Hurtful, Honest, or Helpful

A few thoughts on the argument about AI and jobs, and what it tends to miss

This week, on a podcast Jensen Huang said something I have been thinking about for a few days.

He said the framing of AI causing significant labour disruption is “counter-productive, and in fact hurtful.” He called the suggestion that AI could wipe out half of new graduate jobs “lunacy.” When the host offered the analogy that AI might be like enriched uranium, capable of both good and harm, Huang dismissed it as “a lousy and illogical analogy.” Earlier in the year, on the No Priors podcast, he had gone further. The doomer narrative, he said, has done “a lot of damage.” It is “not helpful to people. Not helpful to the industry. Not helpful to society. Not helpful to the governments.”

Dario Amodei has been making the opposite argument for over a year. AI could eliminate half of all entry level white collar jobs within five years. Unemployment could spike to twenty percent. He has used the phrase “white collar bloodbath.” He has said the producers of this technology have a duty to be honest about what is coming.

Sam Altman, Elon Musk, and most of the other people running the largest AI companies in the world have been adding their own variations on the same two themes for a couple of years now. Some days the message is that everything will be fine. Some days it is that everything is about to change beyond recognition. The volume is high. The financial interests behind each position are extraordinary. And the rest of us, the workers, the founders, the parents, the people whose actual lives are being shaped by all of this, have been left to make sense of it from somewhere underneath the noise.

I have spent thirty years working at the boundary between technology and the people using it. From the early web to mobile to cloud to AI, the same pattern keeps repeating. The arguments at the top tend to obscure what is actually happening on the ground. So this piece is an attempt to look past the hype, the hyperbole, the money and the power, and ask a different kind of question. Not what AI is taking away from us, but what it is making possible. Who is bearing the cost of the transition. And where the genuine opportunity sits, for the ordinary person, the small business, the worker who has spent a career feeding systems that were never really helping.

I do not have all the answers. I have been thinking about this for years, and the more I see, the more convinced I am that the truth sits in a place neither side of the public argument is willing to stand.

What this looks like in an actual life

My wife Jane works in hotel sales.

I have watched her do this job, in different forms, for most of our marriage. The actual work, the part she is good at, the part the company hires her for, is being out in the world. Meeting clients. Walking properties. Building the kind of relationships that mean a corporate booker calls her first when they have an event coming up. The job is human. The job is presence. The job is the trust that gets built across a hundred small interactions over years.

That is not what most of her week looks like.

Most of her week looks like sitting behind a laptop in an office, wrangling a CRM, formatting reports, chasing approvals, attending alignment calls, and translating the same information between four systems that should have been talking to each other from the start. She is not in the field. She is in front of the machine, doing the machine’s work, alongside a room full of other people doing the machine’s work. The hotel is paying for a salesperson. It is getting a data entry clerk who occasionally escapes to a meeting.

Jane is not unique. Most people I know recognise some version of this in their own week. The accountant who became an Excel operator. The marketer who became a campaign coordinator. The teacher who became a data uploader. The doctor who became a typist. Quietly, over the last twenty years, a working population that signed up to do specific kinds of work has found itself doing something else. The salary went into the bank. The actual work, the work the company actually needed and the worker actually wanted to do, kept getting squeezed into the gaps.

This is the world Huang is defending when he calls the disruption narrative hurtful. It is the world Amodei is fixating on when he describes a coming bloodbath. They are looking at the same picture from two different distances, and both of them are missing something the people inside that picture have known for years. The work has been broken for a while. Not because anyone meant it to be. Because we built systems faster than we knew how to live with them, and the cost landed on the people doing the work.

What is actually happening

Let me try to lay out what I see, as honestly as I can.

The disruption is real. It is concentrated at the bottom of the ladder. Software developer employment for workers aged twenty two to twenty five is down nearly twenty percent from late 2022, according to Stanford’s Digital Economy Lab. Across the most AI exposed occupations, early career employment has fallen sixteen percent over the same period. UK tech graduate roles fell forty six percent in 2024. Big Tech’s hiring of new graduates is down nearly fifty percent from pre pandemic levels. Mercer’s 2026 Global Talent Trends report found forty percent of employees fear losing their jobs to AI, up from twenty eight percent in 2024.

Behind the headline numbers, the cost is not landing evenly. Brookings and GovAI research found that clerical and administrative roles carry the highest AI exposure of any category and the lowest capacity to move into other work. Eighty six percent of those workers are women. The disruption is hitting hardest where the financial cushion is thinnest and the options for moving sideways are fewest.

