The Four Margins, Refined: An AI ROI Framework for Hospitality After the February 2026 Print Cycle

It is the last Friday of February. The first full earnings cycle of the year has landed — Toast on the twelfth, Yum on the fourth, Sweetgreen yesterday, Burger King’s Patty announcement and Slang’s Series B in between — and I have spent the better part of a week sitting at the same back-office desk I sat at four weeks ago, rereading my own January Mise essay against what the prints actually said. The honest answer is that the frame I published on January 30 needs work. Not a tear-down; a refinement. The labels were wrong on two of the four. The shape of the argument is the same, the discipline is the same, but the vocabulary has to keep up with the evidence. Mise is supposed to be a working frame, not a museum piece, so this essay is the second iteration.

If you read The Four Margins of AI Spend in Hospitality on the second-to-last Friday of January, you remember the four labels I proposed: labor, throughput, CAC, shrink. That was the right shape and two of the four labels were right. After watching the February print cycle land, I am moving CAC to accuracy and I am moving shrink to guest data. The shape of the argument is unchanged — four levers, one per dollar of spend, contract on the line — but the labels now sit closer to where the public-tape evidence is actually accumulating. I will explain each rename below, the deployment that forced it, and why the new labels do operator-side work the old ones did not.

A brief note on the broader frame before the refinement. The Four Margins of a Restaurant essay — gross, operating, brand, enterprise — I have been threatening to write since January is still coming, later in the spring. That essay sits one level above this one. The four AI-spend margins below are all sub-levers inside the operating-margin column of the bigger restaurant frame. If you want the broader owner’s view, that one is the May piece. If you want the AI-budget view, this is it, updated.

The thesis, restated

There are four margins that hospitality AI can credibly move in 2026. After the February print cycle I am calling them labor, throughput, accuracy, and guest data. 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 — and the print cycle just past has reinforced this, not weakened it — 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.
• Accuracy: reduce the errors — order, prep, inventory, fulfilment, communication — that compound silently into lost margin.
• Guest data: capture, structure, and own first-party guest information that becomes a durable acquisition and retention asset.

Two are cost-side (labor, accuracy). One is mixed (throughput, which produces a labor dividend but is structurally a productivity lever). One is a balance-sheet item dressed as an operating-margin item (guest data, which is technically an enterprise-margin lever expressed inside operating-margin tooling — I will come back to this). All four are real. None is the same as the others. The confusion sits, as it did in January, in operators treating AI spend as one budget line. 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 a real deployment cited in the February print cycle. Where I cite a number I link to its source. Where I am offering a read, not a fact, I say so.

Margin one: labor — Burger King’s Patty and Sweetgreen’s Infinite Kitchen

Labor is still the margin most operators think of when they say AI, and the February prints reinforced rather than complicated the picture. Two deployments are doing the live work here, and they sit at opposite ends of the same lever.

Burger King’s Patty assistant — announced into pilot last week and covered in the Nation’s Restaurant News piece — is running in 500 restaurants, with rollout to all U.S. locations targeted by year-end 2026. The system runs on cloud-connected headsets powered, per the NRN reporting, by OpenAI’s base technology layered with Burger King’s proprietary architecture. Tom Curtis, BK U.S. president, framed the strategic intent in operator-language rather than technology-language: “When managers and team members can focus more on leadership and customer interaction, growth follows.” RBI chairman Patrick Doyle called it a “game changer.” I read the deployment as a labor-margin bet first — the explicit theory of change is that managers and crew spend less of their shift on the operational micro-decisions Patty can intermediate (menu-item availability, prep guidance, drive-thru coaching) and more on the higher-leverage work of running a service period. That is a labor contract. The print to watch is whether crew hours per occasion or manager hours per shift compress meaningfully in the 500-store cohort before the rollout decision lands.

The deeper labor proof-point — the one I cited in January and which Sweetgreen reprinted on its Q4 2025 call yesterday — is the Infinite Kitchen labor figure. Per the Sweetgreen Q4 2025 earnings call, CEO Jonathan Neman repeated the line operators have been watching for two years: “Established Infinite Kitchens delivered more than 700 basis points in labor savings over classic locations of similar age.” The company opened 8 Infinite Kitchen restaurants in Q4 2025, finished the year with 30 IK locations, and has added two more in Q1 2026 (Long Beach and Pike 7 in the DMV market), bringing the current count to 32. The 700-bps line restated again, against a now-meaningful denominator of 32 stores, is the strongest single labor-margin number in the public restaurant tape. The fact that the number has not moved across multiple quarters and a growing store base is itself the signal — labor savings from automation either compound or evaporate, and Sweetgreen’s are compounding.

