Everyone’s spending a fortune making AI smarter. Hospitality and retail solved the part that actually matters, decades ago.
The biggest tech companies on earth will spend close to seven hundred billion dollars this year building artificial intelligence. Microsoft alone is in for about a hundred and ninety billion in infrastructure. Even after spending that, it’s hedging on the intelligence itself. Its own product, Copilot, quietly swaps between ChatGPT, Claude, Gemini, and Microsoft’s homegrown AI, picking whichever one fits the moment.
Microsoft then told its customers not to get attached to any of them.
That means the company pouring the most money into the intelligence itself is betting you won’t care which one answers you. It’s a good bet since you probably can’t tell which one wrote the email you read this morning.
So, if the intelligence itself is going to be a commodity, there needs to be a better differentiator. If what’s under the hood is roughly identical across ChatGPT, Claude, and Gemini, we need to define what actually separates one AI product from another.
The answer brings us back to classic, offline strategy, the kind hospitality and retail have run for a century. Think about the thing a great restaurant has always had over a good one. Every serious kitchen has a good oven, for example, but what separates them is everything that happens around the food. Whether the server can read the table, or they catch the mistake before you do, or whether you leave feeling like someone was on your side.
Hospitality and retail spent a century learning to make a person feel understood. We’re spending seven hundred billion dollars building AI that still hasn’t.
“I just know”
Years ago, I’m at the first bar I owned, an East Village neighborhood joint called Destination. I’m sitting at the bar next to one of my business partners, Mason Reese. He’s watching the bartender build an apple martini, and he turns to me, satisfied, and says, “Now that’s how you make an apple martini.”
I blink at him. Mason doesn’t drink. Has never had one in his life. “How can you tell?”
He looks at the glass, this sweet antifreeze-green concoction, and says, “Look at the color. That’s just what they’re looking for.”
I’d made a lot of apple martinis behind that bar. “Apple pucker and some water would look exactly like that. There’s no way you can tell a good one from the color.”
He looks at the glass again. “I just know.”
Thirty seconds later the drink came back. The customer said it was oversweet and undrinkable.
I think about that “I just know” a lot now, because it’s the exact sound an AI model makes. Exuding total confidence about something it has no actual way of knowing.
To be clear, Mason wasn’t lying. He genuinely believed he could read the drink. He’d just confused the way something looks with whether it’s any good, and there was no tell, no crinkle of the mouth, until the thing came back to the bar.
Bartending taught me the first half of a very simple lesson long before any AI did. It’s one of the more reliable findings in the study of human judgment, our certainty and our accuracy are only loosely related, and machines inherited that flaw almost automatically, sounding just as sure when they’re wrong as when they’re right.
Confidence is not competence, and the person feeling confident is very often the last to find out.
You are not the customer
The deeper mistake underneath Mason’s martini is that he thought he was the customer. He wasn’t. He was a guy who doesn’t drink, deciding he knew what a drinker wanted.
At Charming Robot, the design studio I’ve run for fifteen years, we put a slide in front of every new client that says, “We may use the product, but we are not the customer.” We do it because the confusion is so easy and so constant.
When you’re solving a problem, you reach for your own habits, and your habits aren’t your customer’s habits. “Well, I’d do it this way.” “I’d never do that.” Good for you, but you’re just one person.
When we were building Econofact, a site meant to help regular people understand economic policy, I suggested that we knew most people don’t have time to read a whole article, so we’d give them the key points up top and let that at least give them the TL;DR (too long; didn’t read). The client said, oh, like those bullet summaries at the top of news articles. Yep, I said. And he said, “Oh, I hate those. I find them really distracting.”
This performance of total confidence with zero standing is the same move Mason made and it’s the most expensive mistake in this business, because nobody thinks they’re the one making it.
The AI industry is walking straight into it. The people building these tools are power users, fluent, comfortable, delighted by a clever answer, and quietly certain everyone else feels the same way. Raluca Budiu calls this the false-consensus effect, our built-in assumption that other people are basically us. The most confident people in AI right now are building for people exactly like themselves and calling it user research.
