QR Codes Didn't Kill Waiters. And Neither Will AI.

Automation usually moves backstage. The AI industry is starting to remember.

By late fall 2020, eight months into the pandemic, QR codes were everywhere in New York City. My friends and I would huddle outside in our warmest coats, precisely six feet from any other diners, and scan a square on our phones to pull up the menu. We’d order, we’d pay, and a masked back-of-house employee would slide the food onto our table without a word.

Susana M and me in our COVID dining era ca. November 2020.

I remember thinking that the hospitality industry was cooked (excuse the pun).

The technology had been around for decades (since 1994, to be exact) without ever catching fire. But this time, the speed at which restaurants integrated QR codes with their POS systems — turning every diner’s phone into the front-of-house — felt irreversible.

Given the circumstances, it was a pretty good user experience. Or, at least, it was efficient. You could browse a complete, regularly updated menu with photos and order and pay without flagging anyone down.

Six years later, the waiters are back.

Full-service restaurant employment is still 3.7% below 2019 levels — but the segment added 86,000 jobs in the last twelve months alone. Tens of thousands of customer-facing QR codes have been retired without fanfare.

What didn’t get retired was the POS systems they ran on, the kitchen display software, the reservation flow, the back-end infrastructure that QR codes accelerated. That part of the transformation stuck.

The diner-facing version of the QR code lost to human touch while the plumbing it ushered in won.

The lesson is not that automation failed; but that automation moved backstage, as it so often does.

I’ve been thinking about this a lot over the past week, just as the narrative around AI and the job market seems to be pivoting — hard.

For the last few years, the dominant version of the AI story has been: we are building machines that can do your job. Sometimes that was a warning and sometimes it was a sales pitch and most often it was both.

Dario Amodei warned of a white-collar bloodbath. And virtually every dinner party conversation in San Francisco invoked some version of universal basic income or the permanent underclass. Then last week, as Noah Smith documented in AI’s Big Messaging Pivot, Sam Altman called AI CEOs predicting mass job loss “tone-deaf” and announced that OpenAI is in the business of “augmenting and elevating people, not entities to replace them.”

Some of this is undoubtedly PR. 50% of Americans are now more concerned than excited about AI. The local campaigns to block new data centers are heating up. The regulatory pitchforks are coming out and the CEOs are putting on their cardigans and acting all neighborly.

Derek Thompson framed the deeper question well this week: is AI a normal technology in the lineage of electricity and the internet, or something genuinely new? The bull case for "new" runs through Anthropic's Jack Clark, who puts 60% odds on recursive self-improvement by 2028 — the moment AI starts building better AI on its own and the historical analogies stop applying. Maybe. But focusing on capability is the wrong frame. The waiter didn't beat the QR code on capability. The QR code was better at the measurable part — faster, more accurate, infinitely scalable, doesn't get a cold or a bad mood. The waiter won because the waiter was not just an input to the product. The waiter was part of the product.

Some categories of work are naturally like that: hospitality, therapy, teaching, hairdressing, live music. The arts, obviously, but also a long list of jobs people lazily assume are about information transfer when they’re actually about a person caring about another person in real time. AI could become far superior at the underlying task and still not settle the question, because the task is not always what’s being bought.

That said, the human-touch thesis isn't as universal as its boosters make it sound. People pay extra to take Waymo over Uber specifically to avoid the small talk. ATMs replaced bank tellers. And the QR code didn't lose everywhere — Sweetgreen and Chipotle still run on apps; Starbucks runs on mobile order. The line is more specific than "humans win because they’re nicer." Where the person is just the interface, the interface gets automated. Where the person is part of the product, the person sticks around.

Customer service is the obvious example. Companies aggressively automated CX and the technology works in the narrow sense — chatbots resolve tickets and generate sympathetic apologies at near-zero marginal cost. And yet customer service in 2026 is uniformly worse than it was in 2022. I have spent more time this year begging chatbots to let me speak to a human than I ever spent waiting on hold.

Princeton’s Arvind Narayanan and Sayash Kapoor named a related problem in AI as Normal Technology: “the easier a task is to measure via benchmarks, the less likely it is to represent the kind of complex, contextual work that defines professional practice.” Put simply: benchmarks measure tasks, but tasks are not jobs.

The bull case is that this is just early — that AI customer service in 2026 is what self-driving was in 2018, and the chatbots will be indistinguishable from a great agent in five years. Maybe. Chatbots can already apologize, summarize policy, and route tickets. They cannot empathize or take responsibility. When you’re calling about a lost bag, you’re not only calling about the bag.

AI will absolutely cannibalize tasks, just like QR codes did.

But the mistake is assuming tasks cleanly translate to jobs. They don’t. A job is a bundle of tasks, trust, taste (sorry), accountability, and care.

Sometimes the task is the thing being bought. But often, it’s just the excuse for everything else.