The general narrative in enterprise technology today is that AI replaces people. I want to put the opposite hypothesis on the table. For the rest of this decade, many of the organisations that adopt AI most aggressively will need more people, not fewer, because of what happens to a system when you make one part of it suddenly, dramatically cheap.
That is deliberately provocative, and I am offering it as a hypothesis rather than a prediction based on the jagged capabilities of the tech.
The pendulum has swung
The pendulum has been mostly on the side of tech mania/FOMO-driven adoption: any function not visibly “doing AI” is behind. Anthropic’s chief executive has suggested AI could remove half of entry-level white-collar jobs within one to five years. That is the high end of the range, and it sets the mood in a lot of boardrooms.
Pendulums swing back. What pushes this one back towards the middle is the success of AI colliding with what the hype skips over.
Three things AI cannot do (or do cheaply)
One: the robots are not here, and when they arrive they will be expensive. Most of the work humans do is still physical. As of early 2026, enterprise humanoid robots sell for roughly 50,000 to 320,000 US dollars per unit, and the pilots at BMW, Hyundai and Amazon remain narrow. Tesla’s sub-30,000-dollar target depends on mass production that does not yet exist. Until that changes, manual and care work is protected by unit economics. The WEF Future of Jobs Report 2025 and the US Bureau of Labor Statistics projections both expect absolute growth in farming, construction, delivery, care and the renewables trades to 2030.
Two: AI cannot hold judgement, accountability or originality.
Judgement is the ability to weigh a situation the training data doesn’t see as it lacks the non-language-based, real-world context (more on this in Artificial intelligence is not artificial wisdom). Accountability is the capacity to be answerable: a model cannot be a defendant, cannot lose a licence, cannot be hauled in front of a regulator. Pretend otherwise and you get the moral crumple zone: when a model is in the loop, liability collapses onto the nearest human, usually the one with the least power to have prevented the outcome. And originality, in the sense of work that does not regress to the average, is something current models are measurably bad at. Recent research on creative homogeneity across large language models finds their outputs are far more similar to each other than human outputs are to each other, and that tweaking the prompt or the temperature does not close the gap.
There is a deeper irony here, and it is not new. As I set out in The AI Efficiency Fallacy, the better the automation, the more the supervising human’s skill atrophies, and the harder the rare moment of intervention becomes. Capable AI raises the value of skilled human oversight while making it scarcer. That is the confirmation fatigue and relevance degradation risk from the original sleepwalking piece, seen from the labour side.
Three: AI itself is not free. This is the constraint most often waved away. The big four hyperscalers are on course to spend in the region of 725 billion US dollars on capital expenditure in 2026. The IEA’s work on energy and AI projects global data-centre electricity demand roughly doubling by 2030. Inference costs per token have fallen dramatically, which is real and important, but the infrastructure that makes cheap inference possible remains expensive, concentrated and increasingly constrained by power availability rather than chip supply. An organisation betting its operating model on infinite cheap compute is making an assumption. I go into why the brute-force, centralised approach has material limits in The AI Efficiency Fallacy.
Jevons Paradox
When you make a task dramatically cheaper, you usually get more of it, not the same amount done by fewer people.
Economists call the energy version of this the Jevons paradox. Jevons himself wrote in 1865 that “the very labourers whose labour is saved find their more efficient labour more demanded than before.” The cleaner modern parable is the cash machine. As ATMs spread from the 1970s onward, the number of bank tellers in the United States rose, because cheaper branches meant banks opened more of them. The teller’s job changed, but for decades the headcount grew.
Apply that to knowledge work. If an agentic system can draft ten times the code, ten times the legal memos or ten times the marketing variants, someone still has to verify, choose, approve and own each output, and that verification work scales with volume. Cheaper analysis means more decisions get analysed, so the market for the underlying service expands. And the rare, high-skill human intervention becomes more valuable per hour even as routine roles thin out.
This is the controlled versus agentic architecture spectrum viewed from the staffing angle. Augmentation architectures keep humans in the loop by design and tend to create verification and judgement work. Full automation architectures try to remove it, and frequently just relocate it downstream, where nobody planned for it.