Anyone who knows how to create great service knows better. A great bartender doesn’t read you on the first drink. Walk into a good running store like JackRabbit and the person doesn’t hand you the perfect sneaker, they ask what you’re training for, watch you run, ask where it hurts, and come back with three boxes for you to try.
The expertise comes from the curiosity, questions, the watching, the second and third pass. Niels Van Quaquebeke and Will Felps call this “respectful inquiry,” asking open questions and then actually listening to the answer, and argued it makes the other person feel like a person, competent and known, instead of a problem being processed.
That’s what AI currently skips. It reads the words you typed and answers, once, instead of treating help as something you arrive at together. Nobody’s asking it to be psychic. They should be asking it to be curious.
Service is a monologue. Hospitality is a dialogue.
Danny Meyer, who built Union Square Cafe and Shake Shack and wrote the book most restaurant people quietly consider scripture, draws a line I can’t stop applying to software. Service, he says, is a monologue.
We decide our standards, we set them, we deliver them. Hospitality is a dialogue. It’s listening to the actual person and responding to what you heard.
Put that into the context of AI and it’s almost too on the nose. The AI is a monologue machine, and a magnificent one at that. It delivers its standard output, beautifully, to everyone, having read nothing about you except the words you just typed. It performs service at a very high level, but it does approximately zero hospitality. It can’t read the room, because it can’t see the room.
Will Guidara, who ran the dining room at Eleven Madison Park when it topped The World’s 50 Best Restaurants in 2017, wrote the book Unreasonable Hospitality, where he sharpens the same edge. Service is black and white, he says. Hospitality is the color. His whole philosophy comes down to the phrase, “one size fits one.”
This is the exact opposite of what the AI does. ChatGPT and Claude fit one size to everyone and sound thrilled about it.
You can watch this distinction play out, painfully, in The Bear. The whole show is the gap between technical excellence and actually making a person feel something. There’s an entire episode, “Forks,” where a guy who’s spent his life being competent gets sent to stage at a world-class restaurant and slowly understands that plating the food perfectly didn’t mean the restaurant had done its job.
The real job was the fork placed just so, the reservation remembered, the moment where a stranger feels seen.
For what it’s worth, Meyer doesn’t think this is a restaurant idea. He thinks we’ve moved from a service economy into what he calls a “hospitality economy,” where a superior product stops being a differentiator because someone can always match it, and how you make people feel becomes the only moat left.
He said that about restaurants and retail twenty years ago. Today that reads like a memo about AI.
Everyone will steal from you (and other trust lessons)
Reading the person is one half of hospitality. Trusting them is the other, and that one I learned the hard way. Let me tell you about the worst thing that happened in my bar’s first month.
Early on, we had a cocktail waitress that I’ll call Tara. Another one of our partners didn’t trust her. There was no evidence to support this, but he “just had a feeling.” So one Saturday while she was on a smoke break, he slipped an extra twenty into her register drawer, figuring she’d come up twenty over at the end of the night and pocket it.
She did. He had her fired.
You can claim that he “caught” her, but it was still a disaster. One of our first employees, Josh, still talks about it, how it curdled something in the whole staff, how after that, every shift, one of the partners had to count every penny of his drawer at close and wouldn’t let him touch it. Josh would just stand there thinking, “why am I even here?”
Instead of protecting the business, all that surveillance did was communicate to the staff that we saw them as thieves.
My partner ran it backwards. He demanded proof of trustworthiness first, went looking for the betrayal, and found it, which created surveillance that made things exponentially worse.
The idea that watching people impacts them is something we’ve known for a century. It’s called the Hawthorne effect, named after a 1920s factory study that found workers behaved differently the moment they knew they were being observed. The original claims turned out to be shaky. When Steven Levitt and John List dug up the data decades later, the dramatic version mostly fell apart. But the quieter truth held, and anyone who’s been watched at work has felt it. Josh felt it every night at that register.
More recent work is sharper about which direction the behavior goes. In a pair of studies, Chase Thiel and his colleagues found that employees who know they’re being watched are more likely to break the rules, to cheat, to steal, to slack. Being monitored made people feel less responsible for their own conduct, so they did things they’d otherwise think were wrong.
Watch people like criminals and you help produce them. People tend to become how you treat them.
