The Four Margins of the AI-Era Restaurant

It is the last Friday of August. The newsroom is quiet in the way August newsrooms are quiet: half the staff is at the beach, the other half is filing the back-half of the back-to-school previews, and I have an empty notepad in front of me and the residue of about forty earnings calls in my head. I keep doing the same thing every twenty minutes — refilling a coffee, walking to the window, sitting down again, writing one sentence, scratching it out. The sentence I keep trying to write is some variation of the AI story in restaurants in August was about… and every time I finish that sentence with a single noun, the rest of the month rises up and laughs at me.

It was not about one thing. That is the problem and that is the column.

So I am going to do something I have not done before in these Mise essays, which is to abandon the pretense of a single thread and write the framework I have been talking through with operators all month. I want it on the page where I can argue with it later. I want to be able to come back next month and the month after and say here is where the first sketch was wrong, here is where it survived contact. Treat this as a first draft. I will iterate this thing in public — there is a refinement of this framework due in May where I expect to revise at least half of what is below — and I would rather have something to argue with than nothing.

Here is the contrarian read I keep coming back to, and the thesis of this column: August 2025 was the month every margin in the restaurant business got a different AI thesis attached to it. Labor, food, occupancy, capital. Four margins, four theses, four different sets of vendors selling four different stories. The trade press wrote about each one in isolation. The operator I want to write for has to make decisions about all four at once, with one balance sheet, and that is the framing problem. Most of what passes for restaurant-AI commentary right now is a category error: it pretends there is one AI thesis when there are at least four, and they trade off against each other.

The rest of this essay walks the four margins, names the August evidence behind each, and ends with the question I think every multi-unit operator I know is sitting with right now: which margin do you defend, and which do you attack.

Why “margins” and not “use cases”

A quick aside about why I am insisting on margin language rather than the more standard “use case” framing the consultancies prefer.

Use-case taxonomies — chatbot, drive-thru voice, demand forecasting, dynamic pricing, vision-based portion control — collapse the moment you put them next to a P&L. A drive-thru voice agent is a labor story for one operator, an occupancy story for another (if it speeds throughput at the lunch peak), and a food-cost story for a third (if upsell recommendations shift mix toward higher-margin SKUs). Tagging it as one “use case” hides which dollar line on the income statement is actually moving.

Margins, by contrast, are the lines the operator stares at. Prime cost — labor plus food — is the obvious one. Occupancy is the line nobody calls a margin because it lives upstream in revenue, but every full-service operator on the planet is running an implicit yield on covers per hour and dollars per cover, and the AI thesis on the booking layer is functionally a margin thesis. Capital is the one I keep arguing belongs in the framework because the cost of money — how it gets raised, how it gets paid back, whether you are public or private — determines which of the other three theses you can even afford to chase.

So: labor, food, occupancy, capital. Four margins. Let me take them in order.

Margin 1: Labor — the kitchen is the test bed

The August story I cannot stop thinking about is Sweetgreen’s Q2 print. They missed the top line, the stock got hammered after-hours, and the call itself — which you can read in the Investing.com transcript — was the most explicit articulation I have seen all year of the labor-margin AI thesis from a public restaurant company. They named a roughly 700-basis-point restaurant-level margin uplift at Infinite Kitchen units versus the rest of the fleet. Seven hundred basis points. The traditional QSR labor line is a 25-to-30-point line; moving it by seven points is not optimization, it is a different business model.

Now, before everybody pencils 700 bps into their own model — and I have spoken to two private-equity-backed chains this week who already have — read the transcript carefully. The uplift is reported against the fleet, not against a matched comparable. The IK stores are newer, generally in stronger trade areas, and were built with the layout designed around the robot rather than retrofitted. Some unknown but non-trivial slice of those 700 bps is unit-economics-of-new-Sweetgreens, not unit-economics-of-IK. I would call the IK-specific contribution somewhere between 200 and 400 bps until I see a same-store cohort, and the rest is build-quality and trade-area selection.

Even at the conservative end, that is a real number. And the strategic claim — that the back-of-house is the labor test bed for the AI thesis, not the front-of-house — is the right one. The forthcoming May case study on Sweetgreen’s Infinite Kitchen rollout will go deep on the unit-economics question. For this essay, the relevant point is that Sweetgreen has now publicly committed to making this the centerpiece of the new-build pipeline, which means the experimental period is over and the operating period has started.

