Americans did not stop eating out. US restaurant sales went from $580B in 2010 to a projected $1.5T in 2025. Fine dining grew. Fast food grew. What died was the middle — the places that sold you convenience at a medium price.

This is the pattern. And it is about to happen to your job.

What is convenience work?

A job where a customer pays a professional to do something they could do themselves but don’t want to: book a flight, develop a roll of film, do their taxes, draft a basic contract, buy stocks, run a small ad.

For a hundred years, a huge slice of the middle class did this kind of work. Then technology arrived and unbundled it.

Travel agents were first. Then one-hour photo shops, retail stockbrokers, video rental clerks, tax preparers. Now it is the turn of consultants, marketers, lawyers, and developers.

The interesting question is not whether it is happening — the data says it is. The interesting question is what the restaurant industry can tell us about how it ends.

The restaurant playbook

Here is what happened to American food over the last fifteen years.

The middle died. TGI Fridays, Applebee’s, Red Lobster — all compressed or bankrupt. Corporate cafeterias never came back after the pandemic.

The top got richer. US fine dining revenue went from $15B in 2019 to a projected $17B in 2025. Michelin three-star restaurants grew from 100 in 2010 to 157 today. A tasting menu at Atomix in New York cost $175 in 2018. It now costs $395.

The top gets weirder, not cheaper

New York now has about 140 omakase restaurants. The most exclusive charge $10,000 for a membership — just for the right to book a seat.

The bottom exploded. Roughly three-quarters of US restaurant traffic is now takeout, drive-thru, or delivery. DoorDash owns two-thirds of the US delivery market. Chipotle’s digital sales went from under 10% in 2018 to 35% today. Starbucks grew from 11,000 US stores in 2010 to nearly 17,000 now.

Two things about this pattern matter.

First, the category as a whole grew. Restaurants did not shrink — they bifurcated. More money flowed in. It just stopped flowing to the middle.

Second, what survived at the top was not better food. Much of what Michelin chefs do is within reach of a skilled home cook with a sous-vide and four hours. What survived was the thing you cannot assemble at home.

What fine dining actually sells

Not the food. The room, the staff, the ritual, the story.

Experience that cannot be downloaded.

Keep those two facts in mind. They are the whole essay.

Consulting: firing juniors, hiring AI experts

Now look at consulting, where the same thing is happening live.

Accenture did about $70B in revenue last year. It also laid off more than 11,000 people between May and August 2025.

“We are parting ways with people who cannot be reskilled for AI.”

— Julie Sweet, CEO of Accenture, to staff in 2025

At the same time, the firm’s AI-and-data staff grew from 40,000 to 77,000 in two years, and its AI revenue tripled. The firm is hiring at one end and firing at the other. The people being let go are not the people being hired.

McKinsey tells the same story in different numbers: about 10% of partners gone, around 25% of junior staff gone, 13% of back-office. Meanwhile its in-house AI group tripled in size. BCG’s AI practice grew from nothing to 20% of revenue in two years and is projected to hit 40% by 2026.

The generic work — the template deck, the routine research, the cookie-cutter analysis — is what is being automated. The people who did that work for a living are the ones packing their desks.

The numbers, rounded

• Accenture: 11,000+ layoffs in four months; AI headcount nearly doubled
• McKinsey: 25% cut to junior ranks over 18 months
• BCG: AI consulting grew from nothing to 20% of revenue in two years
• Indian IT (TCS, Infosys, Wipro combined): roughly 42,000 jobs gone in two years And yet — here is the awkward part — most of this AI investment isn’t paying off yet either.

“It used to be that McKinsey or Ernst & Young, they knew everything. They don’t know anything anymore. Nobody has a playbook.”

— Ethan Mollick, Wharton, 2025

A recent Gartner survey found 72% of CIOs are breaking even or losing money on AI. A BCG report found only 5% of firms are capturing real value from it; 60% get nothing. The middle is being cut in the name of a transformation that, so far, most companies cannot execute. That will not save the jobs.