A note on Sweetgreen’s Sweetlane drive-thru format — the Costa Mesa pilot opened in November and is, per Neman’s call yesterday, “performing well” — because the format matters for the labor-margin frame. Sweetlane embeds Infinite Kitchen technology inside a suburban drive-thru. If the format prints comparable labor economics in a higher-throughput, lower-cover-mix environment, it will be the cleanest evidence the industry has that kitchen automation generalises beyond the chassis Sweetgreen built it inside. That is the watch-item for the May print.

The labor-margin contract, written properly, is the same as it was in January and the February prints have not changed it. A defined cost line on the P&L the vendor commits to reducing. A measurable baseline of at least 30 days before deployment. A measurable post-deployment number, same metric, same time window, six months in. If the vendor cannot produce all three components in writing, the contract is not a labor contract; it is a sentiment contract. The Burger King Patty pilot is structured the right way — 500 stores, measurable cohort, defined rollout gate. Operators evaluating any voice-agent or back-of-house assistant in 2026 should ask whether their vendor would accept the same pilot structure RBI accepted.

Margin two: throughput — Toast IQ and (still) the Sweetgreen kitchen

Throughput is the under-named margin and the February cycle did not change that. It produced one major data point, though, which deserves to be read carefully.

Toast disclosed on its Q4 2025 earnings call on February 12 that Toast IQ, the company’s conversational AI assistant launched in late October, has been used by over half of all locations within four months of launch and has handled more than 8 million queries across the platform. CEO Aman Narang’s framing on the call was deliberately operator-language: “ToastIQ help our teams quickly make decisions and turn hours of menu analysis into clear, actionable insights in just minutes.” Half-of-locations adoption inside four months is, on its own, the cleanest adoption print any hospitality-AI surface has produced in the public restaurant tape.

Read that print through the throughput lens. Toast IQ is structurally a productivity lever before it is a cost lever. The promise is that a manager who used to spend two hours pulling menu performance numbers, building an email campaign, or running an inventory analysis can now do the work in minutes. That is throughput at the back-office layer — output per unit of input. The labor savings exist (an hour saved on menu analysis is an hour reallocated) but they are diffuse and unbookable. The throughput gains are concentrated and bookable: more campaigns shipped, more menu adjustments per quarter, more inventory cycles per month. Operators reading the 50% / 8 million print as a labor story will under-budget the deployment. Reading it as a throughput story tells them to measure output-per-manager-week, not manager-hours-saved. Two different contracts.

The kitchen-throughput case, again, is Sweetgreen. The 700-bps labor number from yesterday’s call is the output of the throughput lever, not the throughput lever itself. The lever is the IK’s ability to assemble bowls faster than a classic line at peak, with lower per-cover labor exposure. The Q4 disclosure confirmed something I underweighted in January: the IK is now operating at sufficient scale (30 stores at year-end) that the throughput pattern is no longer pilot-bounded. It is system behaviour. Operators evaluating kitchen-automation in 2026 should be reading the IK number as a productivity benchmark rather than a robotics benchmark — the lever is the productivity gain at peak, and the technology underneath it is implementation detail.

The throughput contract has not changed from January. Tickets per hour at peak. Average ticket-completion time. Mis-fire rate. Output per back-office unit of time on the management-AI side. Ninety-fifth-percentile peak-hour measurement, not trailing-thirty averages. Any vendor who balks at peak-hour measurement does not yet understand their own throughput product.

Margin three: accuracy — Byte by Yum and Patty (and why I dropped CAC)

This is the rename, and I want to do the explanation honestly rather than burying it.

In January, the third margin was CAC. The argument was that AI-driven marketing, loyalty, reservation, and aggregator surfaces all reduce the cost of acquiring a booked or repeat guest, and that operators should contract against cost-per-booked-guest. The frame was correct as far as it went. It was also, on reflection, badly placed — because CAC is one operator-side outcome of a much wider class of AI deployments whose actual common lever is something else. The actual common lever, surfaced clearly in the February print cycle, is accuracy.