In that situation, the bar staff is serving the customer, so the staff is the AI. The customer is the user. And the owner slipping the twenty into the drawer, that’s those of us building these products. We are the ones deciding how much to trust who (or what) we’ve hired to take care of everyone else.
Right now we’re all my partner with the twenty-dollar bill.
In other words, the companies building AI don’t trust it.
They have a hunch it’ll embarrass them, so they wrap it in restrictions, and the caution leaks into every ordinary interaction. Paul Röttger and his colleagues called this “exaggerated safety,” after finding that AI systems refuse plainly harmless requests when the words happen to resemble something dangerous.
They used the example of a system that won’t explain how to kill a running computer process, a completely ordinary programming task, because it saw the word “kill” and assumed the worst. Weirdly, even cooking questions spark the same problem. Ask how to beat egg whites or butterfly a chicken breast and instead of a recipe you could get a little flag of suspicion, as if you might be up to something.
This is the system protecting itself from you instead of helping you, the same suspicion we pointed at Josh when we made him prove his innocence at the register every single night, for a crime he never committed.
The AI doesn’t care about being careful with you, but rather it’s been programmed to be careful about itself.
It doesn’t stop there. Justin Cui and his colleagues tested thirty-two of these systems and found that the ones trained to refuse the most genuinely harmful requests were, again and again, the same ones that turned away harmless ones. Safety and over-refusal moved almost in lockstep.
When you tighten your fist out of mistrust, that reflex bleeds into everything, exactly like it did at the bar. The mistrust doesn’t focus itself to one incident or person, exactly like it didn’t at the bar. Build one part of the thing to cover itself and the whole thing starts hedging, and the person who came in for help gets Josh on a bad night, going through the motions, protecting himself, wondering why he’s even there.
Retail figured out the opposite move, which is that you earn trust by lowering the cost of extending it to you.
The word hospitality, it turns out, comes from the same root as hospital and host. It means the taking in of a stranger. The whole concept is built on extending trust before you’ve been given a reason to.
- Warby Parker mails you five pairs of glasses to try on at home, for free, because they understood the actual barrier was the risk of trusting them with a purchase you couldn’t see first.
- REI trains staff to talk you out of the expensive gear you don’t need, and the co-op will genuinely tell you the cheaper thing is fine, which is why you believe them when they say the expensive thing is worth it.
- Chewy sends handwritten cards and, when a customer’s pet dies, flowers. Nobody asked. That’s the point. They read the situation and moved before the request, which is the exact opposite of something that waits to be asked and then answers everyone the same way.
The road to success is paved with mistakes well handled
Extending trust up front is one move. Knowing what to do when that trust gets tested, when something actually goes wrong, is the harder one.
There’s a well-documented pattern in service research called the service recovery paradox. Coined in 1992 by Michael McCollough and Sundar Bharadwaj, it means that a customer whose problem gets handled well can walk away more loyal than one who never had a problem at all.
The mistake, acknowledged and resolved out in the open, becomes the moment trust is actually forged.
One important part to note is that this doesn’t always hold, and that’s the part that matters most for AI. Vincent Magnini and his colleagues found the paradox mostly shows up under specific conditions, such as when the failure wasn’t severe and the customer hadn’t been let down before.
A one-time stumble, handled with care, builds trust. The same mistake over and over destroys it faster than no recovery at all.
That’s the whole problem with a system that fails confidently, and often. A chatbot that’s confidently wrong gets abandoned.
Meyer put the healthy version plainly: the road to success is paved with mistakes well handled.
Ritz-Carlton built an entire legend around this. Every employee, housekeeper to front desk, is authorized to spend up to $2,000 per guest, per incident, with no need for approval. Just so they can fix a problem on the spot.
Everyone remembers the $2,000, but that’s not the point. What the rule actually does is erase the gap between an employee seeing a problem and being able to fix it, putting the authority where the information already is, with the person standing in front of the guest.
It’s allowing the person closest to the context to be trusted to act on it.