The Chipotle data point lands in the same week and tells you that the assembly-line operators have all learned the same lesson. The Restaurant Technology News write-up of the plancha rollout is worth reading in full, but the headline number is that the new chicken plancha takes the cook time from twelve minutes to about four. That is a three-times throughput improvement on a single station, and it sits inside a stack that also includes the produce slicer and — coming through the pipeline — the digital-make-line vision system that I covered in an upcoming May piece on the Chipotle AI stack.

The thing to notice about the plancha is what it is not. It is not a robot. It is not a generative model. It is dumb stainless steel with better thermodynamics. But it is part of the same labor-margin thesis as Sweetgreen’s IK: the back-of-house is where the labor margin gets recovered, the front-of-house is where the brand gets defended, and the operator who confuses the two is going to spend a year and seven figures putting a chatbot on a kiosk and wonder why none of it shows up in the P&L.

The labor-margin thesis, in one sentence: the back-of-house is the test bed; the front-of-house is the trap. The capex shows up first, the labor savings show up over twelve to twenty-four months, and the payback window — if you believe Sweetgreen’s framing and discount it by half — is inside the 24-month operator horizon I care about for this framework.

I will revisit this in a month. My current bet is that the labor-margin thesis is the strongest of the four, in the narrow sense that it has the clearest causal mechanism, the most public proof points, and the shortest payback. It is also the most capital-intensive, which is why the capital-margin thesis at the end of this essay matters as much as it does.

Margin 1, continued: where the labor thesis breaks

I want to spend one more section on labor before moving on, because the failure modes matter for the framework.

The most common failure mode I am hearing on operator calls is the substitution fallacy: the assumption that a robot or a piece of automated equipment replaces a worker one-for-one and the savings show up as that worker’s fully-loaded cost. They do not. They show up as a smaller crew with a different skill mix, and the new skill mix — the person who babysits the IK, the technician who fixes the plancha, the regional ops manager who now owns five units instead of three — is paid more per hour than the people who were replaced. The net is positive, but the headline-to-net ratio is something like 60 percent. Build that into the model.

The second failure mode is the throughput trap. A three-times-faster plancha is only a margin event if you can sell three times the chicken through it. In a peak-constrained store — most Chipotles, most Sweetgreens at lunch — you can. In a demand-constrained store, the plancha is a labor saver but not a revenue lever, and the math changes. This is where the labor-margin thesis bleeds into the occupancy-margin thesis below: throughput improvements are only as valuable as the demand you have to put through them.

The third failure mode is the maintenance line. Every operator I have talked to who has run an IK or a comparable system for more than six months has the same story, which is that the uptime curve in months one through six looks great and the uptime curve in months six through twelve looks like a heart-attack EKG. The vendors are getting better at this — Sweetgreen’s transcript suggests their second-generation IK is meaningfully more reliable than the first — but the labor-margin model needs a maintenance reserve that nobody is putting in the deck.

Margin 2: Food — the platform is taking the cost

The food-margin thesis is more interesting than the labor one because the August evidence points in a direction most operators are not yet pricing in. The food margin in restaurants — the cost of goods, the spread between menu price and food cost — has been under pressure for three years from input inflation. The AI thesis attached to it in August is not about predicting demand or reducing waste, though both are real. It is about who pays for customer acquisition.

DoorDash printed Q2 results on August 6. The release is worth reading not for the top-line GOV number, which is in the press, but for the disclosures around the advertising business. The trajectory there — they are now describing a roughly $1 billion ad revenue run-rate, growing fast, with most of it sold to restaurants and CPG brands — is the food-margin story.

The way that thesis cashes out is this. For a decade, third-party delivery has been a margin compressor for restaurants: the platform takes 25 to 30 percent of the ticket on a commission basis, and the restaurant absorbs it. The advertising layer reframes that compression as something else. Instead of paying commission as a cost of distribution, restaurants are now paying ad fees as a cost of acquisition — and the platform is increasingly indifferent between the two, because the unit economics for them are similar.

For the restaurant, the difference is large. Distribution cost gets buried in COGS. Acquisition cost is supposed to live above the line and be matched against LTV. If you can convert a commission line into an ad line and the conversion math works — meaning the incremental order from the ad would not have happened without it — you have not lost money, you have moved the marketing budget. If it does not work, you have stacked two costs on top of each other and your food margin gets worse.

The mid-market chains I respect are running this analysis right now and most of them are coming out skeptical. The ad-funded acquisition thesis works for large operators with sophisticated attribution and pricing power, and it does not work for independents who do not have the attribution infrastructure to tell whether the incremental dollar showed up because of the ad or because Tuesday is always a good Tuesday. The platform knows this asymmetry better than the restaurant does. The platform also has the data.

