In 2016, Geoffrey Hinton stood in front of a room full of machine learning researchers and said something close to a curse on his own field. He told them to stop training radiologists. Within five years, he said, AI would be reading scans better than any human. The radiologists were toast.
Eight years later, the radiology job market is on fire. Salaries are up. There's a shortage. Residency programs are full and turning people away.
Hinton wasn't wrong about the technology. He was wrong about what part of the job the technology was going to eat.
AI did get extraordinarily good at reading scans. It now flags nodules, measures tumors, and pre-reads stacks before a human ever opens the file. But the radiologist still gets the call from the referring doctor. She still owns the read. She still signs her name to it. She still calls the patient when it's bad news. She still gets sued if it's wrong.
What AI took wasn't the radiologist. It was the part of the radiologist's day where she stared at a screen for nine hours and did the work the radiologist before her did in the dark.
And that pattern, almost embarrassingly, is showing up everywhere.
The Same Movie In Every Industry
A lawyer still wins the client over lunch. He still argues the motion in court when it matters. But AI did the discovery, drafted the contract, and pulled the case law. We know exactly what happens when somebody tries to skip the human bookend, by the way. A New York lawyer named Steven Schwartz famously filed a brief in 2023 that cited six court cases ChatGPT had made up entirely. The judge sanctioned him. The cases didn't exist. AI did the middle. He skipped the part where a human verifies it. The whole legal system noticed.
A CPA still has the relationship with the business owner. She still signs the return. But AI categorized the transactions, reconciled the books, and drafted the statements.
A doctor still owns the diagnosis and the relationship with the patient. But hospitals are now rolling out AI scribes from companies like Abridge and Suki that listen to the appointment, write the chart note, and draft the billing code in real time. Doctors talk about it like a religious experience. They got their evenings back. The middle of their job, the documentation grind, is gone.
A software engineer still architects the system, still owns the deploy decision, still debugs the thing nobody else can. But AI wrote half the code in between. Mark Zuckerberg said on Joe Rogan in early 2025 that AI would replace Meta's mid-level engineers this year. Marc Benioff at Salesforce said the company isn't hiring any new software engineers in 2025. They're not saying engineering is dead. They're saying the middle of engineering is dead.
A real estate agent still walks the buyer through the house and reads the room when it's offer time. But AI ran the comps, wrote the listing, and built the marketing plan.
AI isn't taking the front of the job. It isn't taking the back. It's eating the middle.
And the middle used to be most of the work.
Why The Middle Was Where The Money Was
For most knowledge work, the value chain has always looked roughly the same. Someone at the top wins the trust to do the work. Someone at the bottom owns the moment of accountability where a real human signs a real name to a real decision. Everything in between, the production layer, was where most of the hours got spent and where most of the salaries got paid.
A law firm wasn't expensive because of the rainmaking partner and the courtroom finisher. Those are the same two humans. It was expensive because of the floor full of associates doing the middle.
An ad agency wasn't priced on the creative director's taste or the account lead's relationship. It was priced on the producers, copywriters, analysts, and project managers running the middle.
Same for consulting, accounting, marketing, software, and a dozen other industries. The bookends were always human. The middle was the business model.
AI didn't come for the bookends. It came for the business model.
The Uncomfortable Part Nobody Wants To Say
Here's where this gets uncomfortable, and I haven't seen many people willing to say it out loud.
The middle is also where careers get built.
You become a senior lawyer by spending five years drowning in the middle. You become a great strategist by running a hundred bad analyses before you run a good one. You become a sharp marketer by writing 500 mediocre headlines before one of them actually lands. You become a senior engineer by debugging things you didn't fully understand for years until one day you did.
The middle wasn't just labor. It was the training ground.
If AI takes the middle, what happens to the pipeline? How does a 22-year-old become a 32-year-old who knows what they're doing? Look at the entry-level software engineering market right now. Computer science graduates from solid programs are sending out 300 applications and getting silence. Junior associate hiring at big law has stalled in a way the partners aren't quite ready to talk about publicly. The on-ramp narrowed and nobody's quite sure what replaced it.
Some industries will figure this out. Some won't. And the ones that don't are going to wake up in 15 years with no senior talent and absolutely no idea how that happened.
We think about this a lot at WE-DO, because we hire young people, and the work that used to be their on-ramp is the exact work AI now does in 30 seconds. You can't just let them sit and atrophy. You have to deliberately push them up the value chain faster than the industry used to, into the parts AI can't do. Judgment. Taste. Client trust. Problem definition. Accountability. The on-ramp gets steeper, not gentler, and you have to coach harder to get them up it.
If you're running a services business and you haven't thought through this, you're not running a services business in a few years. You're a head count problem waiting to happen.