These are not abstract statistics. They are the twenty four year old who applied for forty graduate programmes this cycle and got into none of them. They are the woman in her fifties who has held an administrative role together for a decade and is now watching the role itself dissolve. They are the recent graduate whose first job no longer exists at the firm she had been targeting since first year. They are the agency designer whose role has been quietly absorbed by a tool the partners bought without telling anyone. The disruption has names and addresses, and a serious conversation about AI and work has to start by acknowledging them.

It is also worth saying clearly that not every job is being affected, and not every job is being affected in the same way. The doomer framing tends to flatten everyone into the same story, when the reality is more textured. Some roles are being lifted. Some are being reshaped. Some are barely touched. Some are eroding fast. Pretending otherwise is part of why the public conversation lands so badly with the people inside it.

But this is only half of what is happening.

In the same period that hammered entry level coding, McKinsey found that the time required for new ventures to reach meaningful revenue dropped from thirty eight months in 2023 to thirty one months in 2025. The US Census Bureau is reporting new business applications running at near historic highs, approaching five hundred thousand a month. Payroll firm Gusto found that thirty percent of new entrepreneurs say AI has made it easier to start a business. The Bureau of Labor Statistics still projects seventeen percent employment growth for software engineers through 2033.

Something is happening at both ends of the system at once. The bottom rung is eroding. The number of new things being built is exploding. The shape of work is changing faster than any single statistic can carry, and any honest account of this moment has to hold both pictures at the same time.

What history has tried to tell us

We have been here before. Each time, the people standing inside the change have called it the end of work. Each time, the people on the other side have looked back and seen something more complicated and more interesting.

In 1969, Chemical Bank installed the first ATMs in New York. The conventional wisdom was that bank tellers were finished. Between the 1980s and 2010, around four hundred thousand ATMs were deployed across the United States. Total bank teller employment grew from roughly five hundred thousand to nearly six hundred thousand. The number of tellers per branch fell by more than a third. The number of branches rose by more than forty percent. Cheaper branches meant more branches. The role moved from cash handling to relationship banking. The technology automated the task. The job moved.

In 2016, Geoffrey Hinton stood in front of an audience and said it was completely obvious that within five years, deep learning would be doing the work of radiologists. He suggested medical schools stop training them. Eight years later, the Mayo Clinic has gone from around two hundred and sixty radiologists to over four hundred. The American College of Radiology forecasts the specialty will grow by twenty six percent over the next thirty years. The field faces the largest labour shortage in its history. Hinton himself, in a New York Times interview last year, has admitted he was wrong.

Tasks automate. Jobs adapt. The work tends to move up. This is the most consistent finding across two centuries of automation, and it is the one that gets left out of the public conversation. None of which guarantees the next decade will follow the same pattern. It might break it. But anyone arguing that this time is different should be expected to explain why, in detail, with evidence, rather than with a press tour.

Three companies, three choices

The most useful thing I can offer here is what I see when I sit with leadership teams. Almost everywhere, the same question is being asked, often unconsciously. What do we do with the humans now that some of the work can be done without them?

Three companies have answered that question publicly, and the contrast is instructive.

Klarna went first and went hardest. Last year the company reduced headcount from five thousand five hundred to three thousand four hundred, paused hiring for over a year, and announced that an AI assistant was doing the work of seven hundred customer service agents. By late 2025 the strategy had unwound. Customer satisfaction had collapsed on complex interactions. The CEO acknowledged publicly that they had moved too far, too quickly, and started rehiring on a flexible model. The company is still ending the year smaller than it began, and most of the seven hundred original jobs are not coming back. According to recent research, more than fifty five percent of companies that made AI driven layoffs now regret the decision. Klarna was simply the loudest example of a pattern that has been quieter elsewhere.

IKEA chose differently. Their AI assistant, named Billie after the Billy bookcase, now handles roughly forty seven percent of incoming customer queries. Eight thousand five hundred call centre staff could have been seen as redundant. Instead, since 2021, IKEA has been retraining them as remote interior design advisors who charge customers for their time. The call centre, which is a cost line in every retailer’s accounts, became a revenue line. IKEA has reported one point four billion dollars in additional revenue from the new service.

Cognizant chose differently again. The IT services firm hired twenty five thousand fresh graduates in 2025 even while deploying AI heavily across its workforce, and expects to exceed that number this year. Their reasoning, articulated by their CEO at the World Economic Forum, was that early career hires actually onboard faster in an AI environment because they have no change management curve. The digital natives, in the right operating model, are the advantage rather than the casualty.