Marketing: the agency collapse

Advertising agencies are the most public version of this, because they are big enough to show up in the stock market.

WPP, one of the largest ad-holding companies in the world, cut 9,000 jobs in 2025 — nearly 10% of its workforce. Its stock dropped 60% in a single year. Its CEO was replaced.

A symbolic moment

For the first time in decades, WPP is worth less than its French rival Publicis. Ten years ago that would have been unthinkable.

Omnicom and IPG, two other giants, merged in November 2025. The publicly stated target: $1.5B in cost savings. About $1B of that will come from more layoffs.

What is replacing them? Partly the clients themselves. A survey of 402 marketing chiefs found that 22% had already cut back on outside agencies because of generative AI tools. Another survey put the number at 60%.

The canonical case — and the cautionary one — is Klarna, the Swedish fintech. In early 2024, it reported AI had cut its marketing costs by $10M a year and reduced external agency spend by 25%. Production cycles went from six weeks to seven days. It became the most-cited AI-replaces-agency story in the industry.

Then the other shoe dropped.

“Our AI replacement of customer service led to lower quality.”

— Klarna CEO, May 2025, announcing the rehire of ~700 support staff

Both halves of that story are true. AI replaced creative work well. It replaced customer service badly. Which means the collapse is not uniform. Some corners of knowledge work are closer to the one-hour photo shop (extinct at the middle). Others are closer to the corner hardware store (still there, still useful). Nobody knows the exact map yet.

Law: the top absorbs the tool

Law is the clearest case of the opposite pattern — where the top tier captures the new technology instead of being disrupted by it.

Meet Harvey. It is a software company that helps lawyers write contracts, run due diligence, and file briefs — exactly the work that used to be done by junior associates.

Harvey, by the numbers

• End of 2024: $50M in annual revenue
• January 2026: $190M
• March 2026: valued at $11B
• Customers: 1,300 law firms in 60 countries, including most of the top 100 US firms Harvey should, in theory, be hollowing out those big firms from the inside. It is not.

Profits per lawyer at the top 100 US law firms are up more than 53% since 2019. Tech spending is up 10% a year. Salaries are rising.

Here is why. In a billion-dollar merger, the other side still wants a brand name on the contract. They are not buying the document. They are buying accountability, reputation, and the insurance that if something goes wrong, there is someone to sue.

What you actually buy when you hire a top law firm

Not the legal work. The name on the door.

A 200-year-old licensed firm with a reputation to lose. AI cannot provide that. A mid-tier firm barely can.

So the big firms bought the AI tools, kept the clients, and now share in the software revenue too. The middle of the legal market — the mid-sized firm that made its money on routine contract review and standard incorporations — is the one being squeezed. Basic legal work is moving to end-customer tools like LegalZoom. Bespoke work stays with the top.

This is why lawyers are a regulated profession and ad agencies are not. A licence is a moat. Without it, there is no floor.

Developers: the contested case

The loudest claim in this whole debate is that AI makes software developers dramatically more productive. The evidence here is messier than the headlines suggest.

The most-cited study is a 2022 GitHub experiment. 95 developers were asked to build a simple web server. Those using GitHub’s Copilot tool (an AI coding assistant) finished 56% faster than those without it. Real result. Real effect.

But in July 2025, a different study tried something harder. Sixteen experienced open-source programmers worked on real tasks in their own complex codebases, with and without AI tools.

The surprise result

The programmers expected to be 24% faster.

They felt 20% faster.

They were actually 19% slower.

Both studies are real. They measured different things. The first measured new developers on fresh, small tasks. The second measured experts on messy, mature code. The truth about AI productivity sits somewhere between them, and nobody has pinned it down.

Meanwhile, the money is not waiting. Cursor, an AI coding tool, went from $1M in annual revenue to $2B in two years — the fastest B2B software growth ever recorded. GitHub Copilot has 20 million users.

