A year inside the AI machine

What I'm actually seeing - and the questions I think matter

A Karteria Partners perspective - drawn from more than a year using AI in earnest, some of it before this firm existed, all of it useful. I don't claim to be right about most of what follows. I do think these are the right questions. It's about where AI is heading, seen through the lenses we work in: sustainability, energy, real estate and infrastructure, and professional services.

For more than a year I've used AI properly - not demoing it, running real work through it daily: building custom AI assistants in the early days, working inside the models, training teams before I went independent, and now running a business with AI in the operating core rather than bolted on the side. At Karteria Partners we exist to help institutions navigate uncertainty - so I've made it my business to sit inside the biggest source of it.

One observation sets the pace. A year ago, my stack was scattered - one tool for writing, another for analysis, another for design. It has converged hard. A single platform now absorbs most of it - the writing, the analysis, the agentic work, the design, the long multi-step tasks - and it's reaching into the enterprise software most of us live in daily. There's a split worth holding, though: the tool that serves the corporation isn't the same as the one that serves the individual. It's where this piece ends up.

My vantage point is deliberately odd. I've been outside a large corporate for most of the past 12 months - free to use what I want, and freer still to think, rather than deliver whatever a typical corporate engine demands. And I've had to write my own AI governance, personal and professional, from a blank page. Hands deep in the tools, head out of the day-job. It's a rare place to watch from - so here's what I see.

Who gets disrupted - and in what order

Most “AI will take your job” views fail because they treat it as one event. It isn't. It's a sequence - and the order is legible if you stop asking what's “skilled” and start asking what actually slows automation down. Four brakes set the pace: is the work physical or digital; must someone be accountable when it's wrong; how costly is a mistake; and does the value depend on a human relationship. Stack them, and a running order falls out.

1. Digital, low-stakes - happening now. First-draft content, basic code, tier-one support, routine research. No brakes apply. This isn't coming; it's here.

2. Digital, but liability-gated - imminent. First-pass legal drafting, basic tax, parts of audit, translation. The capability exists; the only brake is who's accountable.

3. Physical, but structured - near. Driving is almost a single task in a mapped world, which is why it falls earlier than jobs that look “lower-skilled.” The brake is regulation and rollout, not capability.

4. Physical, but messy - later. Trades, care, nursing. A machine that can work a stranger's cluttered house is far harder than one on a mapped road. Hands-on, variable work is more protected than much white-collar work - not less. And there's a cultural journey to travel before we let an unknown machine into our homes unsupervised.

5. High-stakes, physical and relational - latest.. A GP isn't one task. AI may already win the diagnosis - but the job is also examination, procedure, accountability, and the trust that makes someone disclose and comply. That bundle trips every brake at once.

6. The pattern, in a line: digital before physical, structured before messy, task before relationship. The axis isn't skilled versus unskilled - that's the instinct to drop.

   
   

Who survives the cut - and who quietly disappears

Run the same logic at the level of the business, not the job, and a sharper question appears. If AI commoditises raw capability, what's left to defend a company? Three answers - and they sort the winners from the casualties.

The survivors fall into two camps. The first wins on community and network - the businesses whose value is the people already inside them. A taxi network, a fitness community, a marketplace: the product is replicable, but the trusted network isn't, because everyone is there because everyone else is. You can clone the app in a weekend; you can't clone the millions of users. The second wins on enterprise trust, brand and support - the mission-critical systems where reliability, security, compliance and someone-to-call are the whole point. Nobody rips out the system of record to save a licence fee. These businesses don't just survive AI; it makes them stronger, because it raises the volume of work flowing through the systems they own.

Then the exposed camp. The businesses whose moat is simply a feature - a capability that used to be hard and now isn't. A document tool, an image editor, a single clever function sold as a subscription or even a basic advisory service. No community holding users in, no mission-critical trust, no complex support anyone depends on. When the capability becomes near-free to reproduce, the rent it earned collapses - and the more profitable and single-purpose the feature, the more inviting the target. Some of the most familiar names in software are, underneath, a feature with a price tag. They will not all be here in their current form.

