Paying Rent to the Oracle

Apologies for a very long read.

There is a rather bleak little joke forming here, although like most bleak little jokes these days it comes with a spreadsheet, a productivity curve and a management consultant nearby pretending he saw it coming.

For years we were told that Britain’s future was services. We did not need to make things, dig things, forge things, grow things, repair things or generate things with any particular urgency. No, we were going to be a clever country. We would advise, consult, regulate, finance, administrate, brand, strategise and, where necessary, hold a workshop about stakeholder alignment.

It was all very modern. We would leave the grubby business of actually producing things to other countries and concentrate on the high-value work. The sort of work done in glass buildings by people with laptops, lanyards and phrases like “circle back” lodged somewhere deep in the soul.

Then along comes AI. Not the old sort, which wrote faintly odd poems and occasionally told you that glue was a pizza topping. This is the new, more worrying sort: agentic AI. It does not just answer questions. It acts. It plans. It writes code, checks code, runs tests, updates files, drafts documents, processes information, books meetings, chases forms, compares options and quietly does half the work that was supposed to keep the services economy nicely upholstered.

And this is where Britain should probably stop looking smug. Because if your economy is built around moving information between boxes, and somebody invents a machine that moves information between boxes faster, cheaper and without needing a wellness webinar, you may have a problem.

There is something faintly theological about it too. AI lives in the cloud, which is already suspicious language. It seems to be everywhere and nowhere. It speaks in every language, answers almost any question, draws on more digitised human knowledge than any person could ever read, and can extrapolate beyond normal human capability in some fields.

It is not God, obviously. It is not wise, moral or omniscient. It is often wrong with the smooth confidence of a senior official explaining why the computer says no. But it is the first machine that gives ordinary people the feeling of consulting something close to an oracle. The old gods spoke in thunder. This one speaks in bullet points and occasionally asks you to upgrade your subscription.

That is the unsettling bit. AI does not merely retrieve what humans already know. In bounded domains, it can go beyond us. Protein folding, materials discovery, weather modelling, logistics, code, drug screening, plasma control. These are not just library problems. They are problems of scale, prediction and search. AI can move through those spaces faster than unaided human thought can follow.

But extrapolation is not wisdom. A machine may become superhuman at solving bounded problems long before it is remotely fit to decide which problems should be solved. It may find the molecule, optimise the system, write the code or produce the strategy, while still having no real grasp of consequence, responsibility or human mess. That is the theological danger. Not that we have built God, but that we have built something powerful enough to seem divine and stupid enough to still need checking.

Which is why the best use of AI may not be treating it as a single oracle at all. The next stage is already being worked on: systems of AIs set against one another, each with a role. One drafts. One criticises. One checks facts. One tests code. One looks for risks. One searches for alternatives. One tries to break the answer before a human ever sees it.

That is not science fiction. Anyone who has used two different AI systems as sparring partners will recognise the principle. A single model gives you an answer. A second model may spot what the first missed. Feed the criticism back, and the result often improves. Industrialise that loop, automate it, give it tools, connect it to data and workflows, and you no longer have a chatbot. You have a synthetic team.

No.1 Son, who actually understands this stuff rather than just prodding it with a stick as I do, started me on this particular line of research. He tells me this is already happening. The clever bit is not one AI giving one polished answer. The clever bit is multiple agents arguing, testing, correcting, building and checking one another at machine speed. It is my little ChatGPT-versus-Gemini parlour trick, but with a workflow diagram and people’s jobs attached. His own cheerful view is that when his job goes, he will either become a farmer or run a trading operation from his bedroom, which may be the most modern career plan I have heard: half medieval peasant, half hedge fund.

That changes the productivity argument. We are not merely talking about one worker with one clever assistant. We are talking about one worker supervising several artificial assistants, critics and testers, all operating continuously. The human role moves from producing the work to setting the objective, judging the outputs and taking responsibility for what survives the argument.

It also changes the risk. A single AI can be wrong. A group of AIs can still be wrong, but more persuasively, after having apparently checked itself. So the human in the loop becomes more important, not less. Someone still has to ask whether the answer is grounded in reality, whether the assumptions are sound, and whether the machine has simply produced a beautifully formatted mistake.

