Sweetgreen Promised 100% Robot Kitchens. Quietly, They Settled on 50%. Here's Why That Number Matters More Than the Original One.

THE CRAFT

In late 2023, Jonathan Neman — co-founder and CEO of Sweetgreen — stood on an earnings call and told investors that within five years every Sweetgreen store would be running a fully automated “Infinite Kitchen,” the assembly-line robot the company had acquired from a startup called Spyce in 2021. The vision was complete automation. One hundred percent. He said the number out loud. The stock went up.

In September 2024, CFO Mitch Reback said, quietly, that the new target was about half of stores. Fifty percent. Not “most.” Not “the majority.” Half. By late 2025, the company was publicly framing 75% of new openings as the floor for Infinite Kitchen deployment — a number that sounds bigger until you notice it’s only new openings, not the installed base, and the retrofit pace for existing stores was weighted heavily to the back half of 2026. By early 2026 the company had opened its first drive-thru location with an automated makeline — a milestone — while simultaneously being on record that the “every store fully automated” plan was no longer the plan.

This is a walk-back. It is also the most useful piece of data any founder-CEO of a $100M–$500M operator is going to get this year about where AI-plus-robotics actually earns its keep and where it doesn’t. I want to walk you through it, because the tech press covered the original 100% promise as a big story and is going to under-cover the walk-back. The walk-back is where the real lesson lives.

THE OPERATOR

The company and the founder

Jonathan Neman co-founded Sweetgreen in 2007 with two Georgetown classmates — Nicolas Jammet and Nathaniel Ru — as a single salad shop on M Street in Washington, D.C. Eighteen years later the company is a publicly traded (NYSE: SG) fast-casual chain with roughly 250 stores, about $676 million in FY2024 revenue, and a publicly stated goal of 1,000 stores by 2030. Neman is still CEO. He is, by the definition of this newsletter, exactly the operator we write for: a founder who still runs the company he started, operating at a scale where every capital allocation decision he makes is a bet the rest of the industry is going to study.

In 2021, Sweetgreen acquired Spyce — a Boston-area startup founded by MIT engineers that had built a physical assembly-line system for making bowls. A robotic arm portions ingredients from rotating wells into a bowl moving down a belt. Human workers prep ingredients upstream and hand the finished bowl to the customer downstream. The middle section — the actual assembly — is the part the robot does. Sweetgreen rebranded the system as the “Infinite Kitchen” and opened its first retrofit store in Naperville, Illinois in 2023. A second in Penn Plaza, New York. Then more.

The original promise and the quiet revision

Here is the actual arc, in public statements:

• Late 2023: Neman tells investors Sweetgreen expects every store to be fully automated within five years.
• May 2024: Fortune profiles the “banner earnings” Sweetgreen is posting behind its Infinite Kitchen stores, citing materially higher throughput, better labor margin, and a cleaner guest experience. The story is unambiguously positive.
• September 2024: CFO Mitch Reback, on an investor call, reframes the target to roughly 50% of stores. The quote: automation is no longer “the plan for every store” — it’s the plan for the stores where the format and geography justify the capital. Public coverage of the revision is thin.
• Late 2025 into 2026: Sweetgreen publicly commits to 75% of new openings including Infinite Kitchens — a number the tech press will read as a continued push toward automation, and a number the operations team will read as a careful segmentation of where the robot actually earns its keep.
• 2026 earnings commentary: Neman expects the pace of retrofits of existing stores to accelerate through the back half of 2026 and into 2027, with the first drive-thru format featuring an Infinite Kitchen going live in California.

Two things are true at the same time. Sweetgreen is continuing to deploy the robot in most new stores. And Sweetgreen is no longer promising to put the robot in every store. The ceiling came down from 100% to 50% of the total fleet within twelve months of the original promise, and the company did not send out a press release to announce the change.

What changed underneath

The Infinite Kitchen produces real operational wins in the stores where it’s installed, and Sweetgreen has been relatively clear about what those wins are: higher throughput at peak (more bowls per minute), more consistent portioning (better food cost discipline), better labor margin (the same store runs with fewer workers per shift), and a measurable improvement in guest-facing wait times. If you are running a high-volume urban lunch store with a predictable menu mix, those wins compound into genuinely better store-level economics.

But the wins are not evenly distributed. In a suburban drive-thru with a different traffic curve, different menu mix, and different labor cost structure, the robot produces a much smaller margin improvement — and the installed capital, which is not trivial (retrofit costs for an Infinite Kitchen are in the high six figures per store, and new-build costs are materially higher than a traditional kitchen), takes much longer to pay back. In some store formats, the math for full automation simply does not work once you run the detailed unit economics. The ceiling of “every store automated” runs into the floor of “this store’s labor was never the problem.”

This is the insight: Sweetgreen did not walk back the technology. Sweetgreen walked back the belief that the technology would pay off equally everywhere. And that distinction is what separates an operator from a vendor. Operators learn where the tool works. Vendors stay on the original promise for another three quarters because the original promise is what sells the next contract. Neman, to his credit, let his team publicly revise the target once the unit-economic data came in, rather than defending the 100% number for its narrative value.

The Craft of AI read

Three things to take from this, and one explicit warning.

Lesson one: the first version of your AI/automation roadmap is wrong, and the sooner you let it be wrong, the better the second version will be. Every founder-CEO listening to a vendor pitch in 2026 is being told a percentage: “90% of your customer service calls can be automated,” “75% of your content can be generated,” “half of your back office is replaceable.” These numbers are directionally useful and specifically wrong. The actual percentage is always lower, and the interesting operational question is where the percentage lands and why — not what the topline number is. Sweetgreen’s walk-back from 100% to 50% is not a failure. It’s a discovery. The operator who pretends the original number was right is the operator who loses money in the last 20% of stores (or workflows, or call types) where the tool never should have been installed in the first place.