The Bookends Are Where The Trust Lives
Here's the part operators keep getting wrong. They look at AI eating the middle and they assume the bookends are next. So they try to push AI into the customer-facing moment too, and it goes badly almost every time.
Klarna spent most of 2024 bragging about how its AI agent was doing the work of 700 customer service reps. The CEO went on every podcast that would have him. Cost down, efficiency up, board thrilled. Eight months later, they were quietly rehiring humans because customer satisfaction had cratered. The AI was fast. It was also wrong in ways that ate the relationship.
Air Canada had a chatbot tell a grieving customer he could apply for a bereavement fare refund after his flight. He couldn't. The chatbot made it up. He sued. The airline argued in court that the chatbot was a "separate legal entity" responsible for its own answers. A Canadian tribunal looked at them and said, no, that's not how this works. The company owns what its AI says. You can't outsource your accountability to a model. Air Canada paid.
McDonald's spent three years piloting AI drive-thru order taking with IBM. Customers ordering bacon when they wanted ice cream. Hundreds of nuggets being added to single orders as a TikTok meme. They killed the pilot in 2024.
Sports Illustrated, once the most trusted name in American sports media, got caught publishing AI-generated articles with fake author bylines and AI-generated author headshots. The brand never recovered from that news cycle. CNET did the same thing in 2023 with finance articles riddled with errors that had to be quietly corrected.
There's a pattern in all of these. The companies pushed AI past the middle and into the bookend, into the moment where the customer expects a human to be responsible. And every time, the customer noticed. Trust collapsed faster than the efficiency gains compounded.
The bookends matter more now, not less, because they're the only thing left that's actually scarce.
What Stays Human
A few specific things, in our experience:
Defining the problem. AI solves the problem you hand it. It's bad at noticing the problem you didn't see, or telling you you're solving the wrong one. The most valuable conversation a client has with us isn't about the audit we ran. It's about the question we asked them that reframed what they were trying to do.
Taste and standards. AI averages toward the median of its training data. If you're trying to build a brand that stands out, you have to set the bar above the median. We see this every week. Clients send us AI-generated content and ask why it isn't performing. It isn't performing because it sounds like everyone else's AI-generated content. The bar got reset and nobody told them.
Accountability. Someone has to sign their name. Someone has to look the customer in the eye when it goes wrong. Someone has to take the call from the regulator, the board, the press. Air Canada tried to argue otherwise and lost.
Trust. Sales that hinge on the buyer believing you. Therapy. Mentorship. Hard conversations with a team. Leadership in a crisis. The medium is the person. Nobody wants their bypass surgery done by a chatbot, no matter what the benchmarks say.
Operating the AI itself. Knowing what to ask. Knowing what to trust. Knowing when to override. Knowing how to direct ten agents at once toward a single outcome that actually makes sense. This is the most valuable skill of the next decade and almost nobody is teaching it deliberately.
What We Actually Do At WE-DO
We run growth marketing for north of a hundred clients, and we've rebuilt how we deliver around exactly this pattern.
The middle is automated, aggressively. SEO audits run live against the data, not a template. Performance dashboards pull from GA4, Search Console, Google Ads, and DataForSEO in real time and assemble themselves. Content drafts are generated, structured, and pre-optimized before a human ever opens the doc. Technical research that used to eat a junior's afternoon now happens in the background while we're on a call. Competitive analyses that took a week now exist in 20 minutes.
That's the production layer. We invested heavily there because the cost of doing it the old way is no longer competitive with the speed and quality of doing it the new way. If we hadn't, somebody else would have, and we'd be the expensive option that's also slow.
But the bookends are aggressively human, and we've actually pushed harder there. We sit with a client and ask the question they hadn't thought to ask. We set the standard for what their brand should sound like and refuse to ship work that doesn't clear it. We make the call on which initiative actually moves the business this quarter versus which one just looks productive on a roadmap. We read the room on a hard call. We stake our name on the recommendation.
When you hire WE-DO, you're not hiring us to do what AI can do. AI is cheap. You can do that yourself. You're hiring us to direct it on your behalf, with the judgment that comes from a thousand prior reps across industries, and the accountability that comes from a team of humans who answer the phone when something matters.
The Trade
The companies winning right now are the ones who got honest about which part of their work was production and which part was direction, and stopped pretending the two were the same thing.
The ones losing are the ones still pricing the production layer like it's 2019, still staffing it like it's 2019, and getting underbid every week by competitors who let AI do the middle and put their humans on the parts that actually compound.
Production gets cheaper every quarter. Direction, taste, accountability, and trust get more valuable. That's the trade. It's not going to reverse.
If you're trying to figure out which side of that line your business is on, we should talk. That's the most useful conversation we have these days, and it doesn't cost anything to have it.
