023 — but the cohort got large enough in the last two quarters that we can start to read a signal.

The signal, as best I can read it, is that AI-driven attach beats rules-based attach by roughly 60 to 120 bps of average check, in the first quarter, and decays to roughly 30 to 50 bps of sustained lift by the third quarter. That is, again, a real number. A 40-bps sustained lift on average check, applied across a chain’s full top line, is the kind of result that earns a CFO call-out. But it is not the 100-plus-bps lift that the vendor pitch implies, and the gap between the pitch and the sustained number is — at the risk of repeating myself — the most consistent predictor of whether an attach deployment will survive its second annual budget review.

What makes the attach pool defensible, despite the decay, is that the direction of the decay is informative. The deployments that decay slowest are the ones that have been integrated into the operator’s content and merchandising calendar — they get refreshed with the menu, the season, the LTO cycle. The deployments that decay fastest are the ones that were sold as fire-and-forget. That distinction maps almost perfectly onto the operator/vendor relationship: the operators who treat attach AI as a tool that requires ongoing editorial work get sustained lift; the operators who treat it as a black box do not.

I am marking the attach pool as real but smaller than advertised. The sustained number is 30-50 bps, not 100 bps, and the moat is in the editorial integration, not the model. That has implications for how the category should be valued — it is closer to a content-tech multiple than an AI multiple — that I will return to when I get to Olo below.

Margin four: Working capital

The working-capital pool is the one I am most excited about and the one I am most likely to be wrong about, in roughly equal measure. Working capital is not where most AI conversations live — it does not photograph well, it does not demo at conferences, and it does not have a vendor cohort with a clean pitch. But it is, in my view, the pool with the most underpriced upside, and September gave us the cleanest data point we have had in months.

The data point is the Sachem Head letter urging a merger between Performance Food Group and US Foods, with Sysco as the implicit comparison set. The activist’s math is straightforward — combine the two, take out duplicative cost, and earn a multiple re-rating against Sysco — but the AI-adjacent version of the math is the one I want to sit with for a moment.

Foodservice distribution is, structurally, a working-capital business. The distributors carry inventory, finance receivables, and earn a spread on the cash-conversion cycle that is, in aggregate, a meaningful share of the segment’s profit pool. Anything that compresses that cycle — better demand forecasting, better inventory placement, faster receivables-to-cash, better terms negotiation with suppliers — drops disproportionately to the bottom line, because the denominator is small and the leverage is high. The largest distributors have been quietly investing in forecasting and routing AI for two or three years now, and the savings have been showing up in the gross-margin line without anyone being especially loud about it.

The Sachem Head letter is interesting because it forces the question into the open. If the activist’s math works at a combined PFG/USFD, it works in significant part because the combined entity can apply AI-driven forecasting and inventory optimization across a larger network than either could alone. The merger thesis is, in important respects, an AI-leverage thesis dressed up as a scale thesis. And the comparison to Sysco — the implicit benchmark — is a comparison about who has the larger working-capital surface to apply the same AI capability to.

The mark to market on the working-capital pool is: largest, quietest, and most underpriced. I do not have a clean public number to anchor it to the way I have FreshAI’s 80 bps for labor or Marriott’s loyalty test for traffic. But I have enough conversations with people inside the distribution stack — and enough of a read on how the comparable working-capital optimizations have priced in retail and industrial — to be confident that the pool is real and that it will, over the next two years, become the loudest pool in the framework rather than the quietest.

It is also the pool where the regulatory overhang matters most. The EU AI Act’s general-purpose obligations applicable since August 2 are most relevant to deployments that move money — forecasting models that drive procurement decisions, routing models that allocate inventory, payment-term models that price receivables. I will not litigate the full regulatory picture here — an upcoming regulatory piece will take that on properly — but operators who are thinking about working-capital AI without thinking about the GPAI obligations are going to discover, in roughly the order of magnitude of the next twelve months, that the compliance perimeter is larger than they assumed.

Olo as the multiple bellwether

I want to pause on Olo for a moment, because the Thoma Bravo close on September 12 is doing more work in my framework than its size alone would suggest.