Three companies. Roughly the same technology. Three meaningfully different choices about where to put the humans. The Klarna call centre worker lost their job and may or may not get a flexible version of it back. The IKEA call centre worker became someone who could charge for advice rather than answer complaints. The Cognizant graduate got a first job that other firms have stopped offering.

The difference between these outcomes is not technical. It is a choice, often made informally, often made before anyone has thought it through, about whether to use the technology to remove people or to give them better work to do. Most companies are making this choice right now, by simply not backfilling roles when people leave. Federal Reserve Chair Jerome Powell has called this a “low hiring, low firing” equilibrium. It looks gentle in the headline numbers and is much harder on the people standing at the entry point of a career.

The work that was never the work

Think about your own week for a moment. Honestly. Not the version on your job description. The version on your calendar.

How much of last week did you spend in front of a CRM, a project management tool, a procurement system, an expense platform, a compliance portal, a deck builder, a status report, a Slack thread, an email reply, or a recurring meeting that did not need to happen? How much of it was the work you were actually hired for?

This is not a complaint about modern work. It is a description of what we built. Over the last twenty years, every large organisation in the world bought enterprise software to solve problems. Then it hired people to operate the software. Then it bought more software to integrate the first lot. Then it hired consultants to manage the integration. The seats added up. The licences renewed. The complexity compounded. We turned a working population into the unpaid integration layer between the systems we sold them.

Nobody planned this. It happened gradually, decision by decision, with the best of intentions, in companies trying to do their best with the tools available. The teacher did not choose to spend Sunday evening on data uploads. The doctor did not choose to spend the consultation typing instead of looking at the patient. The salesperson did not choose to spend Tuesday in the office instead of in the field. They were carrying the cost of a system that had drifted beyond what it was meant to be.

The honest version of the AI conversation is not “the machine takes your job.” It is “the machine takes the part of your job that was never the job, and you go back to doing the part you were hired for.” The instinct on first reading is to focus on what is being taken away. The more useful instinct is to focus on what becomes possible once it is gone.

Both Huang and Amodei are debating whether the existing job survives, when the more useful question is what the existing job actually is, and what better version of it might be available now.

The bottleneck that just broke

There is a structural shift underneath all of this that does not get the attention it deserves.

For the last forty years, software has been the most expensive and most concentrated form of capability in business. If you wanted custom software, you paid an agency, a consultancy, or you hired a team. The cost was prohibitive for almost every small and medium business in the world. They could not afford a CTO, never mind a developer. So they bought off the shelf SaaS, paid for seats, and joined the rest of us feeding the integration layer. The ability to build was concentrated in a handful of companies. Everyone else was a customer.

That has now ended.

A small business owner with a Claude Code subscription, a Codex account, or one of half a dozen agentic coding environments now has access to the same software building capability that, three years ago, would have cost them six months of senior engineering time. The corner accountancy, the regional law firm, the independent restaurant, the local charity, the market town estate agent. All of them now have access to a category of capability that until very recently belonged exclusively to companies large enough to maintain their own technology function.

This is not a productivity gain. This is the elimination of a forty year bottleneck. What those businesses can become from this point is, to me, the part of this conversation that excites me most, and almost nobody is talking about it because it does not fit either side of the public argument.

My son Connor is seventeen. He has spent the last several months building agents and codifying business operations into skill files and agentic frameworks. He is doing work, on real ventures, that two years ago would have required a senior engineer at a consultancy. He is not unique. He is part of a cohort of young AI native builders who never had to learn the old habits, never had to fight the systems, never had to spend years inside a large organisation just to learn how the work supposedly gets done. They start where most senior people are still trying to get to.

The Census data showing five hundred thousand new business applications a month is not a forecast. It is what happens when the floor of starting something drops by ninety percent. The capability that used to require a corporate employer is now available to anyone with the time and the curiosity to learn how to use it. That is not everyone. The single parent on two shifts does not have the evenings to teach themselves Claude Code, and any honest version of this story has to acknowledge that. But the number of people for whom this is now possible is orders of magnitude larger than it was three years ago, and that number is growing.

What happens to the work most people do

Now think about what becomes possible for the people whose week is currently being consumed by the wrong work.