A tiered view of exposure
Synthesising the ILO 2025 update, the OECD, the BLS, the WEF and Goldman Sachs, here is a tier model that helps with strategy conversations. The percentages are task-exposure ranges for 2026 to 2030, not predictions of net job loss. “New-role likelihood” is a qualitative judgement based on historical precedent and current evidence.
| Tier | Work type | Task exposure 2026-2030 | New-role likelihood | Note |
|---|---|---|---|---|
| 1 | Physical and care work (farming, construction, delivery, home and personal care) | Low, under 5-10% | High | Protected by robot unit economics; demographic ageing lifts care demand |
| 2 | Skilled physical trades (electricians, HVAC, machine maintenance, wind and solar technicians) | Low, 5-15% | High | Fastest growth is in renewables trades; site-specific dexterity resists automation |
| 3 | Routine knowledge work (clerical, office admin, tier-1 customer service, basic bookkeeping) | High, 30-50% | Low to moderate | The clearest net contraction; BLS expects AI to reduce office and admin employment |
| 4 | Skilled analytical and production knowledge work (junior consultants, paralegals, routine software, marketing execution) | Medium-high, 30-50% at task level, closer to augmentation at role level | Moderate | The entry ladder thins; the mid-tier expands |
| 5 | Complex, judgement-intensive work (senior strategy, M&A, regulated advisory, clinical diagnosis, enterprise and security architecture) | Low displacement, high augmentation | Moderate to high | Oversight becomes more valuable as automation improves; accountability protects regulated work |
| 6 | Creative origination (R&D, novel design, original research, brand-defining work) | Tools widely used; originating work less displaced | Moderate | Homogenisation risk if AI is used uncritically; genuine originality gains a premium |
| 7 | Accountability-bearing roles (directors, fiduciaries, regulated approvers, safety-case owners) | Very low displacement; decision volume likely to rise | Possibly growing | AI cannot be a defendant; volume increases the load on the human signatory |
The pattern is a barbell. Tiers 1, 2, 5, 6 and 7 are relatively protected or even growing. Tier 3 contracts. Tier 4 does not disappear but is hollowed in the middle, and that hollowing is where the social and organisational pain concentrates.
Buy side and sell side
Consulting is the leading indicator, because it is a pyramid made of exactly the work AI does first.
Sell side. The classic model bills leveraged junior hours: analysts research, build decks, run models. AI does that work now. McKinsey’s internal assistant is used by most of the firm monthly; graduate intakes across the big firms have been cut and starting salaries frozen. The pyramid is becoming what some observers call an obelisk: a thinner base, a fatter expert middle, more equity at the top, and pricing shifting from time billed to outcomes delivered. The risk to the big firms is margin compression and disintermediation by clients who bring the capability in-house.
Buy side. AI does not reduce a client’s need for the things consultants were really selling: independent challenge, fiduciary cover, cross-client pattern recognition, surge capacity, and someone external to be accountable. It does reduce the willingness to pay for slide production and benchmarking that a competent internal team can now do itself. The buy-side question shifts from “how many bodies” to “what specifically am I buying that I cannot create internally”. Tellingly, the major AI vendors are signing alliances with the big consultancies, because deploying this technology at enterprise scale still needs a great many humans on the ground.
In-house knowledge workers. Expect the same diamond shape to spread well beyond consulting: a thinner entry layer, a fatter middle, and rising demand for AI-literate mid-career judgement. Organisations that use AI well will tend to do more things, more products, more analyses, more variants, more compliance checks, and will need more people at the human-critical steps of those pipelines.
Three new ways to sleepwalk
The original eight sleepwalking risks were largely societal and ecosystem-facing. Looking at the same problem through the labour lens surfaces three more, all of them organisational, and all of them easy to walk into with eyes apparently open.
Are we destroying the experience pipeline? If you automate the entry rung, you stop manufacturing senior judgement. A senior lawyer’s instinct was built on years of discovery grunt work. A principal engineer’s intuition came from fixing junior-grade bugs. Automate the bottom of the pyramid for this year’s efficiency saving and you have defunded your own succession pipeline. The cost is invisible on a spreadsheet and severe over a decade. This is relevance degradation operating across careers rather than within a single task. I look at the training side of it in How do we train people in the future?