Watch what an AI product does when it’s wrong. It smooths it over and moves on, because rather than being built to be on your side, it was actually built to keep you engaged. Tristan Harris, who spent a decade warning about this, calls it a race to the bottom of the brain stem. The smartest engineers alive pointing supercomputers at your weaknesses to figure out how to keep you clicking.
When that instinct hardens into craft, it has a name that Harry Brignull coined in 2010 as “dark patterns,” which are interfaces built with a real understanding of human psychology and aimed the wrong way, to trick you into doing what the company wants instead of what you came to do.
This is, of course, the exact inverse of hospitality. It gives the same deep attention to the person, but focused on taking from them instead of taking care of them.
A place built to take care of you handles a mistake differently.
The bartender who comps your ruined drink and remembers it next time is betting on the long game. The restaurant absorbs the cost of that drink because it knows a comped drink tonight can buy a regular for years.
An AI product could do the same thing, and it would barely have to try. It could tell you it’s about to guess before it guesses. It could say “I got that wrong” the moment it knows, instead of when you catch it, and not make you paste the whole problem back in to fix it, since it was the one who broke it. It could refund the wasted time the way a bar eats the cost of a bad pour.
Honestly, none of that requires a smarter engine, but it does require deciding the person is worth more than a transaction. It costs a little more than shoving out a confident answer and moving you along, but it builds the relationship and sets you apart (for now).
The day these systems are built to make that trade, to spend a little now because a person who trusts you is worth more than one you gave an answer to without caring whether it was right. That’s when they will have earned the word hospitality.
Build the dining room
There is no soft conclusion here. That’s the whole point.
ChatGPT, Claude, and Gemini are about to be interchangeable. They all currently feel like similar experiences. The UX field is already saying this out loud. Kate Moran and her colleagues put trust at the top of their list of design problems for AI in 2026, and argued that the interface itself is quietly ceasing to be a differentiator, because anyone can generate a decent one now, so the value moves to the harder, deeper parts of the experience.
Eric Karofsky says it plainer, that the edge won’t belong to whoever has the biggest model, it’ll belong to whoever earns the most trust, because AI’s real test was never intelligence.
Amelia Wattenberger keeps pointing out that everyone’s obsessing over the AI itself and almost nobody’s focusing on the experience around it.
The service-quality researchers had the vocabulary for this decades ago. When A. Parasuraman, Valarie Zeithaml, and Leonard Berry built SERVQUAL in 1988, still the standard way to measure service quality, they broke it into five dimensions. Two of them are assurance, the ability to inspire trust and confidence, and empathy, individualized attention to the actual person.
Trust and reading the individual aren’t nice-to-haves in that framework. They’re two-fifths of the whole definition of quality.
They’re exactly what a confident AI, answering everyone the same way, skips.
The retailers who win figured this out ages ago.
- Apple sells you a phone and then hands you the Genius Bar, a human trust-and-repair layer bolted right onto the product, so the relationship doesn’t end at the sale.
- Lululemon calls its floor staff educators and trains them beyond just closing a transaction, but rather to build a relationship.
- Trader Joe’s just took the number one spot in the country for supermarket customer satisfaction, and it didn’t do it with lower prices alone, but rather with warmth and empowered staff.
None of these companies has a better oven than their competitors. They built a better dining room.
Patrick Neeman argues that the AI failures everyone chases, the hallucinations, the wrong answers, trace back to invisible groundwork nobody wanted to fund. Hospitality is exactly that kind of invisible groundwork. Nobody claps for it and, yet, it’s the whole reason you go back.
Guidara has a line about how hospitality isn’t actually a restaurant thing. Most of the economy is service work now, he says, and anyone can choose to be in the hospitality business simply by deciding to care as much about how they make people feel as about the thing they’re selling.
Seven hundred billion dollars can’t buy that. You can spend your way to a smarter AI, but you cannot spend your way to one that gives a damn about the person using it. That part is a real choice.
It’s not even a mystery what the choice looks like, because we’ve all been on the receiving end of it when interacting with an actual human.
It’s the AI that asks a question before it assumes it’s you. It’s the one that says how sure it is, and means it. It’s the one that gets to know you across a conversation instead of meeting you fresh and slightly confused every single time, like Drew Barrymore in 50 First Dates. It’s the one that admits the miss out loud and fixes it without being chased. It’s the one that hands you to a human the second it’s out of its depth, cleanly, instead of confidently walking you off a cliff.