Watch Byte by Yum. On its Q4 earnings cycle covered by PYMNTS this month, Yum disclosed the platform numbers: up to 85% reduction in stockouts, up to 75% reduction in aggregator ordering failures, and up to 10% improvement in consumer satisfaction. CFO Ranjith Roy described the platform as “the only multi-brand, multi-market QSR technology platform built by restaurant operators for restaurant operators,” and CEO Chris Turner framed the data-ownership thesis explicitly: “Owning our core digital and technology platforms gives us an edge over the competition.” The Smart Ops bundle is in 7,000 restaurants; the Digital Ordering bundle is in 18,000; the total Byte-product footprint is 38,000 restaurants at year-end 2025.

Now read the three Byte numbers. Stockouts are an accuracy problem (the system inaccurately predicted demand or inaccurately tracked inventory). Aggregator failure is an accuracy problem (the integration inaccurately transmitted an order, an item, or a price). CSAT improvement is the operator-visible symptom of accuracy gains — the guest got what they ordered, when they ordered it, the way they ordered it. The three numbers compose into a single thesis: the lever is accuracy, and accuracy compounds into both cost reduction and revenue retention simultaneously.

Patty fits the same frame from the labor end. Per the NRN piece, the system’s specific functions include automatic menu-item removal when products become unavailable, real-time guidance on food preparation, and drive-thru audio analysis for order accuracy and coaching. Two of those three are explicitly accuracy levers — the item-removal function reduces order-failure errors at the moment the guest is told what is available, and the drive-thru audio analysis is, by its own framing, accuracy coaching. The labor-margin frame I used on Patty above is the promised contract; the accuracy-margin frame is the underlying lever. Operators evaluating Patty-equivalent deployments in 2026 should hold both contracts simultaneously and measure both.

Why is accuracy the right rename for CAC? Three reasons.

First, accuracy is the lever and CAC is one downstream outcome of it. A reservation platform that accurately matches a guest to a table reduces no-shows, increases repeat bookings, and lowers effective CAC — but the lever is accuracy of the match, not the marketing dollar. A loyalty surface that accurately identifies the right offer for the right guest at the right time lowers CAC — but again, the lever is the accuracy of the targeting. By naming the lever directly, the contract becomes cleaner. You measure the accuracy first (error rate, match precision, prediction quality) and the CAC second. The accuracy number is the leading indicator; the CAC number is the lagging one. Mis-name the lever and you measure the wrong thing.

Second, accuracy captures the shrink work as well. In January I named shrink as a separate margin. The February evidence suggests shrink is, more honestly, the inventory face of accuracy — an inaccurate par level produces over-prep, an inaccurate count produces shrink, an inaccurate forecast produces waste. Byte’s 85% stockout reduction is the inverse-shrink number written in QSR vocabulary. Collapsing shrink into accuracy is the cleaner taxonomy because it puts inventory accuracy, order accuracy, and prediction accuracy under one lever. (I am promoting guest data into the freed-up slot, below.)

Third, accuracy is the language operators actually use in service. Walk a kitchen at the Saturday-night push and you will hear accurate and inaccurate a dozen times. You will hear CAC zero times. The frame should match the language of the work, not the language of the deck.

The accuracy contract is the cleanest of the four to write. Before deployment: error rate on the target metric (stockouts per location-week, order-failure rate, mis-pour rate, forecast variance), measured for at least 60 days. After deployment: same metric, same definition, 90 days post-go-live. The vendor who delivers accuracy improvement shows you an error line that compresses. The vendor who does not, does not. There is, as I wrote in January about shrink, essentially no room for sentiment in the contract — the math is the math. Byte by Yum’s up to 85% stockout figure is a vendor-disclosed range, not a contracted floor; operators should read it as the upper bound of the achievable, not the median. The median is the contract.

The trap on accuracy is the up to qualifier. Vendors will quote you the best cohort’s improvement and let you assume your fleet will match. Demand the median number, not the up to number. A vendor whose median is half the headline is not a fraud; they are an honest reporter of variance. A vendor whose median is the headline has either run a one-store pilot or is selling you back-of-deck math.