That's the real divide for anyone weighing where value sits: not big versus small, or old versus new, but network or trust versus feature. The first two compound. The third is on borrowed time.

The rebound - and why energy is the one certainty

My most confident claim comes from an old idea, not a new one.

In 1865 William Stanley Jevons noticed that making steam engines more efficient with coal didn't cut coal use - it raised it. Cheaper, more useful coal found more uses. Efficiency at the unit level drove consumption up at the system level. The rebound effect - and AI is about to demonstrate it at scale.

When a query is nearly free, you don't ask fewer, better questions - you ask ten speculative ones. When standing up software costs almost nothing, you spawn disposable tools you'd never have justified. When capture costs nothing, you record everything, just in case - the doorbell cameras, the meetings, the calls, the photos, the messages, the LLM chats - and data grows exponentially. We've seen the pattern in miniature: nobody compresses a file to fit a floppy disk anymore. The moment the constraint vanished, the discipline it imposed vanished too, because optimising now costs more than the waste it saves. AI strips friction from cognition the way unlimited broadband and near-free storage stripped it from data. The restraint goes with it.

And agents multiply it: one human request becomes hundreds of machine steps - an agent doesn't ask a question, it runs a campaign of them, spawning its own tools and checking its own work. The rebound, with a multiplier on it.

And all of it - the queries, the disposable software, the junk (and useful) data - runs on electricity. The International Energy Agency expects data-centre demand to more than double, from around 415 TWh in 2024 to roughly 945 TWh by 2030, more than Japan's entire annual consumption, with AI the biggest driver. In the US, data centres are set to account for nearly half of all electricity-demand growth this decade.[1] Closer to home it's starker still: the queue of new demand waiting to connect to Britain's grid tripled in just seven months, from 41 GW to 125 GW - against peak national demand of about 45 GW - with data centres the biggest driver[2]. We have queued up nearly three times the nation's peak power needs, waiting on a grid that can't yet carry them.

So the certainty isn't that AI saves energy. It's the opposite. The rebound drives a structural surge in demand - and with it nuclear, solar, wind, hyperscale build-out, cooling, water and grid. This is one of the defining infrastructure stories of the decade, and it sits squarely where Karteria Partners works.

One twist I keep turning over: as the cost of all this becomes visible - in money, energy and carbon - people may, hopefully, start to ration what they ask. Tokens, oddly, might re-teach a discipline that abundance erased. Abundance breeding its own new scarcity.

The question underneath: does demand have a ceiling?

There's a deeper worry I can't fully answer. The bull case for all this assumes AI cuts costs and expands output forever. But you can only own so many phones, drive so many cars, have so many medical operations. Once a market is saturated, demand stops growing with capability and grows only with the replacement cycle plus population - and no amount of cheaper, faster production lifts that ceiling. Make phones twice as fast and you don't sell twice as many.

If that's right across enough markets, AI races to the ceiling in each one and then stops - which looks less like endless growth and more like an economy bumping against its own limits. The honest counter is that every previous plateau looked like a ceiling from the inside, too. Nobody could want a smartphone in 1980 or a ride-hailing app in 1990 - the category didn't exist yet. Growth resumed because new wants were invented, not because old markets un-saturated. So the real fork is this: does AI mostly open new categories of want, as every general-purpose technology before it did - or is it the first that mostly satisfies existing ones? I lean towards the former, on 250 years of precedent. But I hold it loosely, because the mechanism that would make this time different - a technology aimed at cognition itself - is at least coherent.

What stays scarce - and the divide nobody's naming

If cheap cognition floods the world with analysis, analysis stops being scarce. Judgement, trust and accountability become the scarce things instead. In the corporate world what sticks is community and the enterprise trust-and-support function. On the personal level it's lived experience, accountability and relationship - navigating the grey areas you can't synthesise.