Lesson two: segmentation is the entire game in a mixed-format operator. Sweetgreen has traditional urban lunch stores, suburban stores, drive-thrus, airport kiosks, and a growing catering channel. Each is a different unit economic. The Infinite Kitchen works best in the highest-volume, most predictable-mix formats. If you are running a mid-market business with more than one “format” — whether that means store types, product lines, customer segments, service tiers, or geographic markets — your AI-plus-automation strategy has to be segmented by format from day one. A single percentage applied across the company is a hand grenade. The operator who segments first and deploys second beats the operator who deploys first and segments second.

Lesson three: the walk-back is the moment you learn who the real operators are. Watch how Neman handled the revision. No press release. No big acknowledgement. A CFO comment on an investor call that reframes the target. The vendors in the space continue to market the technology at 100% deployment; the operator runs it at 50%. This is not cynicism — this is operating discipline. The narrative hasn’t caught up with the data yet, and Neman is neither trying to correct the narrative publicly nor trying to sustain the original number for its PR value. He is just running the business against the actual math. When you are in the walk-back moment on your own AI project — and you will be — this is the posture to model. Revise the target quietly. Don’t call a meeting to announce the revision. Let the unit economics become the strategy.

Warning: do not mistake the walk-back for a retreat. Sweetgreen is not “giving up on automation.” Sweetgreen is deploying automation more aggressively in 2026 than in 2024 — the difference is that the aggression is now targeted. 75% of new builds. Retrofit acceleration in the high-volume stores. The first drive-thru with automation. What changed is not the direction; it’s the precision. Your board is going to hear “Sweetgreen walked back from 100%” and read it as “automation isn’t working.” That read is wrong, and you need to be ready to correct it in the room. The technology is working. The all-stores version of the vision is what isn’t working. Those are different sentences and the distinction matters.

Things to consider

1. Segment your operation by format before you deploy any AI tool at scale. If you have different store types, product lines, customer tiers, or geographic markets, the AI/automation case is going to be different in each. Start with the one where the math is most obviously positive. Do not roll anything out to the whole company until you have proven the unit economics in the best segment first.
2. Write down the “walk-back number” before you start. At what installed percentage of your operation would the ROI start to underperform? 50%? 70%? 80%? If you do not have an honest answer, you are not going to recognize the walk-back moment when it arrives. Define it now, while the original enthusiasm is still in the room.
3. Identify the worst-fit segment for the tool you’re considering, and deliberately do not deploy it there first. Most operators make the opposite mistake: they deploy the tool in the segment that needs the most help, under the theory that the upside is biggest there. It is not. The upside is biggest where the conditions match the tool’s strengths. Deploy into the easy wins first, learn, then consider the hard ones.
4. Who on your team is authorized to walk a commitment back in public? Neman’s walk-back went through his CFO on an earnings call. The CFO had the institutional standing to revise the number without making it look like a crisis. Who has that role in your company? If the only person who can walk back a commitment is you, you will walk it back too late or not at all, because walking back your own commitments is emotionally hard. Build in the distance now.
5. Read the Sweetgreen walk-back language carefully. The framing is not “automation didn’t work.” The framing is “automation works in the formats where the economics support it.” That is the exact sentence your board needs to hear when your first AI deployment comes up short of its original promise. Borrow the language. It works.

THE WORKBENCH

Here’s the tactical takeaway for your Monday.

Build a one-page segment matrix for the AI/automation tools you are currently evaluating. Rows: your operational segments (store types, product lines, customer tiers, geographies — whatever the natural cut is in your business). Columns: the tools under consideration. Cells: the honest number for what the ROI looks like in that specific segment with that specific tool.

Most operators doing this exercise for the first time are going to find three things. First, the ROI is genuinely positive in a smaller number of cells than the vendor’s pitch implied — usually the two or three most favorable segments. Second, the ROI is neutral or negative in more cells than the vendor’s pitch implied — usually the segments the vendor hopes you’ll deploy into anyway, because every deployed instance is revenue. Third, the cells where the ROI is biggest are rarely the cells where the pain is loudest. The loudest pain usually lives in the worst-fit segment, and the vendor’s pitch deck will point there first.

Print the matrix. Tape it to the wall next to your desk. Before you sign any contract, look at the matrix and ask yourself one question: if I deploy this tool only in the cells where the ROI is positive, does it still make sense to buy it? If yes, you have a deal. If no, you were about to buy a company-wide tool to solve a segment-specific problem, and you should negotiate a segment-specific deployment with pricing to match.

Sweetgreen’s board probably has a version of this matrix on a wall somewhere. They almost certainly have it now. The 50% number is the matrix.

THE QUESTION

Here’s what I want to know from you this week.

When was the last time you let a public commitment your company made about AI, automation, or technology be quietly revised based on real data — and what was the walk-back number?

I am not looking for war stories about projects that failed. I am looking for the moment your team looked at the unit economics and said “the original target was too ambitious, and here’s what we actually think the number should be,” and then someone made the decision to revise the number rather than defend the original. That is a very specific moment. It is a sign of operating maturity, and I suspect it is more common at the $100M–$500M founder-led companies in this audience than at the big enterprise tech companies that get written about for their AI bets. Prove me right, or prove me wrong.

Hit reply, or send a note to grant@thecraftofai.com. Names and details will stay private. I read every reply.

— Grant grant@thecraftofai.com