Olo went private at roughly 3× revenue. That is, by software multiple standards, an unremarkable number. By restaurant-tech multiple standards, it is the most informative print we have had in a year. Olo is the cleanest pure-play restaurant-tech comp on the public tape — it sits adjacent to all four of my margin pools without owning any of them. It is in the labor pool by virtue of order-handling rails; in the traffic pool by virtue of direct-channel infrastructure; in the attach pool by virtue of upsell-surface placement; and in the working-capital pool by virtue of payment-flow proximity. Olo is the polyhedron that touches all four pools without specializing in any of them.

The 3× revenue mark, therefore, is the closest thing we have to a blended multiple for restaurant-tech that is broadly exposed to the four margin pools without being a pure play on any of them. That mark is a floor, not a ceiling. It tells us that a pure-play deployment that can defensibly claim one of the four pools — with a real number behind it, not a pitch deck — should price at a meaningful premium to Olo’s blended number. And it tells us that the pure plays that cannot defensibly claim one of the four pools should price at a discount, sometimes a steep one.

This is, I think, the single most useful framing for the next two quarters of restaurant-tech deal flow. The frame is not “is this AI?” — almost everything is AI now, in some form. The frame is: which of the four pools does this deployment defensibly own, and is the implied multiple above or below Olo’s 3× floor? A voice-AI deployment with a credible 50-bps labor claim should not trade at Olo’s multiple; it should trade at a meaningful premium. A reservation-tech deployment with a vague “AI-enhanced” pitch and no defensible pool claim should trade at a discount. The honest market will sort these over the next two to four quarters, and the sorting will look — I would bet — almost exactly like the four-pool frame predicts.

I want to mark one caveat: Olo at 3× is a take-private multiple, and take-privates are systematically lower than public-comp multiples because the seller is accepting a control premium discount in exchange for liquidity and an exit from quarterly disclosure. The fair-comp adjustment is probably 0.5 to 1.0 turn higher than the headline number. Even with that adjustment, Olo at roughly 3.5× to 4× revenue is the cleanest blended-multiple reference we have.

The Darden negative-case test

Now for the part of the essay that I have been arguing with myself about all week. If the four-margin frame is right, it has to be falsifiable. It has to predict not only which deployments will compound but also which non-deployments will fail. And the cleanest near-term test of that falsifiability is Darden.

Darden’s Q1 FY26 print showed +10.4% revenue growth without a flagship AI deployment. Darden does not have a voice-AI drive-thru — it is not in the daypart for that. It does not have a direct-loyalty-meets-conversational-booking surface like Marriott. It has not been the visible face of any of the attach-rate experiments in the Toast/Bon Appétit cohort. It is not in the foodservice-distribution stack where the working-capital pool lives. By the lights of my own framework, Darden should be underperforming — or at least, it should not be growing top-line at 10.4%.

So is the framework wrong?

I do not think it is, but I want to spell out the reasoning carefully, because the temptation in any framework essay is to wave away the negative case. Darden’s +10.4% is not a counterexample to the four-margin frame for three reasons.

First, the four margins are defensibility claims, not growth claims. The frame predicts which deployments will compound durably; it does not predict that operators without those deployments will shrink. Darden has been growing for reasons — menu execution, value perception, capital allocation, dividend discipline — that have nothing to do with AI and that are not going to be priced into a four-margin frame in any version. The frame would only be wrong if it predicted that non-AI operators could not grow, and it does not predict that.

Second, Darden is operating in a daypart and segment mix where the four-margin AI levers are weakest. Full-service casual dining is, structurally, a labor-pool deployment of last resort — the voice-AI substitutions that work in QSR drive-thru do not work in a seated full-service context, where labor is a service touchpoint, not a transaction friction. Traffic-pool AI works less well in a destination-occasion segment than in a routine-meal segment. Attach-pool AI works less well when the menu is already curated by a server. And working-capital AI is more relevant to Darden’s supplier ecosystem than to Darden itself. The frame predicts, correctly, that the four AI pools are less material to Darden’s P&L than to a comparable QSR or hotel operator. It does not predict that Darden cannot grow.

