The Four Margins of AI Spend in Hospitality

It is the second-to-last Friday of the month — Mise cadence — and I am writing this from the same back-office desk where, for fifteen years, I used to read the daily report at eleven-fifteen at night. The desk is the same. The job is different. Two years ago the line item I argued with was the produce invoice. Today the line item I argue with is AI. Or, more honestly: today the line item I argue with is a stack of pitch decks from twenty-two different founders, each of whom is convinced that the right way to spend my AI budget is on their product. That is the difference between owning a restaurant group and editing a publication that covers them. The decisions are smaller; the volume of decisions is larger; the vocabulary is, if anything, worse.

I want to spend this Mise essay arguing that there are exactly four margins that hospitality AI can move, that almost every operator I talk to conflates them, and that the only AI bets worth writing a cheque for in 2026 are the ones that carry a measurable contract against one of the four. Framework first, evidence second, opinion third. This is editorial doctrine — the frame is mine, the examples are real and citable, and I will name where I am drawing the line between fact and read.

A note before the frame. I will come back to a more comprehensive Four Margins of a Restaurant essay later in the year — gross, operating, brand, enterprise. That essay is the lens this one sits inside. This essay is narrower: it is the four margins that AI spend can plausibly affect, sitting inside the operating-margin column of that bigger frame. The two essays are siblings, not duplicates. If you want the broader operator’s frame, that one is coming. If you want the AI budget frame, this is it.

The thesis

There are four margins that hospitality AI can move. I will call them labor, throughput, CAC, and shrink. Each has a different lever, a different evidence standard, and a different failure mode. A vendor who cannot tell you, in a single sentence, which of the four they move is selling you a story. A vendor who claims to move three of the four is selling you a worse story. The job of an operator in 2026 is to learn the four-margin vocabulary, demand a contract against one of them per dollar of spend, and refuse the rest.

The four, in one sentence each:

• Labor: reduce the hours required to deliver the same service.
• Throughput: increase the covers, transactions, or guests served per labor-hour, per square foot, or per minute of peak.
• CAC: lower the marketing and acquisition cost per booked or repeat guest.
• Shrink: reduce waste, theft, over-pour, over-prep, mis-portion, and inventory loss.

Two are cost-side (labor, shrink). Two are revenue-side (throughput, CAC). All four are operating-margin levers in disguise. None of them is the same as any of the others, and the confusion almost always sits in operators treating AI spend as one budget line rather than four. It is four budget lines. If you are not splitting the line, you are not reading the P&L.

The rest of this essay is one section per margin, anchored to an AI deployment that is real, in-market, and citable as of this morning. Where I cite a number I link to its source. Where I am offering a read, not a fact, I say so. Where a public deployment is forward-indicating but the print has not yet landed, I flag it as something to watch — operators making 2026 budget decisions today should not rely on numbers that have not yet been disclosed.

Margin one: labor — the McDonald’s drive-thru case

Labor is the margin most operators talk about when they say AI. It is also the margin most often promised and least often delivered. The National Restaurant Association’s 2025 data puts median labor cost at roughly 36.5% of sales for full-service and 31.7% for limited-service operators — a band wide enough that even a small AI-driven shift in scheduling or order-taking or coordination registers as real money on a $4M sales line.

The canonical labor-margin AI deployment is voice ordering at the drive-thru. McDonald’s has been the most public test case in the industry. Through the IBM-partnership era, the company piloted automated order-taking across roughly 100 US restaurants before ending the partnership in mid-2024 and signing a multi-year strategic alliance with Google Cloud (announced December 6, 2023) to rebuild the stack with generative AI for both crew tools and customer-facing surfaces. I will come back to the McDonald’s case in detail in a forthcoming desk review, but for the four-margin frame the relevant point is narrow: the labor lever is real, the technology is not yet mature enough to deliver on it in the highest-noise, highest-variance environment in QSR, and the second wave — with Google Cloud as the substrate and a more measured pilot footprint — is the one operators in 2026 should be watching, not the first wave.