Margin four: guest data — Slang’s Series B and Toast’s multi-unit play

This is the addition, and the harder argument. In January I had four margins; I am keeping four margins; but the fourth slot is now occupied by something the January frame did not name: guest data as a margin in its own right.

Why add it. Because the February print cycle made it harder to ignore that an entire class of AI deployments is structurally about capturing and owning first-party guest information — and that this capture is the actual contract, even when the vendor pitch is in labor or accuracy vocabulary. The Slang AI Series B announcement is the clearest February example.

Per the Slang AI press release, the company raised $36 million Series B ($28M equity, $8M debt) led by US Venture Partners, bringing total funding to $68M. The product is now deployed at 2,000+ restaurant locations globally. Guest satisfaction on agent-handled calls is reported at 95%+. Phone reservations are reported as 2x. The numbers operators will pull from the headline are the operator-side ROI claims — Slang has published case-study figures of up to twentyfold ROI for operators, and customer testimonials including Dan Simons of Founding Farmers Restaurant Group describing how “Slang AI has fundamentally impacted our business” by capturing missed reservations and event leads while freeing staff to focus on in-person service. The DineAmic Hospitality case study Slang has circulated in operator channels references 13x ROI, six-figure incremental reservation revenue, and roughly two thousand dollars per month in displaced host labor at the deployed properties.

Read those numbers carefully. The 13x ROI is genuinely large. The displaced host labor is real. The incremental reservation revenue is real. But the line that should make every operator stop and reread is the one in Slang’s own platform disclosure: 25 million customer calls from 10 million unique guests. That dataset — the structured transcripts, intents, preferences, and reservation patterns of ten million identified diners — is the asset the company is being funded to build. The operator-side labor contract is the wedge. The guest-data asset is the durable position.

This is not a critique of Slang. The product is good, the labor savings are real, and the dataset position is reasonable for a vendor to pursue. It is a clarification about which margin the operator is actually paying into. A restaurant deploying Slang is buying a labor-and-accuracy contract on the operator side and contributing a row to a guest-data position on the vendor side. Both transactions are happening. Operators who think they are only buying the first transaction are paying for the second one without recognising the trade.

The same frame applies to Toast IQ at the multi-unit layer. Toast’s Q4 disclosure was that 8 million queries have been sent across half its locations in four months. The single-operator value is the productivity lever I named under throughput above. The multi-unit value, and the reason Toast’s roadmap explicitly extends toward “autonomous agents handling complete business functions” (per Narang’s commentary), is the cross-customer dataset Toast is building of how restaurant operators actually run their businesses. That dataset is the moat. Narang said the quiet part out loud on the call: ToastIQ’s “advantage derives from embedded restaurant domain expertise rather than generic AI.” The domain expertise is the data.

This is the deeper rhyme with the AmEx-Resy-Tock and DoorDash-SevenRooms moves I have written about in passing — and which my colleague Oliver has written more directly in his anti-AI-premium buy-side note. The platforms acquiring reservation and guest-data infrastructure are not buying restaurants; they are buying the data layer above restaurants. Vendors selling AI surfaces into operator workflows are, in many cases, doing the same thing one row at a time. The guest-data margin is the lever this category of deployments is actually pulling.

What does the guest-data contract look like, written for an operator? Three components. First, data ownership clarity: who owns the guest record, who owns the call transcript, who owns the inferred intent, and who can use that data for which downstream purposes — including for the vendor’s own model training. Second, portability: if the contract ends, what data leaves with the operator and in what format. Third, reciprocity: what the operator gets in cross-customer model improvements, what fraction of the platform’s data position the operator benefits from, and whether the vendor’s network-effect compounding accrues even partially to the participating operators. These are the questions a 2026 contract for any AI-driven guest-communication or guest-personalisation product should answer in writing.

The trap on guest data is treating it as a side-effect. The Slang and Toast prints make the point: this is not a side-effect. It is the asset that funds the next round, builds the moat, and makes the vendor acquirable at a multiple no operator-side ROI calculation could justify. Operators are paying for the operator-side ROI. They are not paying for the guest-data asset, which they are contributing to and not capturing. The four-margin frame, post-refinement, names this directly.

Why CAC moved and shrink moved

Two of the four label changes deserve direct accounting. I do not want to silently rebadge the frame.