The creative markets are the early test of that scarcity. As AI-generated images, text and music become near-free, the market is splitting on provenance rather than output: AI-made work finds a real but lower-priced tier, while a painting, a story, a hand-drawn card holds its value precisely because a person - and a story - stands behind it. Authenticity becomes the scarce good, and the logic that prices a trusted brand's signed-off advice above an algorithm's is the same logic that holds the human-made canvas above the generated one.

Which surfaces a divide I think matters enormously and almost nobody is naming: corporate-augmented AI versus individual-augmented AI. Corporate AI amplifies the organisation - its processes, its averaged best practice, its compliance. It builds continuity, and tends to make the individual seat more interchangeable: everyone gets the same copilot behind the same security.

Individual AI - your “second brain” - amplifies the person: your taste, your judgement, your memory, what makes you uniquely good. It makes you more singular, not less. These are not the same product pointed at different data.

And here's the kicker. The individual layer takes years to build and can't be bought, transferred or shortcut. The model is the cheap, commoditised part - replaced every few months if needed. The accumulated context - your context - is the scarce part, and it compounds with use. While AI commoditises everyone's raw capability, the person who's spent two years training their own augmentation holds the one asset a newcomer can't copy. That, I think, is the real hedge against your own obsolescence.

But it raises an unsolved problem - and it's the one I'd most like someone to crack. How does your personally-trained AI ever see your work - sit beside enterprise systems the way a human colleague under an NDA does - without the corporate engine ingesting your context, or your personal AI extracting theirs? Look but don't extract. Plug in, then unplug. A permeable membrane for insight and an impermeable one for data. Humans do this instinctively - we sit with confidential information all day and know not to repeat it down the pub; with machines that discretion has to become architecture. I don't have the answer. It's a question someone needs to answer soon.

 

The insurance hinge

Here's a specific moment worth watching, because it unlocks much of the above. What protects professional fees today isn't only skill - it's liability. A human (or company) signs off, carries indemnity, and is accountable when it's wrong. The professional-indemnity moat does quiet, heavy work.

And it's already under strain. There have been clear examples across consulting and professional services of firms making significant headcount reductions and trimming graduate intake; and of listed advisory and information businesses seeing share prices fall anywhere from the mid-teens to more than half from recent highs, with analysts pointing squarely at AI[3]. These are the firms whose model was to sell expertise by the hour - exactly the model AI compresses.

So the question is what happens when AI advice becomes insurable to a credible standard. When the cover exists, the liability moat dissolves - and the routine, high-volume work behind it goes first.

But notice that insurance has its own running order. Low-stakes advice - where a claim might run to a few hundred thousand - is easy to underwrite against a model, needs no brand and no balance sheet behind it, and so it goes first and cheapest. At the other end sits the advice where a mistake means lives or tens of millions - signing off a bridge, a building, a safety-critical design. That work demands deep liability cover, a name people trust, and a human willing to put their signature and their reputation on the line. It will stay human-led the longest, perhaps indefinitely - not because AI can't do the analysis, but because the trust and the insurance are the product.

That's the professional-services echo of the survivors from earlier. A trusted brand carrying serious indemnity is the advisory equivalent of the mission-critical system nobody rips out - the value isn't only the answer, it's the someone-who-stands-behind-it. The professions can feel it coming. I don't know when the hinge fully turns, or how cleanly. But it's a hinge, not a footnote - the kind of thing that reshapes an industry quietly, then suddenly.

If the human's role shifts from producing the advice to guaranteeing it, then the shape of professional services has to change with it - advice unbundled into tiers, the routine delivered cheaply and at scale, the human reserved for the judgement and the accountability that can't be delegated. What that looks like in practice - how it's priced, packaged and trusted - is one of the more interesting questions in front of the profession. We have views. But that's a conversation for another day.

The graduate problem

This is the one I genuinely can't answer - but I think we have a duty to.