This is where policy has to begin, not end. The Government should not simply ask how Britain can “lead the world in AI”, which is the sort of phrase that should immediately make everyone check for missing wallets. It should ask who owns it, who powers it, who is displaced by it, who benefits from it, who audits it, who is liable when it goes wrong, and who is left outside pressing their nose against the glass while being told to retrain.

The jobs most exposed are not, at first, the traditional practical ones. The plumber still has to turn up. The electrician still has to know which wire is live. The farmer still has to deal with weather, soil, machinery and animals, none of which are especially impressed by a beta release. The builder still has to build the wall, although preferably not in the style of a national infrastructure project.

But we should not kid ourselves that physical work is automatically safe. Robots can weld. Machines can lay, cut, lift, sort, scan, spray, inspect and assemble. AI can design the joint, check the image, plan the route and optimise the process. The clean, repetitive, predictable parts of physical work are exposed too.

So the protected work is not simply “manual labour”. That is too crude, and it sounds like something said by someone who has never had to make anything fit. The valuable bit is skilled physical judgement in messy environments. It is the person who can deal with awkward access, damaged materials, one-off repairs, bad drawings, poor weather, odd noises, hidden corrosion, tired machinery and the small but important fact that reality refuses to behave like a training dataset.

Steel, water, soil, heat and animals do not care what the strategy document says. They do not care about vision statements, quarterly objectives or whether someone has used the word “transformation” in a meeting room with frosted glass. They just sit there being awkwardly real.

Which means technical education stops being a side issue. It becomes national policy. Apprenticeships, further education, adult retraining, engineering colleges and practical routes into skilled work should not be treated as the consolation prize for people who did not go to university. They are part of the country’s survival kit. If we need people who can install, repair, maintain, fabricate, build, care, farm and engineer in real conditions, then we need to fund and respect the institutions that produce them.

The vulnerable layer is the polite, educated, spreadsheet-and-email middle. The people who process, format, review, schedule, summarise, chase, compare, draft and produce internal documents that are read by seven people and understood by none. Junior analysis, routine legal work, compliance checking, customer service scripts, HR processes, report writing and “let me pull together a deck”.

AI loves that stuff. Not because it is wise, because it is not. But much of the modern services economy is not wisdom either. It is procedure, paperwork and pattern. It is applying rules to text and moving the result onwards. That is exactly where agentic AI becomes dangerous.

So government should map the exposed jobs properly, not with a cheerful press release about “opportunities” and a stock photograph of a woman smiling at a tablet. Administrators, clerks, call-centre workers, junior analysts, paralegals, coders, routine project managers, compliance staff and all the people whose work is mostly information-processing need to be visible in policy before the redundancy letters arrive. Wage insurance, transition support, retraining grants and regional employment plans sound dull. Dull is good. Dull is what you want before people start throwing furniture.

The programming world is already getting an early look at this. The people closest to the machinery can see what is coming. A good developer with agentic tools can produce multiples of his previous output, not because he has suddenly become a genius, but because the machine is now doing the dull implementation, testing, fixing, refactoring and boilerplate that used to fill the day.

This is the warning. When bright young programmers start saying their own job, as currently defined, may have a two-year shelf life, it is probably worth listening. They are not textile workers in 1980 being patronised by politicians with retraining leaflets. They are the people building the machines that are coming for the leaflets.

The corporate response will, of course, be magnificent. First, it will be banned. Then it will be quietly used by the people who get things done. Then management will wink at it. Then there will be a policy. Then there will be a taskforce. Then someone will produce a slide saying “AI-native transformation journey”, and everyone will pretend that this was the plan all along.

Even the bosses should not feel too smug. A great deal of management is not leadership in the Churchillian sense. It is moving information upwards, instructions downwards and blame sideways. Agentic AI can do quite a lot of that without needing a reserved parking space.

If an AI can track projects, summarise progress, flag risks, assign tasks, compare performance, draft reports and remind people what they promised last Thursday, then the question becomes rather awkward. What is the manager adding? In good cases, judgement, experience, coaching, accountability and the ability to resolve human mess. In bad cases, a calendar invitation and a tone of mild disappointment.