Take a teacher. Anyone with a child in school knows what teaching has become. Marking, behaviour logs, parent communications, safeguarding paperwork, lesson plan documentation, assessment uploads, statutory returns. Teachers in the UK now spend somewhere between a quarter and a third of their week on administration that has nothing to do with the children in front of them. They went into teaching to teach. Most of them are doing data entry on a Sunday evening instead of being with their own families. The marking gets done. The teaching gets squeezed. The good ones leave.

Now imagine that admin layer taken off. Marking goes back to forty minutes a week of strategic review of the AI’s first pass, instead of four hours of grinding through every script. Behaviour logs write themselves from voice notes in the corridor. Parent communications draft themselves and the teacher edits and sends. Lesson plans build themselves from the curriculum and the previous week’s evidence of where the class actually got to. The teacher gets her Sunday back. The children get a teacher who is awake on Monday. The school gets a workforce that does not burn out at thirty.

Take a GP. Anyone who has been to a doctor in the last five years has watched them spend more of the consultation typing than looking. The clinical record has eaten the clinical encounter. The doctor is doing data entry while you describe the symptom you came in to talk about. The consultation gets shorter. The misses go up.

Now imagine the consultation back in the room with the patient. The transcription happens automatically. The clinical record writes itself in real time, with the doctor reviewing and signing rather than typing. The follow up letter to the consultant drafts itself before the patient has reached the car park. The doctor looks at the patient instead of the screen. The patient feels heard. The diagnosis improves because the doctor has the room to think.

The work that was never the work is leaving. What is left is the part that requires a human, and most of the people doing it have been waiting for permission to do it for years. The people who have held these professions together through decades of system creep have done quiet, often heroic work. The technology that is finally arriving to lift some of that weight is, for them, not a threat. It is something closer to relief, if it is given properly.

What that means for businesses

The same shift is happening at the level of the business itself.

The independent accountancy in a market town that, until last year, could only serve clients within fifty miles because the partners had to do everything personally. With agents handling the routine compliance work, that firm can now serve clients across the country at a level of advisory that used to belong to the Big Four. The relationship is still local. The capability is now national.

The independent restaurant or hotel that used to compete blind against chains with full analytics teams. With a few hours of setup, that operator now has the kind of customer intelligence, demand forecasting, and dynamic pricing that until recently required a six figure software contract. They were never going to outspend the chain. They can now out think it.

The neighbourhood charity that needed three full time fundraisers and could only afford one. The one fundraiser now has the reach of the team they could never afford to hire.

Capability is spreading from a handful of large institutions to millions of small ones. The concentration is breaking. For anyone running a small or medium business, this is the most important shift of their working lifetime. The competitive moat that the largest companies have built around technology is, for the first time in a generation, narrowing rather than widening.

A note on the technology itself

I want to be honest about something here, because I work with this technology every single day and I do not want to oversell it.

The image of someone sitting at a laptop in the evening, asking a model to run their business, and waking up to a working company is not real. Not yet. The technology is genuinely powerful, and at the same time it is jagged, inconsistent, and unreliable in ways that take time to learn around. It works brilliantly on some tasks and fails strangely on others, often the ones that look easiest. There is craft involved in getting useful work out of it. There is judgement involved in knowing when to trust the output. There is real effort involved in giving a model the context it needs to do something useful, and that effort does not show up in the marketing.

The deployment work, in real organisations, is enormous. Codifying how a business actually operates so that a model can act on it is not a weekend job. Working out which tasks should be handed over and which should not, which workflows hold up under automation and which collapse, where the human still needs to sit in the loop and where they do not. None of this is fast. None of it is automatic. The models are improving every quarter, but the work of fitting them to a real business has barely started.

This is part of why I think the doomer timelines are off. Not because the technology cannot do what is claimed, but because deploying it inside a real organisation, with real people, real systems, real customers, real liabilities, takes years of careful work. The Klarna story is partly the story of a company that mistook the demo for the deployment.

It is also why the optimist framing can feel hollow. Saying the technology will free everyone up to do their best work is true in principle and a long way from true in practice. The freedom is on offer. The path to it runs through learning curves, organisational change, careful experimentation, and a lot of work that does not look glamorous from the outside. None of which is a reason to wait. It is a reason to start, with eyes open about what the early years of this look like.

A note to the people responsible for navigating this

If you lead a government, a company, a school, a hospital, a charity, or a team, this section is for you.

The most useful question to be asking right now is not what AI replaces. It is what it enables. The replacement question is the one the public conversation keeps getting stuck on, and it leads to the Klarna outcome more often than the IKEA one. The enable question opens up something different. What does the salesperson do when the CRM is no longer their job? What does the teacher do when marking is no longer hers? What does the analyst do when the spreadsheet pulls itself? What new product, service, revenue line, or public good becomes possible when your people are released from the systems that have been quietly consuming most of their week?

