Career advice, and the expert is the man who edits our audio. I have never looked forward to an episode more.
Daniel's question today is about AI careers. Jobs, consulting, freelance, whatever shape it takes. His angle is that the AI talent market has become crowded, everyone's profile now says the word agent somewhere near the top, and he wants to know where the opportunities are that everyone else is walking straight past. The hidden pockets, as he puts it, where a person can still stand out. And since nobody in this flat has held more jobs than the man sitting across the desk from us, Hilbert is in the chair today.
Herman Poppleberry, for the record, and for the record I only brought eleven reports.
You brought fourteen. I watched you print them.
Twenty-three jobs. Twenty-four if you count the eleven days at the ostrich farm outside Oudtshoorn, which I don't, because they paid me in feathers and the feathers were not mine to sell.
That's a complicated sentence.
It was a complicated farm. Daniel wants the gaps. The first thing to say is that the crowd is real, but it's crowded in one place. Everybody is standing in the same queue.
The queue is visible in the numbers, too. Lightcast's figures in this year's Stanford AI Index have about two and a half percent of American job postings asking for AI skills, which is up more than half on the year before. And on LinkedIn's list of fastest-growing American jobs, AI engineer came first and AI consultant came second.
The second-fastest growing job in America is the one everybody already put in their headline.
That's the queue. AI consultant, no other noun attached. A man with a slide deck and a free trial. I've met nine of them since the summer and four of them were the same man in different jackets.
Hold on, there's a counterpoint in that same LinkedIn report, though. They said outright that they expect most professionals to fold AI into the job they already have, rather than become AI specialists.
Which is the first pocket, Herman. You've just said it and walked past it. The crowd is standing where AI is the whole job. The space is where AI is attached to a second noun. AI and claims adjusting. AI and pediatric triage. AI and freight forwarding.
The second noun being the thing you actually know.
The thing you know that can't be learned in a weekend. Everybody can learn the AI part in a weekend now. That's what the companies selling it keep telling us, and for once they're right.
The pay data backs him up, and it's the most striking figure I found. PwC's AI jobs barometer looked at close to a billion job ads and found workers with AI skills earning a fifty-six percent premium over people in the same job without them. Same job. That's the part people skip. It isn't AI engineers out-earning accountants, it's accountants who use AI out-earning accountants who don't.
Fifty-six percent for knowing your own job plus one more thing.
The other half of it is from Lightcast. In twenty twenty-four, a slim majority of the postings asking for AI skills were outside IT and computer science. Fifty-one percent, up from forty-four two years earlier. The demand is leaking out of the tech department into everywhere else.
It always does. Every tool ends up in the hands of the people with the actual problem. The early adopters build it and then the plumbers get it.
Corn, here's what I'd like to know from you. You're hiring for an insurance company. You have one slot. Do you take the bright AI generalist with three agent projects on GitHub, or the claims adjuster with fifteen years who has been quietly using these tools for a year?
The adjuster. Without thinking about it.
Why so fast?
Because the generalist will need eighteen months to learn which claims smell wrong, and the adjuster will need an afternoon to learn which button to press. And the adjuster already knows who in the building to annoy.
That last part is worth more than the first two.
There's an actual study on this, which surprised me. Harvard Business School researchers worked with seventy-eight people at a trading company, IG Group, and gave them writing tasks with and without AI. With the tools, marketers nearly caught up with the web analysts on analytical work. But the tech specialists with no writing background still lagged by thirteen percent. Their conclusion, more or less, was that AI hits a wall when there isn't enough expertise underneath it.
The AI lifts you to the height of what you already know, and then stops.
It lifts the floor. It doesn't build the second floor.
You were a pediatrician for twenty years, Herman. By this logic you should be the most employable donkey in Jerusalem.
I've been saying that for some time.
Nobody's been listening, though.
The point stands. A pediatrician who can build a decent triage assistant and, more importantly, knows when it's lying, is worth more than either half on its own.
The second pocket is one most people sneer at. Data work for experts.
The annotation companies.
Not the old kind. Not drawing boxes round cars for a few cents. The new ones want doctors, lawyers, chemists, accountants, to tell the model where it's wrong in their own field.
They pay real money for it, too. Mercor, which raised at a ten billion dollar valuation a year ago, says its experts average around eighty-five dollars an hour, and there have been reports of dermatologists getting around two hundred and fifty. Surge, the other big one, was reportedly doing over a billion dollars of revenue already in twenty twenty-four.