What does the 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 strongest in-market labor proof-point is one I have written about before and will write about again: Sweetgreen’s Infinite Kitchen. Per CEO Jonathan Neman, IK-equipped stores have been running with a labor-cost advantage of “more than 700 basis points” against classic Sweetgreen stores of comparable age — a number that has been restated quarter after quarter through 2025. The IK is technically a throughput device that produces a labor dividend; I will come back to this in the throughput section. But the 700-bps figure is the cleanest single labor-margin number in the public restaurant tape, and the fact that it has held across multiple quarters is itself the strongest signal in the category. I covered the deployment in our Infinite Kitchen case study.

The voice-agent category is the labor lever’s other live front. The Voice Agent Maturity Curve essay — the second framework essay in this column, coming later in the spring — will lay out the five stages of voice-AI maturity and which operators should buy at which stage. For the four-margin frame today, the simple version: a Stage 2 reservation-taker (Yelp Host, Slang AI, SoundHound) is a labor contract dressed as a tech 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 covers the phone. That is a labor contract. Read it as one. Demand the baseline, demand the post number, and refuse to pay for vibes. The forthcoming Voice Agent Maturity Curve walks through which voice deployments are credibly at the labor-saving stage today and which are still aspirational.

The trap on labor is that the savings can appear as a one-time number rather than a recurring number. A vendor will quote you the host hours you displaced in month one and not the host hours that crept back into the schedule by month six because the agent fails on the long tail of intent. The labor margin contract has to be measured at month six, not month one. If a vendor is unwilling to be measured at month six, you have learned what you needed to learn about their conviction.

Margin two: throughput — the Sweetgreen kitchen case

Throughput is the most under-named margin in hospitality AI. It is the margin that increases the output per unit of input — more covers per labor-hour, more transactions per square foot, more order completions per minute of peak. It often produces a labor dividend, but it is not a labor margin in the strict sense. It is a productivity margin. The lever is different. The evidence is different. The failure mode is different.

The cleanest live throughput case in hospitality AI is, again, Sweetgreen’s Infinite Kitchen. The throughput claim is the one Sweetgreen has been making longest and most consistently: an IK store can run with materially fewer labor-hours per cover than a classic Sweetgreen store, because the bowl-assembly robotics handle the steps that used to require two to three crew members in a brigade line. The 700-basis-point labor advantage I cited above is the output of the throughput lever, not the throughput lever itself. The throughput lever is the robotic kitchen’s ability to assemble bowls faster than a human line at peak, with lower error rates and less per-cover labor exposure.

The reason this distinction matters: an operator who buys an Infinite Kitchen expecting a labor contract will be disappointed when the labor number takes eighteen months to settle. An operator who buys it expecting a throughput contract will measure the right thing — bowls per minute at peak, error rate, cover-mix at the IK window versus the classic line — and will know within a single quarter whether the deployment is working. The Sweetgreen team has been disciplined about which number they put at the top of the slide on each earnings call, and the discipline is itself instructive: the labor number is the durable number, but the throughput number is the leading one. Operators evaluating any kitchen-automation deployment in 2026 should be asking for the throughput print first and the labor print second.

The other throughput frontier — and one I think is more underweighted by the operator community than the kitchen-automation conversation — is the KDS layer. Kitchen display systems with intelligent prep-time prediction, dynamic ticket reordering, and cross-station coordination are a software-only throughput lever. No robotics, no capex. Several Toast IQ surfaces fall in this category: the Toast IQ smart AI assistant launch (October 29, 2025) extended Toast IQ from feature-level intelligence to assistant-level intelligence across roughly 148,000 customer locations. 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.

The trap on throughput is that the margin appears in peak-hour operations and disappears 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. Operators who measure throughput on the trailing-thirty average will miss the contract entirely. Measure it on the peak-hour ninety-fifth percentile. Vendors who do not understand that distinction are not yet ready to sell you a throughput product.