CAC moved to a downstream metric. It is still real and still measurable, but it is not the lever — it is the outcome of accuracy at the targeting, matching, and personalisation layer. Operators measuring CAC after deploying a reservation AI or a loyalty AI will see the number move, but they will not be measuring the right cause number. Measure accuracy first. The CAC compresses if and only if the underlying error rate compresses. The reverse implication does not hold cleanly, because CAC can move for brand reasons and seasonality reasons that have nothing to do with the AI. Accuracy is the leading indicator. CAC is the lagging one. The frame names the leading indicator.

Shrink moved inside accuracy. Shrink is a subset of accuracy: inaccurate forecasting, inaccurate counting, inaccurate par-setting, inaccurate prep produces shrink as a downstream cost. The Byte by Yum stockout figure is shrink-adjacent (the inverse case — accuracy of supply matched to demand). Collapsing shrink under accuracy is taxonomically cleaner and gives operators one lever to budget against rather than two overlapping ones.

I want to be honest about what is lost in the rename. CAC as a separate margin was useful precisely because it forced an operator to budget against marketing-AI as a distinct line from operations-AI. The new frame loses that separation. The compensating discipline, post-refinement, is that the accuracy contract has to be written with sub-lines: order accuracy, inventory accuracy, prediction accuracy, targeting accuracy. If you run one accuracy budget for the four sub-lines combined, you will under-fund the right one. Split the line one level down.

Why guest data is on this list and not on the next one

A reasonable objection: guest data is an enterprise-margin lever, not an operating-margin lever. Owning a first-party guest data asset increases the value of the business at exit; it does not, on its own, move the operating P&L next quarter. By the strict logic I used in January, guest data should belong to the broader Four Margins of a Restaurant frame coming in May — the one with gross, operating, brand, and enterprise columns — and not to the AI-spend frame for 2026.

I considered that. I am putting guest data on this list anyway, for one reason: the AI surfaces sold to operators in 2026 force the guest-data trade whether the operator is reading for it or not. A reservation AI, a voice agent, a personalisation engine, a marketing-AI surface — every one of these involves the operator’s guest data flowing through and contributing to the vendor’s dataset. Treating that flow as out-of-scope for an AI-spend budget conversation is how operators end up paying for an asset they do not capture. The frame has to name it inside the AI budget, even if the strategic frame in May puts it in the enterprise column.

When the May essay lands, the four-margins-of-a-restaurant frame will give guest data its own column under enterprise margin. The AI-spend frame and the restaurant frame are siblings, not duplicates. Both are needed. The vocabulary in this frame, post-refinement, anticipates the May one without collapsing into it.

How the February cycle confirmed three things from January

Before the budgeting section, three points from January that the February prints confirmed and that I want to name explicitly so the corpus has them in one place.

One: adoption is the leading indicator, contract is the lagging one. I made this point in January about Sysco’s AI360 — 95% weekly active usage inside ninety days was the signal. Toast IQ’s 50% of locations inside four months is the same signal at a much larger denominator. A surface that reaches half its installed base inside four months of launch is a surface doing real work; one that gets stuck at 30% adoption is not, regardless of the headline ROI claim. Operators reading vendor pitches in 2026 should ask the adoption number first.

Two: the lever names matter more than the technology names. Vendors will pitch you agents, copilots, LLM-powered platforms, intelligent automation. None of those words tell you which margin moves. The four-margin labels — labor, throughput, accuracy, guest data — are operator-side language and they survive vendor-side fashion cycles. Use the operator-side language in every meeting. The vendor will catch up.

Three: the 700-bps figure is still the strongest single number in the public tape. Sweetgreen restated it on yesterday’s call against 32 stores, up from a much smaller denominator two years ago. The number’s durability is the signal. The number itself is the headline.

How to budget AI spend in 2026, post-refinement

The operator-actionable version. Six steps, refreshed from the January version.

Step one: split the AI line into four, with the new labels. Labor AI, throughput AI, accuracy AI, guest-data AI. Pick a dollar budget for each. The split 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, and your binding constraint on the data side. Read the trailing-twelve P&L. Ask which of the first three lines (labor, throughput, accuracy) is the worst-performing relative to the benchmark for your concept. Separately, ask whether you currently own your first-party guest data or whether it sits inside vendor systems you do not control. The first three are operating-margin questions. The fourth is an enterprise-margin question. Both feed the AI budget.