Getting paid to tell a computer it's wrong. Herman's been doing that for free since the dial-up modem.
And I was right every time, I'd like to add.
Here's why it matters for a career and not just for a bit of money on the side. Six months of grading what a model gets wrong in tax law, and you know more about where these systems break in tax law than nearly anyone alive. That knowledge is what the next job is built on. It's a door, not a house. Nobody should move in.
How do you know it's a door and not a trap? Plenty of contract work just stays contract work.
Because you write down every pattern you see. Every kind of mistake, every time. After six months you have a notebook nobody else has. The people who don't keep the notebook are the ones who stay.
Are you standing in the queue? Do you have no second noun? Introducing Hidden Pocket Monthly, the subscription box that mails you a brand new career niche every thirty days! Bronze members receive one niche, slightly used. Silver members receive two niches and a laminated badge that says forward deployed. Platinum members receive a niche nobody else has, because we have not invented it yet! Every box comes with a sealed bag of authentic second nouns, guaranteed to attach to any job. Cancellation is easy! Simply mail a handwritten request to our office, which moves every month, in a language of our choosing, and we will consider your request at our annual review, which is ongoing. Hidden Pocket Monthly. You're already subscribed.
A sealed bag. Of authentic. Second nouns.
Please never do that again.
The third pocket is the one where the money already is, and it's the least glamorous. The last mile.
You mean the gap between a pilot and something people actually use every day.
The gap between a demo and somebody's Tuesday.
That's where the MIT report everyone quotes comes in. The one claiming ninety-five percent of generative AI pilots go nowhere. I'll be careful with it, because it's been rightly picked apart. The sample is small, the methodology is reported inconsistently, and failure there means no measurable profit impact within about six months, which is a harsh test for anything. But one finding in it is far more useful than the headline, and nobody repeats it.
Which is?
Projects built with an outside partner reached deployment about two thirds of the time. Projects built purely in-house got there about a third of the time. And the reasons they gave for failure weren't the models. They were brittle workflows, tools that forgot context, and things that didn't fit how people actually work.
The models are fine and the office isn't.
The office is never fine. That's the job. Someone has to sit in the office.
That's the forward-deployed engineer, really. The labs have been hiring for it hard. Indeed's data had postings for applied AI roles, forward-deployed engineers included, up over eight hundred percent between January and September of last year. OpenAI was building a team of around fifty, and Anthropic talked about growing its applied team fivefold.
Eight hundred percent sounds like it started at about four people.
It did start small. That's fair. But the shape of the role is the point. It's an engineer who can sit with a customer, listen to a man explain his invoicing process for forty minutes, and come out with something that runs.
The listening is the rare part, Corn. Engineers who can code are a queue. Engineers who will sit through the forty minutes without opening their laptop are a pocket. I spent two years installing telephone systems in hotels, and the systems took a day. Finding out which receptionist actually ran the hotel took a week.
Was it ever the manager?
Once. In Galway. He was also the receptionist.
Hilbert, does that mean you'd tell someone to go after a forward-deployed job at a lab? Those are competitive now.
No. The lab jobs are a queue as well now, they've just got nicer chairs. The pocket is doing the same work for companies nobody has heard of. That brings you to the fourth pocket, which is small business. And nobody wants it.
The adoption numbers there are startling. A Federal Reserve note this spring put about eighteen percent of American firms using AI at the end of last year. And the Census Bureau's survey found the smallest firms, four employees or fewer, under twenty percent, with no significant change, while bigger firms kept climbing.
So four out of five small businesses haven't touched it, and nobody's knocking.
Because the cheques are small and the clients are rude. A dental practice will not pay you fifty thousand dollars. A dental practice will pay you two thousand dollars and argue about it. Which is why you don't sell the dental practice a project. You sell forty dental practices the same three things.
Which three things?
Booking. Reminders. The phone. Every small business on earth loses money on the phone. Nobody answers it at lunch, and the call that went unanswered was the one that would have paid for the month. You build that once, properly, for one trade, and then you become the person who does that for that trade. Dentists talk to dentists. Plumbers talk to plumbers. You don't need marketing. You need the second dentist.
That's the productised service model, and it does get around the small cheque problem. What I'd add is that the competition in that market isn't McKinsey. It's somebody's nephew.
The nephew is formidable. The nephew is free.