Margin three: CAC — the Toast and Sysco cases

CAC — customer acquisition cost — is the third margin, and it is the one where the vendor category is most crowded, the operator confusion is highest, and the evidence is hardest to read. Loyalty platforms, marketing-AI platforms, reservation-aggregator AI, email and SMS personalisation engines, the entire surface of “we will help you bring guests in” — all CAC theses. All using the same vocabulary. Most measuring different things.

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 it 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.

The strongest current proof-point in CAC-side hospitality AI is the back-door enablement work Sysco has been doing with AI360. On the Q2 fiscal 2026 earnings call this past Tuesday (January 27, 2026), CEO Kevin Hourican disclosed that “95% or more of our colleagues are using the tool weekly” — a roughly 90-day-old deployment that has reached near-universal adoption inside the sales force. The CAC frame on AI360 is not a guest-acquisition frame; it is a customer-acquisition-and-retention frame on the distributor side. The Sysco sales rep is paid to land and keep operator accounts. AI360 is the tool that tells the rep which accounts to call, what to recommend, and which “swap and save” substitutions to surface. The contract is measurable in two numbers Hourican has been disclosing across calls: rep retention and operator-loss rate. Both have improved since deployment. That is a CAC contract with a measurable print.

Reframe Sysco’s deployment in operator-side language and you get a useful generalisation: a CAC AI tool that reaches 95% weekly active usage in 90 days is a tool that is doing real work; one that gets stuck at 30% adoption is a tool that has not earned its keep regardless of the headline ROI claim. Adoption is the leading indicator. The contract is the lagging one.

The CAC frontier I am watching most closely on the guest-facing side is the Toast surface. Toast has been deepening its loyalty and marketing-AI footprint across 2025; the Toast IQ smart AI assistant announcement (October 29, 2025) extended the platform’s AI surface across roughly 148,000 customer locations. The most useful operator-side CAC question to ask of any Toast IQ surface — or any equivalent CAC-side product from a competitor — is the cost per repeat visit number. Toast has been disclosing recurring gross-profit growth and retention metrics on its quarterly calls; the Q1 2026 print, due in early May, will be the next data point operators should mark on the calendar. Forward-looking note: that print is post-Jan-30 and so should not anchor any budget decision being made today. It should anchor the conversation in May.

The trap on CAC is that the lever is shared with the brand. 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. Refuse to pay.

The reservation-aggregator CAC battle — DoorDash, Resy, OpenTable, Tock and the wave of AI-driven personalisation surfaces being layered on top of each — is the noisiest theatre in this category right now. The competitive dynamics are evolving fast enough that I will not try to summarise them here; the early-2026 trade press is full of the back-and-forth, and operators evaluating any reservation-platform AI feature should read those pieces with the CAC contract in mind. The contract is the same as for any other CAC tool: cost per booked guest, before and after. A reservation platform whose AI feature does not reduce that number is not selling you a CAC tool. It is selling you a feature.

Margin four: shrink — the inventory case

Shrink is the fourth margin and the most under-named. It 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 is doing real, less-hyped work and where the evidence standard is unusually clean — because shrink is, by definition, a counting problem, and AI does counting well. The lever is inventory intelligence: the joining of POS sell-through data with delivery receipts, prep counts, par levels, and waste logs to produce a real-time theoretical-vs-actual variance for every SKU on the cost line.

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 Sysco AI360 surface I cited under CAC also has a shrink-adjacent component on the back-door side: the “swap and save suggestions” Hourican described on the Q2 2026 call — data-science-driven product substitutions cuisine-specific to the restaurant. The framing on the Sysco call was that the substitutions benefit the customer, Sysco’s margin, and the rep’s compensation simultaneously. The operator-side benefit is shrink-adjacent: a cheaper substitute that performs identically at the cover is a shrink-equivalent gain on the cost line. The deployment is too new for a print, but the structure of the contract is the right one — a measurable cost-of-goods improvement against a defined SKU.