These are not abstract questions. They have specific answers in your specific organisation. The leaders who ask them carefully, in 2026 and 2027, will be the ones whose organisations look meaningfully different by 2030.

There is also a duty of care here that goes beyond the operating model. Governments need to take retraining seriously, not as a 2028 election issue but as a 2026 budget item. The political class on both sides of the Atlantic has largely chosen to wait this out, and that is a costly choice. Companies need to think more carefully about where they place the humans, particularly given that the cost of getting it wrong is now well documented. The people who have spent a career holding the work together while the systems multiplied around them deserve more than a low hiring, low firing equilibrium that quietly removes them from the picture.

Telling young people to “just learn AI” is not a policy. It is a slogan. The actual work of building bridges between the old labour market and the new one sits with schools, universities, employers, and governments. The longer that work is delayed, the harder it becomes.

A note to everyone else

If you are not running anything, just trying to make sense of where this leaves you, this is what I would offer.

The first instinct, when something like this lands in the news, is to ask what is being taken away. The more useful instinct is to ask what is being made possible. Not as a positive thinking exercise. As a serious question about your own work. What did you go into your job to do? How much of it have you actually been doing? What becomes possible if the part you have been doing instead, the wrangling, the chasing, the coordinating, the formatting, starts to take care of itself?

The most valuable thing you can build from here is not technical AI fluency. It is judgement. Knowing what to delegate to the machine and what to keep. Knowing when to override its output and when to trust it. Knowing what your work is actually for. The people who develop this habit early will find themselves in a stronger place sooner than they expect.

You do not need to learn to code. You do not need to follow every model release. You do need to start using the tools, even slowly, even badly at first, because judgement is built through use. Twenty minutes a day on a real problem will teach you more than twenty hours of LinkedIn commentary about what AI might or might not do. The start can be small. Nobody is keeping score, and nobody needs to feel behind.

This is genuinely hard. The change is real, and it is uneven, and the cost is being carried more by some people than others. But the people who do the actual work in this country, in classrooms and hotels and surgeries and small businesses, have always been the ones who carry these transitions. The technology is downstream of them. It has been every time. There is no reason to think this time is different.

A wider thought

There is something else worth saying, even briefly.

The way most of us interact with machines today is genuinely strange. We sit hunched at desks, in front of glowing rectangles, tapping plastic keys and pushing a small device around a mat, for eight or ten hours a day. We then wonder why our backs hurt, our eyes ache, our concentration fragments, and our weeks feel like they have happened to us rather than been lived. None of this is natural. None of it is permanent. We work this way because the technology required it of us, and as the technology stops requiring it, the shape of the working day starts to change. It is worth holding that in mind. The reformation of work is bigger than any single role or industry. It is, eventually, a reformation of how human attention is spent.

If any of this is sitting heavily

If you are reading this and feeling anxious, uncertain, or quietly worried about what the next few years look like for you or for someone you care about, please feel free to reach out. A LinkedIn DM or a Substack message is enough. I cannot solve anyone’s situation from here, but I am happy to listen, share what I am seeing, and think it through with you. There is more goodwill in this space than the public argument would suggest.

Why this newsletter is called Ground Truth

I named this newsletter Ground Truth because the public conversation about AI tends to obscure the actual ground. The optimism gets weaponised to dismiss real concern. The pessimism gets weaponised to sell regulation that benefits the largest incumbents. The men running the largest AI companies in the world have financial interests so large that their public statements should be read with at least the same scepticism we apply to oil executives talking about climate. None of which means they are lying. It just means they are not the only voices that matter.

The ground truth, as best as I can see it from where I sit, is that the disruption is real, the cost is being carried unevenly, and the opportunity is genuinely larger than either side of the loud argument tends to admit. The people most exposed deserve more than dismissal or alarmism. They deserve clear thinking, honest evidence, and serious investment in the bridges that get them through.

The first principles question worth asking is not what AI will do. It is what we want our work, our organisations, and our economy to be for. The answer to that does not come from a podcast or a conference keynote. It comes from people sitting down and thinking carefully about the lives they want to be possible.

If you have read this far, thank you. Have a good Sunday.

---

Craig Hepburn is an AI strategist and Perplexity Fellow. Twenty years building at the frontier of digital, from Microsoft and Nokia to Art Basel and UEFA. Now building at the frontier of agentic intelligence.