Senior executives are not magically protected either. Market analysis, competitor scanning, financial modelling, scenario planning and board-paper drafting are all AI-friendly work. The remaining human value is choosing between imperfect options, taking responsibility, understanding people, spotting what the model has missed, and having the courage to act when there is no spreadsheet-shaped certainty.

That is why governance matters. Multi-agent AI systems will draft, check, test, criticise and act. Some will be useful. Some will be wrong in ways that look reassuringly professional. Government needs rules on audit trails, accountability, safety, procurement, data protection, liability and human responsibility. “The AI said so” must never become an acceptable defence, especially in government, healthcare, policing, finance, welfare, employment or anything else where a tidy error can ruin someone’s life.

That may be the most amusing reversal. The people who spent years asking whether workers were “adding value” may soon find a machine asking the same question about them, only faster, cheaper and without laughing at the phrase “leadership journey”.

And then comes the awkward question. If much of the services middle is exposed, and only the skilled, messy, physical work is relatively protected, who runs the world?

It may become a struggle between those who control artificial cognition and those who still command physical competence. But the numbers matter. There will not be millions of people controlling frontier AI models, chip supply chains, data-centre networks and global platforms. That world will be narrow, capital-heavy and probably rather pleased with itself.

There will, however, be millions of people keeping the physical economy alive: installing, repairing, farming, building, wiring, plumbing, maintaining, nursing, fabricating, driving, inspecting and dealing with the endless awkwardness of matter.

That gives the physical side a different kind of power. Not the concentrated power of ownership, but the dispersed power of necessity. The AI controller can scale one system across millions of users. The electrician, farmer, technician or carer cannot be scaled in quite the same way, because the work is local, embodied and annoyingly real. The pipe bursts in one house. The roof leaks on one building. The field is wet in one valley. The old person needs care in one room. Reality is not a platform, however much someone in California may wish it were.

The cloud still needs a roof. The oracle still needs electricity. The model still needs cooling. The digital economy still needs copper, concrete, steel, water, land, food, grid connections, skilled hands and people who can fix things when reality declines to reboot.

This is where ownership policy becomes unavoidable. If a handful of companies own the models, chips, data centres, platforms and access rights, they will own the tollbooths of the new economy. Competition policy, data rights, public-interest AI infrastructure, sovereign capability and taxation of extraordinary AI rents are not dull technical details. They are the difference between a productive society and a rent-collecting machine with a chatbot interface.

This is not a romantic return to the village blacksmith. The future artisan may use drones, sensors, diagnostics, software, robotics and AI. But he still has to know what happens when theory meets mud, rust, heat, fatigue, weather, wear and a client who changed the requirement yesterday but forgot to mention it.

The danger is obvious. The people who control the AI may try to own everything and rent access to everyone else. The people who keep the physical world running may discover they are more numerous and more essential than they were told, but numbers only become power when they have bargaining strength, training, ownership and political voice.

And then there are the people left behind. Because there will be people left behind. Every technological revolution produces hymns to productivity from those close enough to the money, and small, soothing phrases about “transition” for those whose job has just evaporated.

The usual answer will be retraining, because it always is. Retraining is the comforting word used by people whose own job is not currently being retrained out of existence. Some of it will work. Plenty of people will learn to use AI, move into practical trades, maintain infrastructure, care for others, manage systems or build small businesses.

But not everyone can become an AI supervisor, nuclear engineer, heat-pump installer, nurse, farmer, electrician or artisan baker with a waiting list. A serious society has to admit that. Otherwise we end up pretending that a 54-year-old accounts clerk, a redundant junior solicitor, a call-centre worker and a middle manager with a laminated leadership certificate are all just one online course away from becoming robotics technicians.

This is where universal basic income starts to look less like a student union fantasy and more like a riot-prevention strategy. That sounds brutal, because it is. If AI and automation massively increase productivity while stripping income, status and bargaining power from millions, the question will not be whether we can afford redistribution. It will be whether we can afford the alternative.

UBI would not be charity. Nor would it be a reward for idleness, as the usual suspects would immediately honk from the cheap seats. It would be a claim on the common productivity of the system: the machines, models, energy networks, public research, infrastructure, law, education and accumulated human knowledge that made the wealth possible in the first place.