None of this is an intelligence problem. It comes down to hospitality, and hospitality is a choice made by someone who decides a person is worth it.
Microsoft is betting you won’t care which AI answers you, but that’s not a bet I’d necessarily take. Scott Magids and his colleagues found that a customer who feels emotionally connected to a brand is worth somewhere between 25 and 100 percent more than one who’s merely satisfied.
Satisfied is what you are when the answer was fine and connected is what you are when someone takes care of you. Those are different feelings and only one of them creates loyalty.
Right now, the intelligence is already good enough, but what nobody has built yet is the AI you’d actually come back to because the last time you used it, it felt like something was finally on your side.
Whoever figures that out will have earned the one thing seven hundred billion dollars can’t buy. They’ll have your trust, given freely, because for once something gave you a reason to.
References and further reading
On hospitality and service as craft:
- Danny Meyer, Setting the Table, on “service is a monologue, hospitality is a dialogue,” the hospitality economy, and hospitality meaning the other person is on your side (via The Commonwealth Club / KQED)
- Will Guidara, Unreasonable Hospitality, on “service is black and white, hospitality is color” and “one size fits one”; ran the dining room at Eleven Madison Park when it topped The World’s 50 Best Restaurants in 2017
- The Bear, “Forks” (FX), on the difference between technical competence and making a person feel seen
On trust, surveillance, and recovery:
- Chase Thiel and colleagues, “Monitoring Employees Makes Them More Likely to Break Rules” (HBR), monitored employees break more rules, not fewer
- Paul Röttger and colleagues, “XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models” (NAACL 2024), AI refuses plainly harmless requests when the wording resembles something dangerous
- Justin Cui and colleagues, “OR-Bench: An Over-Refusal Benchmark for Large Language Models” (ICML 2025), across thirty-two systems, the ones that refuse the most harmful requests also refuse the most harmless ones
- Vincent Magnini and colleagues, on the paradox’s limits, it holds only when failures are occasional and not severe
- The Ritz-Carlton $2,000 rule, via Micah Solomon, Exceptional Service, Exceptional Profit (Forbes)
- Harry Brignull, who coined “dark patterns” in 2010, interfaces built to trick users against their own interest
- Tristan Harris, on the extractive attention economy and the “race to the bottom of the brain stem” (Senate Commerce testimony)
On confidence, judgment, and getting it wrong:
- Our certainty and our accuracy are only loosely related (Psychology Today)
- Miao Xiong and colleagues, “Can LLMs Express Their Uncertainty?”, ask an AI how sure it is and it comes back overconfident
- Raluca Budiu at the Nielsen Norman Group, on the false-consensus effect, assuming other people are basically us
On experience, not intelligence, as the differentiator:
- Kate Moran and colleagues, “State of UX 2026” (NN/g), trust is a top design problem for AI and the interface is no longer the differentiator
- Eric Karofsky, “2026: The Year User Experience Finally Rewrites the Rules of AI” (CMSWire), the edge belongs to whoever earns trust, not whoever has the biggest model
- Amelia Wattenberger, “Why Chatbots Are Not the Future of Interfaces,” on how underbuilt the experience layer still is
- Scott Magids, Alan Zorfas, and Daniel Leemon, “The New Science of Customer Emotions” (HBR, 2015), emotionally connected customers are 25 to 100 percent more valuable than merely satisfied ones
- Niels Van Quaquebeke and Will Felps, “Respectful Inquiry: A Motivational Account of Leading Through Asking Questions and Listening” (Academy of Management Review, 2018), asking open questions and truly listening makes people feel known, not processed
- A. Parasuraman, Valarie Zeithaml, and Leonard Berry, “SERVQUAL” (Journal of Retailing, 1988), assurance and empathy are two of the five dimensions of service quality
- Patrick Neeman, “Information architecture is the foundation AI is starving for” (UX Collective), AI’s visible failures trace back to invisible groundwork nobody funded
On the retailers who lead on experience:
AI has a hospitality problem money can’t fix was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.