Third — and this is the part I am most interested in — Darden’s growth without a flagship AI deployment sets the opportunity-cost benchmark for the rest of the sector. Every operator considering an AI deployment now has to answer the question: would I be better off doing what Darden is doing — operational excellence, menu discipline, capital allocation — rather than spending the engineering and integration budget on an AI rollout that may or may not move my P&L? The honest answer for most operators is both, in some proportion. The honest answer for some operators is Darden’s playbook is better for me. And the four-margin frame helps surface that answer rather than burying it.

The negative-case test, in other words, is not whether Darden’s growth disproves the frame — it does not — but whether the frame is useful in deciding when not to deploy. I think it is. An operator whose unit economics do not have material exposure to any of the four pools should think very hard before committing to an AI roadmap that costs more than it returns. The frame is permissive, not prescriptive: it tells you which pools exist and how to value the deployments that target them. It does not tell you that you must fish in all four pools.

What I expect to revise

I want to close by being honest about where I expect this framework to need revision, because the worst version of a framework essay is the one that pretends to be finished.

The four pools, as I have drawn them, are probably too clean. In practice, the most interesting deployments touch two pools — Marriott’s loyalty-meets-conversational-booking surface is arguably both traffic and attach; FreshAI is arguably both labor and attach (the upsell suggestion is part of the voice agent’s job). I have written each pool as if deployments live cleanly in one, but the next two quarters will probably push me toward a frame where the pools are vectors with primary and secondary exposures, and the defensibility test is whether the primary exposure is large enough to justify the deployment on its own.

The labor pool’s 80-bps mark is probably too generous as a generalized number. I expect the next four or five QSR rollouts to land between 30 and 50 bps, and I expect the analyst conversation to migrate from “look at FreshAI’s lift” to “look at how the lift compressed.” That compression will look like bad news in the short term and like a healthy market in the medium term.

The traffic pool’s Marriott mark is unproven and may not land for two to four quarters. If it lands weak, the pool will get quiet, and I will have to rewrite this section. If it lands strong, the pool will absorb a substantial share of the next year’s hospitality-AI investment and the framework will need to give it more weight.

The attach pool’s 30-50 bps sustained number is the most likely to drift. I have anchored it to the Toast/Bon Appétit cohort because that is the cohort I have the most operator conversations from, but the cohort is small and the editorial-integration moat is hard to measure. A larger sample over the next two quarters could move the number either way by 20 bps without surprising me.

The working-capital pool is the one I am most uncertain about in absolute terms but most confident about in relative terms. I do not have a clean mark for it. I have a thesis — that it is the largest and quietest pool — and I have a structural argument for why that thesis should be right. But I do not have a number. The number will come, I think, from the PFG/USFD/Sysco arc rather than from a clean operator disclosure, and it will come in pieces rather than in one print.

And the regulatory overhang — the EU GPAI obligations, the patchwork of state-level deployments in the US, the disclosure conventions that have not yet stabilized — is going to move every pool’s defensibility, and not always in the directions the optimists predict. I am setting that aside for a regulatory piece that will run separately.

The bet

The bet I am making, by committing this framework to the page rather than to a notebook, is that the four pools are stable even if the marks inside them move. The shape of the frame — labor, traffic, attach, working capital — is, I believe, the right cut. The numbers inside the frame will iterate monthly. The shape, I think, will not.

I want to be wrong about as little of that as possible, which is why I am going to revisit it on a recurring cadence through the end of the year and into 2026. The October version will absorb whatever Q3 prints land between now and then. The November version will incorporate the second wave of voice-AI labor disclosures that I expect to come with the Q3 conference calls. The December version will close the year with a full mark on all four pools and — if the working-capital pool has produced any disclosure-grade numbers by then — the first serious revision to that pool’s mark.

The four-margin frame is, at the end of all this, a way of refusing to confuse motion for progress. Every quarter has more AI announcements than the last. Most of those announcements are not deployments, and most of the deployments are not in any of the four pools, and most of the deployments that are in the four pools have not yet earned their place there. The point of a scoreboard is not to celebrate the score; it is to make it harder to pretend the score is something other than what it is.

September gave us, for the first time this year, enough data to start keeping that scoreboard honestly. That is — to use the word I have been avoiding all essay — a useful thing.

I will see you in October.

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