The same dynamic shows up in a smaller form in the Yelp Host product, which crossed my desk in August and which I am still chewing on. Yelp is — and has always been — primarily a search-and-discovery surface. Host is their attempt to bundle a reservations layer on top with AI-assisted guest management, and the pricing structure is an upsell on top of the existing advertising relationship. The food-margin angle is the same as the DoorDash one: the platform owns the demand, the restaurant pays for access to its own customers, and the AI layer is the wrapper that justifies the new price point.

This is the thesis that I expect to age the worst of the four. Operators with leverage will push back on it within twelve months. The platforms will adjust pricing, the take-rate-versus-ad-rate calculus will normalize, and the food-margin pressure from delivery will revert to something closer to where it was before. The AI thesis here is not actually about AI; it is about who controls the demand graph, and the AI features are window dressing on a distribution fight that has been going on since 2017.

If I had to name the food-margin thesis in one sentence: the platforms are turning their take-rate into an ad-rate, and the AI features are the justification, not the cause.

Margin 2, the small-operator footnote

One thing the food-margin section above understates: there is a real AI thesis on food costs at the operator level, separate from the platform fight. Inventory forecasting, dynamic recipe costing, spoilage prediction, supplier negotiation. The tools have gotten meaningfully better in 2024 and 2025, and the payback math on a credible inventory-management system is in the 9-to-15 month range for a mid-size chain.

I am not putting those tools in the food-margin section of the framework because they are not the August story, and this is a framework essay about August 2025, not about everything that is happening in restaurant tech. The platform-versus-operator distribution fight is what the month was about. The inventory tools are real and matter; I will come back to them in a future essay. Treat this as a known gap in the first draft.

Margin 3: Occupancy — the demand layer is consolidating

The occupancy-margin thesis is the one most people miss because occupancy is not a line item on a restaurant P&L the way labor and food are. But every full-service operator I work with thinks about occupancy constantly. Covers per hour. Dollars per cover. Two-top versus four-top utilization. No-show rate. Walk-in conversion. These are the levers, and the AI thesis attached to them in August 2025 is a consolidation story, not an optimization story.

The August catalyst is the Toast-Amex partnership, which has not been written up well enough yet for me to link to a definitive source, but which is being talked about across the industry. The shape of it: Toast as the dominant POS in independents and mid-market chains, Amex as the bank with the highest-LTV diners and a long-running reservations product through Resy, Tock now sitting inside the Amex perimeter, and a new layer that routes demand from cardholders to restaurants based on availability, ticket size, and — this is the AI bit — predicted yield.

The Toast side of this is on the table because of the Q2 print, which is a strong quarter and gives them the latitude to do something ambitious on the demand side. The Amex side has been telegraphed for two years through the steady accumulation of dining-related assets (Resy, Tock, a series of partnerships). What is new in August is that the two sides have started talking to each other in ways that produce a routable signal: a cardholder opens the Amex app, the Amex app knows which restaurants have inventory at 7:30pm on a Thursday because Toast told it, and the inventory routing has a yield component on top — meaning the cardholder gets directed not to the closest restaurant with availability but to the one where their predicted spend matches the restaurant’s predicted yield.

For the restaurant, this is the most consequential AI thesis of the four, and the one I have the most trouble pricing.

On the positive side: a yield-routed cardholder is, in expectation, a higher-LTV cover than a walk-in. If the routing is well calibrated, the restaurant gets a Thursday-night four-top that spends $400 instead of $220. The annualized impact across a fleet is enormous; for a 50-unit full-service chain, even a 5 percent lift on the routed share — call it 20 percent of covers — is high-seven-figures of incremental revenue at full-restaurant margin.

On the negative side: the demand graph belongs to the routing layer, not to the restaurant. This is the same problem as the DoorDash ad thesis, except worse, because the dining occasions Amex routes are the high-LTV occasions that the restaurant most needs to own directly. If the routing layer is the way a 30-something Manhattanite finds their Thursday dinner, the restaurant has lost the marketing relationship with that diner and is paying — directly through Amex partnership fees, indirectly through Toast — for the access.

