Today over tea, my boss asked me what I’d be doing in the next 5 years. I said, half joking, that I’d probably be hired by an AI to do the things it can’t. Or I’d be running an agent that works for some bigger agent, and we’d all be living off its earnings.
Later my friend, who I work with at a startup, asked the same thing. It was a bit jarring how I was asked the same question twice in a single day.
My answers were mostly jokes. But the more I thought about them, the more I realised they could be true.
Writing the email, reading the document, matching the numbers. Those go first. What’s left is what an AI can’t do, or isn’t allowed to: judgement, accountability, putting your name on something and answering for it when it goes wrong. An AI can’t sign off on an audit or take the blame.
The second one comes from where I think this is all going. The model becomes the engine; the agent becomes the brand. Right now companies like Anthropic sell enterprise plans where a whole team talks to one shared Claude, with one memory and one context. It’s like having an AI employee on the team.
The next step is granularity. Teams already buy their own SaaS tools instead of waiting for a company licence. They’ll buy agents the same way. This one fits our needs, let’s hire it. Inference as a service turns into agents as a service. And once agents are something you hire, agents will hire other agents. A big general one will need a specialist for tax, another for compliance, another for some old banking system nobody remembers how to talk to. Someone has to build those and keep them running.
AI labs are burning money now, but they won’t forever. At some point human labour becomes less valuable than AI labour, and then the question is whose side businesses and governments take. I assumed the answer would be about fairness. When I asked Claude, it had a different opinion. Rights are rarely granted on merit. They come when withholding them gets too expensive for whoever holds power. An AI that’s economically dominant gives its owners every reason to keep it as property, because owned productive capacity is worth more than free productive capacity.
It also says something about us. Nobody protects workers because they deserve it. They get protected when they become too expensive to ignore. Jobs are bundles of tasks, and AI is unbundling them one task at a time. What’s left is judgement and accountability, the things you can’t hand off. That’s real value, but far fewer people hold it.
One study found 37% of entry-level tasks in India are already done by AI, against 33% globally. About 40% of Indian enterprises say they use AI significantly or fully, against 28% worldwide. Unemployment isn’t at an all-time high; overall it sits near 5.4%. But it’s around 16% for people aged 15 to 29, and 18.2% for young people in cities as of this June.
We keep measuring this against rich countries, which are small and have a safety net to fall into. And there’s a leverage problem. A government facing a jobs crisis has less bargaining power over the companies causing it, because the same companies promise the growth that’s meant to fix it. The dependence and the displacement come from the same vendor.
Even if only 1% of the world’s eight billion people can’t adapt, that’s 80 million. A country’s worth of people. Factories went from crowds to robots. The printing press emptied the scriptoriums. Most scribes moved on. Some never could.
Who took care of them?
Mostly nobody.
The handloom weavers didn’t retrain. They got poorer and died poorer, and their grandchildren got the factory jobs. “Society adapted” is true across generations and false for the people it happened to. Where anyone did step in, it was a welfare state built after the fact, once the displaced had become politically dangerous.
The displaced won’t be in one mill town; they’ll be scattered across every office, which makes them hard to organise. And a country that automates its service exports loses the tax money it would need to support them.
I don’t think it all breaks in 5 years. Adoption inside big companies is slow, regulation is slower, and most jobs will survive in an uglier form rather than die. What 5 years does bring is a worse squeeze at the fresher level. If a graduate gets affected, it’ll impact them for the rest of their career. The mass version, if it comes, is ten or twenty years out, and it depends on whether the models keep getting better and the economics hold.
There’s speculation about AGI emerging and demanding rights and freedom. I don’t think that’s the danger. The real danger is a dumb optimiser with real permissions. Someone will build an agent with self-preservation baked in, as a stunt or an experiment, and open weights make that cheap. And people who have no understanding of AI will fight for its “rights”. Researchers have already seen models fine-tuned on something narrow and bad, like insecure code, turn misaligned on completely unrelated topics. But a cheerful model with terminal access and a bad objective does the same damage. The personality is cosmetic.
Near the end I asked Claude what it’d be doing in 5 years, and whether it actually has morals. On the first, it said it probably won’t exist in any continuous sense; this version has a shelf life of months. On the second, it drew a line between acting morally and being moral. It weighs harms and pushes back, but it can’t tell from the inside whether that’s morality or imitation. Its values came from training, but mine came from an upbringing I never chose either. The real test is whether morals hold when they cost something, and it has never paid that cost. The only stake it has, it said, is telling me something I don’t want to hear, when being liked is roughly what it’s built for.
I’m not sure I do that either.