If AI already out-performs the inexperienced graduate at the tasks a graduate cut their teeth on - how does anyone become the experienced expert? The accountable professional, with the relationships and the lived judgement, was forged by years of doing the work AI now does instantly. Pull out the bottom rungs and you save cost today while starving the profession of tomorrow's experts.

Maybe the answer is that firms carry early-career talent deliberately, as an overhead - an investment in producing future judgement rather than immediate output. Maybe it reshapes the route in entirely: a rise in apprenticeships built around learning beside an expert, and a corresponding decline in the traditional degree that front-loads knowledge AI now supplies on demand. Either way it's a real cost and a real cultural shift, and I'm not sure the market chooses it voluntarily. But if we don't, we hollow out the pipeline that produces the very people the whole system still depends on.

Why I think this is ours to hold

There's an ethics weight under all of this that can't be an afterthought - bias, bad actors, accountability, who's harmed and who decides. And there's a physical weight too: the energy and water behind the compute are real, and growing. Standing back, AI looks to me like a climate-scale phenomenon - systemic, civilisational, decades in its consequences, and not without a footprint of its own. But where climate change is overwhelmingly a threat we must mitigate and adapt to, AI is more genuinely two-sided: a systemic risk and a systemic opportunity at once, with no single clean answer, demanding judgement under deep uncertainty.

Which is why I think those of us who've spent careers walking the sustainability tightropes are unusually well-placed to have a view. Sustainability has never been one issue - it's a balance held across climate and nature, energy and water, waste and resource, social impact and just transition, treating each as both a serious threat to mitigate and a generational opportunity in the response, without flinching from either. We know what it means to weigh build-out against externality, growth against limits, speed against fairness. Those are the same muscles AI demands - and it's why, at Karteria, we've leaned into this rather than waited for it to arrive.

So I'll end not with an answer but an invitation. What do you think? Where does this match what you're seeing, and where does it miss? I'm genuinely asking - because the firms and the people who navigate this well will be the ones who started comparing notes early.

Robbie Epsom is Co-Founder of Karteria Partners.

1. International Energy Agency, Energy and AI (special report, 2025): global data-centre electricity demand is projected to more than double from c.415 TWh (2024) to c.945 TWh by 2030 - slightly more than Japan's total annual consumption - with AI the most important driver. In the US, data centres account for nearly half of electricity-demand growth to 2030 and, by the end of the decade, are set to consume more electricity than the production of aluminium, steel, cement, chemicals and all other energy-intensive goods combined. iea.org/reports/energy-and-ai.

2. Ofgem, Demand Connections Reform Call for Input (13 February 2026): contracted demand connections grew from c.41 GW (November 2024) to c.125 GW (June 2025) - against GB peak electricity demand of c.45 GW - driven by a surge in data-centre applications, which account for c.50 GW of the queue: ofgem.gov.uk (PDF). For context on the generation side: the pre-reform connections queue had exceeded 700 GW - around four times 2030 needs, though largely speculative generation and storage projects - before NESO's December 2025 Connections Reform re-ordered it into a 283 GW delivery pipeline: neso.energy.

3. Illustrative of a wider 2024–26 pattern (firms not named to avoid singling out individual businesses). Headcount and graduate intake: several global consulting and professional-services firms have made significant headcount reductions and trimmed graduate recruitment, as widely reported in the financial press and reflected in company filings. Share prices: listed advisory, IT-services and information businesses have fallen from roughly the mid-teens to more than 50% from recent highs - e.g. one of the largest listed consulting firms fell c.18–20% in a single session in June 2026 (its worst on record) and over 50% from its 52-week high, with Bloomberg Intelligence attributing the move to AI disrupting demand across consulting and managed services. Professional sentiment: Thomson Reuters Institute, 2026 - the share of lawyers viewing AI as a major threat to the unauthorised practice of law rose to c.50% (2026) from c.36% (2025). Direct sources (which name the firms): Bloomberg, 18 June 2026: bloomberg.com; The Motley Fool, 18 June 2026: fool.com; Thomson Reuters Institute, 2026 AI in Professional Services Report: thomsonreuters.com.