And the benchmarks? Here is where things get awkward.

The benchmark trick

There is a standard test called SWE-bench that measures how well AI can solve real software bugs. Top models now score 87–94% on it. Impressive.

In September 2025, a harder version — SWE-bench Pro — was released, with tasks the AI had never seen before.

The same model that scored 87% on the original scored 46% on the new one.

The headline numbers overstate the capability. The real number is closer to “about half.”

The pattern everywhere

Pull back, and this has happened before. Many times.

The timing is remarkably stable.

The 13–18 year rule

From the moment a new technology arrives to the moment the middle of an industry collapses: about 13 to 18 years.

• Digital cameras: 1994 → one-hour photo gone by 2013
• Napster: 1999 → streaming won by 2015
• Expedia: 1996 → Blockbuster bankrupt by 2010 One detail matters. Where there is no licence required — photo, video, travel — the middle collapses 85–95%. Where there is a licence — law, accounting, real estate — the middle only compresses 10–30%.

Where do agencies and consultancies sit?

No licence. Just reputation, scale, and client relationships.

Lighter moats. Expect the harder version of the pattern.

But are the robots actually good enough?

The strongest argument against all of this is simple: the AI tools are not yet good enough to replace the work.

The most honest measurement comes from a 2024 Carnegie Mellon test called TheAgentCompany. It gave AI agents real office tasks — the kind a junior employee would get. The best AI finished about one in three of them. By mid-2025 the number climbed to roughly one in three fully and two in five partially.

So: can an AI replace a junior office worker today? No. Not at half the tasks. Not even close.

But here is the trap in that argument.

The threshold is not 100%

The middle does not collapse because AI does 100% of the work.

It collapses because AI does enough of the work that the old cost structure stops making sense.

If a law firm used to need ten junior associates to review contracts, and AI now does 40% of that work, the firm needs six, not ten. Four jobs disappear. The firm survives. The four junior associates do not.

That is the threshold. Not “AI can do the job.” Just “AI can do enough of the job that the math changes.” That threshold has already been crossed in consulting, in agency creative, in routine legal work, and in small-business advertising.

What this means for you

If you are a knowledge worker — a consultant, a designer, a marketer, a developer, a lawyer, an accountant, an analyst — the question to ask yourself on Monday morning is not whether your industry will be disrupted. It will.

The question is narrower and more uncomfortable:

The Monday-morning question

Which part of what you do could your customer now buy directly, without you?

If the honest answer is most of it, you are in the middle of the restaurant, and the middle is where the pattern hits first.

If the honest answer is the parts that pay my rent are the parts nobody can buy without me — the judgement, the taste, the relationship, the accountability, the thing that would make your client pick up the phone at midnight — then you are probably at the top, where fine dining grew while the chains burned down.

What has survived, in every previous round of this pattern, is the thing the customer cannot assemble themselves.

Three examples of what survives

• For the lawyer: the name on the door when a merger breaks down.
• For the travel agent: the 2am call in Lisbon when the flight is cancelled.
• For the chef: the room, not the recipe. Your version of the name on the door, the 2am call, the room — find it. That is the only part of your job that was ever worth paying for anyway. The rest was convenience.

And convenience, as the restaurant business already learned, was never the product. It was the packaging. The packaging is what AI is about to eat.

What changed in this revision

Added 14 pull-quote / callout boxes throughout the text, so the reader gets visual rest between number-heavy paragraphs.

Why this helps

• Heavy-number paragraphs now have a visual anchor. A reader can skim the numbers and catch the punchline from the callout, or read the prose for depth.
• Four human quotes added (Sweet, Mollick, Klarna CEO) — the previous draft had none, which made the essay feel abstract. Quotes give the argument voice.
• The key arguments are now visible at a glance. Someone scrolling through can read only the pull-quotes and still get the full thesis.
• Word count roughly unchanged (~2,000 words), but perceived density is much lower because the page now breathes.