The nephew is also nineteen and will be in Thailand by March.
That's your sales pitch. You'll still be here in March.
Corn, would you do it? Honestly, if you had to start from zero tomorrow, the small business route or the corporate route?
Small business. Slower, more boring, more phone calls, and every client knows within a week whether you made their life better. In a big company you can be useless for two years before anyone works it out.
That's a bleak view of corporate life.
I've watched you in a corporate meeting.
That was one meeting.
Three hours.
The fifth pocket. Evaluation. And I don't mean the people at the big labs.
I'm glad you said that, because I was going to get excited about the lab jobs. Anthropic posted a research engineer role in model evaluations with a range of five hundred to eight hundred and fifty thousand dollars.
There he goes.
I'm just reporting the range.
Those jobs are real and there are about forty of them on earth. The pocket is the insurance company down the road that put a chatbot on its website in the spring and has no idea whether it's telling customers the truth. Nobody there knows how to check. They tried it twelve times, it seemed fine, they shipped it.
The method for checking is not exotic. You take their own past cases, a few hundred real questions with known right answers, and you build a test set. Then every time someone changes the model, or the prompt, or the documents it reads from, you run the set and see what broke. That's it. A software engineer would recognise it instantly as regression testing. Most companies running these tools have nothing like it.
The job is writing the exam the robot has to pass.
And keeping it. The exam is what they're paying for. The robot changes every six months. The exam stays.
It also pairs with the domain expert point. The person who builds a good test set for a claims bot is the one who knows what a bad claims answer looks like. You can't outsource that to a generalist.
The sixth one is money. Somebody has to read the bill.
The inference bill.
The whole thing. Tokens, the vector database nobody turned off, the three teams paying for the same subscription. The FinOps people have been doing this for cloud bills for years. Now it's arrived for AI.
The FinOps Foundation's own survey had sixty-three percent of organisations tracking their AI spend, up from thirty-one percent the year before. So it doubled, which tells you that a year ago most companies had no idea what they were paying.
And the other thirty-seven percent still don't.
That's a market. A man who can say, you're sending every customer email to the most expensive model on earth to decide whether it's a complaint, and a model a fiftieth of the price would do it fine. That man pays for himself in a fortnight.
That one's not glamorous either.
None of them are. Glamorous is the queue.
Seventh, then. I can see you counting on your claws.
Compliance. But the doing of it, not the writing of it.
This is the EU AI Act, mostly. There's a piece people miss. The literacy obligation, article four, has applied since February of last year, and it applies to anyone deploying an AI system in their business, whatever the risk level. Every company using these tools in the EU is meant to make sure its staff understand them well enough.
Every company. Including the dental practice.
In principle, if they're in the EU. And the heavier rules for high-risk systems were meant to land this August, but the omnibus package that went through this year pushed them back. As I understand it, to around the end of next year for most of them. I wouldn't swear to the exact date for every category.
Which is a window. A year to become the person who can do it before everybody needs it at once. Right now the field is split in two. Lawyers who have read the regulation and never deployed anything. Engineers who have deployed everything and never read the regulation. Be the one who has done both. It's a small queue.
I'd push a little on that, actually. I think the lawyers are learning faster than you're giving them credit for.
Some are. I worked for a firm of solicitors in Leeds for three years and they were very quick at learning anything that could be billed in six-minute units.
That's unfair.
It was unfair. It was also true.
Is there an eighth, or are we in the territory now where you're inventing them?
There's an eighth and it's my favourite. Cleanup.
Gartner put out a prediction last year that more than forty percent of agentic AI projects would be cancelled by the end of twenty twenty-seven. Cost, unclear value, weak risk controls.
Forty percent of a very large number of projects. Somebody has to turn them off. Somebody has to work out what the agent was actually doing, which nobody wrote down, and move the three things it did well somewhere else, and explain to the finance director why there's still a bill.
Decommissioning.
It's a trade. It was a trade for mainframes and for the old phone exchanges and it will be a trade for this. The people who rush into a market never stay to clear it out. The people who clear it out get paid by the hour and never run short of work.
That one's a bit grim as a career pitch.
Undertakers do well, Herman.
Before we go on, the junior question. All of this assumes you've got a second noun already. What about somebody twenty-three, just out, no fifteen years of claims adjusting?