Inventory-AI is where most of the shrink-margin pitch volume is concentrated. The vendor landscape is fragmented, the language is messy, and the proof-points are unevenly distributed. An operator evaluating any inventory-AI tool in 2026 should run the contract math first and the technology demo second. The four questions: Which SKUs does the tool cover? (You will be amazed how many tools cover beverage but not centre-of-plate, or vice versa.) What is the integration with my POS and my distributor invoices? (No data joining, no contract.) What is the baseline variance number you will measure against? What is the post-deployment variance target, with a date attached?

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 is what tells you whether the AI is doing work. The actual line is what tells you whether the supplier is doing work. Two different questions.

Where this frame helps

The four margins compose. An honest hospitality AI tool moves one of them. An ambitious one moves two. A vendor claiming three is selling you a worldview, and the worldview is rarely backed by a print. The frame helps because it forces the vendor to specify, in your first meeting, which line on your P&L the contract attaches to.

Three places the frame is immediately useful inside an operator’s week.

First, the AI budget line itself. If you have a single AI budget in your 2026 P&L, you cannot tell which of the four margins your spend is producing. Split the line into four sub-lines — labor AI, throughput AI, CAC AI, shrink AI — and assign every contract to one of them. At the end of the year, you will know which lever your business actually moved. This is a five-minute change in your chart of accounts that will pay back in your second AI budget cycle.

Second, the vendor diligence call. Open every vendor call with: which of these four margins does your product move? If the vendor cannot answer in one sentence, end the call. If the vendor answers with three or four, follow up with: if you can only move one, which would you choose to be measured on? The answer to the second question is the only one that matters. A vendor who refuses to pick is a vendor who has not done the work.

Third, the post-deployment review. Six months after any AI deployment, run the contract. Did the labor number move? Did the throughput number move? Did the CAC number move? Did the shrink number move? If yes, write the next cheque. If no, kill the contract regardless of how much you like the team. Hospitality is an industry that has tolerated too many interesting vendors and not enough productive ones. The four-margin frame is the corrective.

Where this frame fails

Three places the frame is incomplete and the operator should know it.

First, brand-margin work. The four margins above all sit inside the operating-margin column. They do not capture the brand-margin work that an AI surface can do for an operator with strong concept clarity — the way a custom-persona voice agent at a luxury hotel becomes part of the property’s voice, or the way a recommender system at a wine programme reinforces a sommelier’s editorial point of view. These are real benefits. They are not captured here. The forthcoming Voice Agent Maturity Curve essay’s Stage 4: Brand Voice tier is where this work lives, and the broader four-margins-of-a-restaurant frame I am working toward later in the year will give brand its own column. For today: if a vendor is selling you a brand asset, do not force it into one of the four operating-margin buckets. Recognise that you are buying something different and budget it from a different line.

Second, enterprise-margin work. Some AI deployments are not operating-margin moves at all — they are enterprise-value moves, intended to make the business more valuable to a future buyer or to public-market investors. The deepening of Toast’s AI surface is, properly read, partly an operator-side ROI play and partly an enterprise-value-of-Toast-Inc. play. Both are real. Operators evaluating a vendor whose AI roadmap is genuinely an enterprise-margin play (i.e., the vendor is investing to deepen their moat against future competitors and to command a higher acquisition multiple) should ask honestly whether the operator-side contract is strong enough to justify the deployment on its own terms. If the answer is no, walk. The vendor’s enterprise upside is not your operator-side upside.

Third, agentic-coordination work. The most credible long-arc AI bet in hospitality — what I will call Stage 5 in the voice-agent maturity curve, and what is appearing as an emerging category in distributor and supplier coordination — is not yet measurable on a 2026 contract because it does not yet exist at scale. Back-office labor reduction at the chain level, supplier-call automation, vendor-coordination agents — these are 2028–2030 categories, not 2026 categories. Operators should be tracking them; they should not be writing six-figure cheques against them today. The vendors selling Stage 5 capabilities in 2026 are mostly selling 2028 promises with 2026 invoices.