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. The contract structures are different for each margin: labor contracts measure baseline-to-post hours; throughput contracts measure peak-hour ninety-fifth percentile; accuracy contracts measure error-rate compression; guest-data contracts measure data ownership clarity, portability, and reciprocity. Pick the vendor whose contract you trust, not the one whose demo was prettiest.

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. This was true in January and the February prints did not change it.

Step five: at the six-month mark, run the contract honestly. If the metric moved, write the next cheque. If it did not, cancel. Hospitality has tolerated too many interesting vendors and not enough productive ones. The four-margin frame, refined, is the corrective.

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

Where this frame still fails

Three places, restated and refined.

First, brand-margin work. The four margins above all sit inside the operating-margin column of the bigger restaurant frame. They do not capture brand-margin work — the way a custom-persona voice agent at a luxury property becomes part of the brand’s voice, or the way a recommender system at a wine programme reinforces a sommelier’s point of view. The forthcoming Voice Agent Maturity Curve essay’s Stage 4: Brand Voice tier is where this work belongs. If a vendor is selling you a brand asset, do not force it into one of these four buckets. Budget it from a different line.

Second, enterprise-margin work beyond guest data. I have promoted guest data into the AI-spend frame as a fourth margin because the operator-side budget conversation cannot avoid it. But guest data is one of multiple enterprise-margin levers an AI roadmap can move — the others (vendor lock-in, network effects, multiples-expansion for the vendor at exit) sit further from the operator’s quarterly P&L. They belong to the May Four Margins of a Restaurant frame. Read the broader frame when it lands.

Third, agentic coordination at scale. What I called Stage 5 in January — back-office labor reduction at the chain level, supplier-call automation, vendor-coordination agents — is not yet measurable on a 2026 contract because it does not yet exist at scale. Toast’s stated roadmap toward “autonomous agents handling complete business functions” is the closest live signal, and even there the productisation is one to two years out. Track it. Do not write six-figure cheques against it today.

What to watch in the second-quarter print cycle

A short watch-list, refreshed.

• Toast’s Q1 2026 print, due early May. The next ToastIQ adoption number — the four-month-from-launch figure was 50%, the seven-month-from-launch figure will be the better signal — and any disclosure on multi-unit data usage will be the cleanest data point in the tape.
• Sweetgreen’s Q1 2026 print, due early May. Whether the 700-bps figure holds against 32+ stores. Whether Sweetlane prints comparable economics in a drive-thru chassis.
• Yum’s continued Byte disclosure. The stockout, aggregator-failure, and CSAT numbers were presented as up to figures in February. The Q1 print should give us the median.
• Burger King’s Patty pilot. The 500-store cohort is the right pilot size for a labor-margin contract. Watch for any disclosure on crew-hours-per-occasion in the May earnings cycle.
• Slang’s deployment trajectory. 2,000 locations now, the dataset position is 25M calls and 10M unique guests. The next funding event or platform-disclosure will tell us how the guest-data asset is being valued and whether operator-side contracts are evolving to reflect it.

The forward reference, restated

This essay is the second of three framework pieces I am building toward this year. The first was the January AI-spend frame, which this essay refines. The third — landing in May, sitting in the cover-story slot — is the broader Four Margins of a Restaurant essay: gross, operating, brand, enterprise. That frame is the lens this AI-spend frame sits inside. The four AI-spend margins above are sub-levers inside the operating-margin column of the bigger frame, with one important exception: guest data, which is structurally an enterprise-margin lever that I have promoted into the AI-spend frame because the operator-side budget conversation cannot avoid it. When the May essay lands, the seams between the two frames will be visible. For today, the operator-actionable version is this one.

If you read the January essay and are wondering whether to redo your 2026 AI budget against the new labels, the honest answer is yes. CAC and shrink are still real measurements; they are no longer the levers. Labor, throughput, accuracy, guest data — those are the levers the February print cycle made unmistakable. The vocabulary is the work. The contract follows the vocabulary. The cheque follows the contract.

— Eitan writes Mise. He founded TableTransfers after selling his last restaurant group. Tips: tips@tabletransfers.com.