The occupancy-margin thesis in one sentence: the demand graph is consolidating into a routed yield layer, and the operator’s choice is whether to ride it or fight it. I do not yet know which side of that choice is correct. I am going to spend the next two months trying to figure it out. There is an upcoming May piece on Marriott’s AI deployment that I think will help, because the hotel industry went through this exact transition with the OTAs ten years ago and the lessons should carry. There is also a forthcoming framework piece on voice agents that bears on the front-of-house side of the occupancy story, because the voice layer is increasingly where the demand-routing decision gets surfaced to the diner.

For the September scoreboard I expect to write at the end of next month, I want to track three things on this margin: what percentage of covers at a sample of partner restaurants are being routed versus walked-in versus directly booked; what the spread is on ticket size between the three; and what the partnership economics look like once the platforms start to price aggressively.

Margin 3, the QSR variant

The QSR variant of the occupancy-margin thesis is the drive-thru voice agent, and I want to flag it briefly even though it is not the August story.

The voice agents are getting better — there is a piece in the May queue on McDonald’s drive-thru AI that walks the deployment in detail — and the through-line on the framework is similar. The voice agent is functionally a yield router: it gets the order, it knows the queue depth, it can upsell or hold based on capacity, and it routes the customer to a finished window with predicted accuracy. The labor story attached to it is real but smaller than the occupancy story, because the marginal labor saved by a voice agent at the speaker is a fractional FTE per store, while the marginal throughput unlocked at the peak hour is meaningful.

This is the one place where the labor and occupancy theses are clearly fighting each other in the vendor pitch decks, and the operator I care about is the one who can tell the two apart. A voice-agent vendor selling on labor savings is, in most cases, leaving the bigger number on the table. A voice-agent vendor selling on throughput is selling the right product against the right margin.

Margin 4: Capital — the public-market exit is closing

The fourth margin is the one nobody calls a margin, and it is the one that determines whether you can chase any of the first three.

The August evidence is the Olo deal. The Q2 release is the last public quarterly the company will file, because they are going private at $10.25 per share — roughly a $2 billion enterprise value — to Thoma Bravo. The deal closes pending shareholder vote and standard conditions. Olo was one of the small handful of restaurant-tech public companies, and the take-private moves the second-largest pure-play (after Toast) off the public market.

The capital-margin thesis attached to this is what I have been thinking through all week.

The shallow read is that public markets are not currently willing to fund restaurant-tech at attractive multiples, and the private market is, so the rational play is to take the discount-to-fair-value cash today and let private capital reset the cost base, the product roadmap, and the AI investments without quarterly scrutiny. There is something to that, but it understates the strategic content of the deal.

The deeper read is that the AI capex cycle in restaurants — the next two to three years of spend on the labor, food, and occupancy theses above — does not fit cleanly in the public-market quarterly rhythm. The investments are lumpy, the payback windows are 12 to 24 months, and the headline numbers in the interim look worse than the prior steady state. A public company that tries to lean into the AI thesis on labor automation has to absorb the capex up front, defend a margin compression in the year of installation, and trust the market to wait for the run-rate. The market has not waited well in 2024 or 2025.

Going private buys time, and time is the scarce input. The Thoma Bravo thesis on Olo is not, I suspect, a clever financial-engineering thesis. It is a “we will fund three years of AI investment outside the quarterly window and bring it back to the market when the run-rate is clean” thesis. If that bet works, it sets a template, and we should expect to see at least one more public restaurant-tech name follow them off the public market inside the next twelve months.

For the operator, the capital-margin thesis is the most important of the four and the least directly actionable. You can choose how to invest in labor automation, you can choose your platform strategy, you can choose whether to ride the routed demand layer. You cannot choose, by yourself, whether the public market is open for restaurant-tech. But the consequences of that question determine which of your vendors will be around in three years, what their pricing will look like, and whether the partners you are betting on are forced sellers or patient builders.

The capital-margin thesis in one sentence: the public window is closing for restaurant-tech AI capex, the private window is opening, and the vendor consolidation that follows is the dominant structural risk on the operator’s three-year horizon.

The regulatory tax — and why it sits across all four

I cannot finish the framework without mentioning the EU AI Act, because the relevant provisions started to bite on August 2 and the operator I am writing for is going to have to think about them whether they have European exposure or not.

The Commission’s own announcement sets the date. The provisions that landed on August 2 are the general-purpose-AI-model transparency and documentation requirements; the full set of obligations phases in over the following two years. For a US-only operator, the direct legal exposure is minimal. The indirect exposure — through vendors who serve both markets and who will harmonize their compliance posture rather than fragment it — is meaningful. Every vendor in the stack will spend the next twelve months adjusting their documentation, their model cards, their data-sourcing disclosures. Some of that cost will get passed through. Some of it will slow product velocity. None of it will obviously help any of the four margins above.