What to watch in 2026

A short list of forward-indicators every operator should mark on the calendar:

• Toast’s Q1 2026 print, due early May. The recurring-gross-profit growth number and any updated commentary on Toast IQ adoption will be the cleanest CAC-side data point in the public restaurant tape.
• Sysco’s continued AI360 disclosure cadence. Hourican has been unusually generous with the adoption number on each call. Watch whether the productivity per rep and operator-loss rate numbers continue to improve quarter-on-quarter.
• The reservation-aggregator competitive arc. DoorDash’s SevenRooms integration is twelve months old this spring; the AI surfaces being layered across the OpenTable and Resy bases are evolving fast enough that 2026 trade coverage will be the right place to follow the CAC-side competitive dynamics. The mid-Q1 reads in the major business press will be the watch-list.
• Sweetgreen’s Q4 2025 print, due February 26. The Infinite Kitchen labor and throughput numbers are the single most-watched live AI deployment in fast-casual. The 700-bps figure restated again will be a strong tell; revised lower will be a stronger one.
• Hotel-side AI spend disclosure. Multiple major hotel groups have been increasing their AI line on annual capex disclosures; the survey work coming out of the Q1 industry conferences will be worth reading carefully when it lands. Forward note: hotel-side AI surveys often blur the four-margin lines worse than the restaurant tape. Read them with the frame in hand.

How to budget AI spend in 2026

I will close with the operator-actionable version. Six steps. They are the steps I would walk through if I still owned my restaurant group today and was sitting at a desk with twenty-two pitch decks and a finite cheque-book.

Step one: split the AI line into four. Labor AI, throughput AI, CAC AI, shrink AI. Pick a dollar budget for each. The split itself is the work — most operators will discover, in the first hour, that they have implicitly been spending 80% of their AI budget on one margin without ever having decided that was the strategy.

Step two: identify the binding constraint in your operating margin. Read your trailing-twelve P&L and ask which of the four lines is the worst-performing relative to the benchmark for your concept. If your labor line is tracking high, the labor budget gets the largest cheque. If your food cost is tracking high, the shrink budget gets it. Do not spread the AI budget evenly across the four margins because that feels prudent. The four-margin frame is a triage frame, not an equality frame.

Step three: for the margin you are spending most against, do a vendor bake-off. Two vendors, same scope, same contract structure, same measurement window. Pick the one whose contract structure you trust most, not the one whose demo was prettiest. A vendor who balks at a 90-day measurement window has told you what they think of their own product.

Step four: build the measurement before the deployment. Baseline the metric you are about to move. Write it down. Date it. The single most expensive mistake operators make in AI procurement is deploying first and trying to baseline after. By the time you remember to measure, the post-deployment number has nothing to compare against, and the vendor’s claim becomes unfalsifiable.

Step five: at the six-month mark, run the contract honestly. If the metric moved, write the next cheque and consider expanding scope. If the metric did not move, cancel. Hospitality has tolerated too many AI experiments that quietly stayed on the P&L for a second year because no one wanted to admit the first year had not worked. Do not be that operator.

Step six: at year-end, look at the four sub-lines and decide whether the split was right. The whole point of the four-margin budget structure is that it tells you, retrospectively, which margin your AI capital actually compounded against. Use that signal to redo the split for next year. The frame compounds across budget cycles. The vocabulary is the lever.

The forward reference

This essay is the first of two framework pieces I am building toward this year. The second — coming later in the spring — will lay out a more comprehensive Four Margins of a Restaurant frame: gross, operating, brand, enterprise. That frame is the lens this AI-specific frame sits inside. The relationship between the two essays is straightforward: the four AI-spend margins above are all sub-levers inside the operating margin of the bigger frame. The bigger frame will explain why operating margin alone does not tell you whether the last six months of work have made your business more valuable. This frame, the AI-specific one, is the operator’s tactical view; the bigger frame is the owner’s strategic view. Both are needed. Neither is sufficient on its own.

If you are reading this in January 2026 and trying to write next year’s AI budget, this essay is the one you need. If you are reading this six months from now and trying to decide whether to sell the restaurant group at all, the second one will be. The vocabulary is the lever for both.

— Eitan writes Mise. He founded TableTransfers after selling his last restaurant group. Tips: [email protected].