Of course UBI has problems. It costs money. It could be inflationary if badly designed. It does not give people purpose, status, community or pride. A monthly payment is not a civilisation. But a civilisation that allows a small priesthood to own the oracle, while millions are told to polish their CVs for jobs that no longer exist, may find that the monthly payment was the modest option.

The alternative is not necessarily quiet poverty. It may be political rage, social fracture, extremism, sabotage, riots, or eventually something worse. Hungry people are one problem. Humiliated people with time, education, grievance and a broadband connection are quite another.

So perhaps UBI, or a citizen’s dividend, or a negative income tax, or universal basic services, becomes less a utopian scheme and more a pressure valve. Not perfect. Not sufficient. But part of the price of keeping a high-automation society governable.

That is where politics should be paying attention. Not to another “world-leading AI strategy” with a foreword by someone photographed near a server rack, but to ownership, energy, technical education, infrastructure, planning, wages, apprenticeships, redistribution and who captures the gains.

Because if we get it wrong, the future will not be clever people freed for higher work. It will be a small priesthood controlling the oracle, a practical class holding the physical world together, and millions wondering why they are poorer in a world supposedly made more productive.

Which brings us to energy. AI needs data centres. Data centres need electricity. Electrification needs more electricity. EVs, heat pumps, industry, rail, batteries, cooling systems, factories and digital infrastructure all need reliable power. The fight over AI is therefore also a fight over energy. Whoever controls the intelligence layer will still depend on whoever controls the power layer.

Renewables are essential, but they do not abolish darkness, still air or winter evenings. Solar is lovely at lunchtime in June. Less so at tea time in January when every home is drawing power, every heat pump is working hard, and a data centre somewhere is helping a marketing assistant generate six alternative ways to say “delighted to announce”.

A serious country needs firm low-carbon power. Grid upgrades, storage, demand management, interconnection and nuclear are not optional extras. They are the physical basis of an AI-heavy, electrified economy. Nuclear is not glamorous. It is slow, expensive, over-regulated and politically painful. In Britain we seem able to turn it into a twenty-year saga, spend twice the budget and then act surprised that France, China and Korea know how to pour concrete.

But if the digital economy is going to expand, nuclear stops being an ideological hobby and becomes plumbing. Boring, vital, heavy, difficult plumbing.

And that is the irony. The more ethereal the economy becomes, the more it depends on very unethereal things: copper, concrete, uranium, pylons, cooling water, skilled engineers, fabricators, land, planning consent and people who can actually fix something when it breaks.

For decades, we nudged the clever children towards screens and told the practical ones not to worry, there would always be something for them. Now the screen has learned to do screen work, and the practical world is looking rather less second-class.

The safe future may not be “learn to code”, because the machine has learned to code. The safe future may be to learn to use AI, understand systems, know where the energy comes from, and have at least one skill that involves judgement in the real world rather than just fluency in the digital one.

And even with AI, the safe approach may not be one machine giving one answer. It may be several machines disagreeing while a human with some judgement sits in the middle, listening, prodding, checking and occasionally asking whether everyone has gone mad.

Which is a slightly awkward conclusion for a country that sold off half its industrial base, neglected technical education, made housing unaffordable, tangled infrastructure in planning treacle, and then proudly announced that its future was services.

The required policy is not mysterious. Own some capability. Tax some rents. Build the grid. Build firm low-carbon power. Fund technical education. Protect displaced workers. Share productivity gains. Regulate high-risk AI properly. Keep humans accountable. Stop pretending that innovation policy is a glossy brochure with a launch event and some lanyards.

We have built something that looks a bit like an oracle, runs on electricity, feeds on human knowledge, predicts beyond human reach, and still cannot be trusted to know when it is talking rubbish.

The question is not whether Britain can write another strategy about it. Of course it can. Britain can write a strategy about a kettle.

The question is whether we still know how to build, power, own and govern the real machinery underneath the miracle, and whether we have the wit to share the gains before the people left outside decide to stop asking politely.

Otherwise we may find ourselves living in the cloud, paying rent to the oracle, and waiting for someone practical to come round and fix the lights.