I am not going to spend much more on this here because the practical implications are still developing. The framework point is that the regulatory tax is a cross-margin drag — it shows up in vendor pricing for the labor thesis, in platform fees for the food thesis, in compliance overhead for the occupancy thesis, and in due-diligence cost for the capital thesis. It is not a margin of its own. It is a weight on the whole framework.

The bet: which margin to defend, which to attack

Here is the operator decision I have been working toward all essay.

You have one balance sheet. The four AI theses above all want some of it: capex for the labor automation, ad budget for the food fight, partnership fees and analytics spend for the occupancy routing, and reserves to survive the vendor consolidation on the capital side. You cannot fund all four equally. You have to pick which margin to defend — meaning, where to protect the existing economics against erosion — and which to attack — meaning, where to deploy capital aggressively to build a structural advantage.

My current bet, knowing how much I expect to revise this:

Defend the food margin. The platform fight is going to compress food costs and acquisition costs further before it stabilizes. Defense here means refusing to let the platform-ad spend balloon past a hard cap as a percentage of revenue, maintaining direct ordering channels, and treating the AI features the platforms sell you as bargaining chips, not value adds. Do not attack on this margin. The platforms have more leverage and a longer time horizon than you do.

Attack the labor margin. Capex for back-of-house automation has the clearest causal mechanism, the most credible public proof points (Sweetgreen, Chipotle), and the shortest payback. Attack here means committing to a multi-unit pilot in 2026, building the maintenance reserve into the model from day one, and being honest about the substitution-fallacy and throughput-trap failure modes. The labor thesis is the one where the operator can build a durable advantage on a 24-month horizon.

Watch the occupancy margin. Too early to commit. The Toast-Amex layer is forming but not yet priced; the alternative — a direct-bookings-and-CRM strategy — is more expensive and slower but preserves the marketing relationship. The right move is to keep both options open through Q1 and pick a side once the partnership pricing is visible. I will revisit in the September scoreboard.

Survive the capital margin. The vendor consolidation that is coming over the next twelve to eighteen months is going to leave some operators stranded on platforms whose acquirers do not love them. The defense here is contract optionality: shorter terms, data-portability clauses, dual-vendor strategies for anything that touches the demand graph. None of this is glamorous. All of it matters.

That is the framework, first draft. Defend food, attack labor, watch occupancy, survive capital. Four sentences for four margins, each of which deserves a full essay and will get one.

A note on what this is, and what comes next

I want to close with a methodological note, because I think the way I am building this framework matters as much as the framework itself.

This is the first sketch. It is going to be wrong in at least three places. I have a May refinement of this framework already on the calendar, and between now and then I am going to test the four-margin structure against every operator conversation, every earnings print, and every vendor pitch that crosses my desk. The framework lives or dies on whether it survives those tests. If it does, by next May it will be sharper, the margin definitions will be tighter, and the bets will be re-priced. If it does not, I will say so plainly and write a new one.

The second methodological note is about what I am explicitly not doing. I am not building a maturity model. I am not ranking AI vendors. I am not trying to be predictive about which platforms will win on which surface. The framework is a decision tool for the operator who has to allocate one balance sheet across four margins, not a market map for an investor or a feature checklist for a procurement team. There are good versions of all three of those documents being written elsewhere; this is not one of them.

The third note is about cadence. I am going to revisit this framework in every Mise essay through the end of the year — there is a September scoreboard already drafted that will pressure-test the four-margin call against next month’s data, an October stack-rank that re-prices the labor thesis, a November ledger that consolidates what we have learned across Q3, and a December reservations-endgame piece that focuses on the occupancy margin specifically. By the time the May refinement runs, this framework will have been through six months of iteration. The first sketch is the one that takes the longest to write and the one that is hardest to be right about. I would rather be partially wrong in public than wait six months for certainty I will not get.

The coffee is finally cold and the office is finally empty. Tomorrow is Labor Day weekend; Monday I will be on the road. Next week we get back to the regular cadence — earnings recaps, vendor analysis, operator interviews. But I wanted this on the page before September starts, because September will be a busy news month and the framework is what I plan to read everything through.

Defend food, attack labor, watch occupancy, survive capital. We will see how much of that survives the next thirty days.

— Eitan is editor-in-chief of TableTransfers. Tips: [email protected].