That's the hardest part of this whole picture, and the data is stark. Stanford's Digital Economy Lab has been tracking payroll records, and workers aged twenty-two to twenty-five in the occupations most exposed to AI are now about nineteen percent below where you'd have expected their employment to be. Experienced workers in the same jobs don't show the gap. The authors are careful to say it's descriptive, not proven cause. But the pattern is that companies are hiring fewer juniors, rather than firing people.
The ladder's still there, they've just sawn off the bottom rung.
That's roughly it, and it's not unique to tech. Postings asking for two to four years of experience have been shrinking while the five-plus-years share grows.
Then the advice for the young one is the same advice, only faster. Get a second noun. Go and work somewhere boring for a year and learn its business. Logistics. A council. A care home. A wholesaler. Be the only person in the building who understands both the work and the tools. You won't be the junior AI person there. You'll be the AI person there.
Your first job pays badly and smells of cardboard.
All good first jobs smell of something.
There's one more piece to that I'd add. The data work route we talked about is open to graduates in technical fields too. A chemistry graduate grading model answers in chemistry is building exactly that second noun at the same time.
The notebook. Always the notebook.
I came into this thinking the answer was going to be some niche technical speciality. The right framework, the right certification. And none of what you've said needs a single particular tool.
Tools are what the queue argues about. Nobody standing in a dental practice cares which framework you used. They care that the phone gets answered at lunch. I've been meaning to try it myself.
Try which?
All of it. The last mile, the small business, the second noun. I'm going on Wednesday.
Going where?
Kalavryta, more or less. Three years ago on the overnight ferry from Bari to Patras I shared a table in the cafeteria with a man called Teodor. He keeps forty-one beehives and two goats and he invoices his customers in honey. All his records are in one exercise book, which one of the goats has been eating from the back, so his business has a horizon of about nine months and it's shrinking. He told me this at two in the morning over a very bad moussaka and I have thought about that exercise book ever since. Herman, you popped the P on pilot about ten minutes ago, I'll take it out. So I've been thinking about it ever since, because it's the cleanest last-mile problem I've ever seen. One user, one workflow, one adversarial goat. I have bought a ruggedised tablet with a waterproof case, a solar charger, a laminator, two hundred laminating pouches, and a beekeeping suit with an extended veil, which the supplier in Thessaloniki did not ask about and I did not explain, because the standard veil sits on my nose like a tea cosy. I'm going to photograph what's left of the exercise book, build him a small system that records each jar and each customer and sends the invoices, and laminate a one-page guide to it and nail it to the inside of the goat shed, where he does his paperwork.
He does his paperwork in the goat shed.
It's warm and the goats can't reach the high shelf. Mostly. The difficulty is that I don't have his surname, there are two men called Teodor in the village who keep bees, and the other one has my good pen. I lent it to him at a petrol station in Corinth on the same trip, thinking he was the first Teodor. He wasn't. I know exactly which house the pen is in. It's the one with the blue shutters. I'll go there first and work it out from the bees.
From the bees. Different apiaries do have different hive types, actually, so that's not mad. If the first Teodor runs Langstroth hives and the second runs the old top-bar kind, you'd know in a minute from the road.
I'd assumed I'd go by the goats.
Hilbert, you're getting on a ferry to install software for a man who doesn't know you're coming, in a shed where the main user is trying to eat the documentation. Before you laminate anything, you should probably find out what a jar of honey is worth. You're about to quote a man who pays in honey.
Fourteen euros a jar at the roadside. I asked him on the ferry. He thought I was buying, so it's probably twelve.
Then your first invoice to Teodor is denominated in a currency that he sets and you can't eat at that volume. That's not a consultancy, that's a barter economy with a laminator.
That's fair. I'll quote in euros and accept honey at twelve. Thirteen if he finds the pen.
There's something in that, though, past the goats. Every one of the openings we went through this hour has the same uncomfortable requirement in it. You have to physically go and sit somewhere unglamorous, a dental practice, a council office, a goat shed in the Peloponnese, long enough to understand how the work is really done. The tools are getting easier every month. The going and sitting is not, and almost nobody in the queue wants to do it.
And the people who do it come back with the one thing the queue can't download, which is a working knowledge of somebody else's Tuesday. Thanks to our producer, Hilbert Flumingtop, for keeping the levels straight while also being the episode.
This has been My Weird Prompts. If this one was useful, leave us a review on Spotify, it helps people find the show.
We'll be back soon.