False savings? This is the trap of cutting headcount in anticipation of AI savings, then being overwhelmed at the human chokepoints by the volume AI itself generates. You removed the reviewers, but agentic systems are now producing ten times the output that needs reviewing, approving and owning. It is asymmetrical overload, except the overload is self-inflicted. Slowing the agent to match review capacity is one architectural response.
Defaulting to average output. Because model outputs regress towards the statistical average, an organisation where everyone leans on AI for everything will drift, gradually and unnoticed, towards sameness. Each individual artefact looks perfectly competent. The aggregate effect is the slow erosion of whatever made the organisation distinctive.
What the customer sees
All three of these risks are internal, but they do not stay there. Put them together and the customer gets a worse product from a company that believes it has become more efficient.
Cory Doctorow coined enshittification for platforms that get steadily worse for their users while value is extracted from them. I am using it more loosely here, for the slide in quality that follows when AI is relied on in places where a person used to notice things. The service still answers, but with something plausible and generic, and the edge cases that used to be someone’s job are now nobody’s.
Klarna is a classic example. After saying its AI assistant was doing the work of 700 customer service agents, its chief executive told Bloomberg in 2025 that cost had been too dominant a factor, that the result was lower quality, and that the company was hiring people again. Customers say much the same when asked. In a Gartner survey of 5,728 customers, 64 per cent said they would prefer companies not to use AI in customer service, and 53 per cent said they would consider switching to a competitor over it. That is what people told an analyst firm rather than what they did, so I would treat the numbers with care. And when the output is wrong, the brand still owns it: a Canadian tribunal held Air Canada liable for what its website chatbot told a customer, after the airline argued the chatbot was responsible for its own actions.
Quality and reputation are slow to build and slow to show up in a business case, which is why this cost gets missed. The saving from removing people shows up in this year’s numbers. The customers who leave, and the reputation for being hard to deal with, show up later.
Things that could prove this wrong
Pace could surprise in either direction. Robotics could move faster than the unit economics suggest; if a credible humanoid drops below roughly 25,000 dollars with reliable two-year deployment, the Tier 1 protection argument weakens for some manufacturing and logistics work. The macro productivity picture is contested: Goldman’s 7-per-cent-of-GDP estimate and Daron Acemoglu’s task-based model, which puts total factor productivity gains at well under 1 per cent over a decade, cannot both be right (the productivity fallacy has the sources). The current employment data is noisy and only partly about AI. And the induced-demand argument depends on demand for the underlying services being elastic; for most professional work it historically has been, but that is an assumption.
The ATM story also carries a lesson: bank-teller numbers grew for decades, and then mobile banking killed the role anyway. A first wave can grow a role for years before a second, complementary wave removes it. The 2026 to 2030 window I am describing may be the grow phase. The kill phase for some of these roles could simply arrive in the 2030s.
What this means
If you are an architect or technology leader, map your operating model against the tiers, and be honest about which of your own functions sit in the hollowing Tier 4. For every AI deployment, name the human signatory and budget properly for their training and tooling, because that is your moral crumple zone whether you have planned it or not (the reference architecture covers how to design that escalation). And before you sign off a business case built on headcount reduction, model the opposite scenario alongside it: the one where AI volume means you need more people at the points only people are capable of. The alternative is a serious risk of enshittification.
If your AI business case assumes you will need fewer people, have you stress-tested it against the false savings scenario? And looking at your own organisation, which tier is most at risk of losing its experience pipeline, and what, concretely, are you doing about it before the rung is gone?
Sources are linked inline. This piece builds on Are we sleepwalking into tomorrow’s AI challenges? and connects to the cone of possibilities and the controlled-versus-agentic architecture spectrum explored in earlier Architect Tomorrow work. The AI Efficiency Fallacy takes the oversight argument further, with the sources for Bainbridge’s ironies of automation and Elish’s moral crumple zone.
AI use: AI augmented this piece rather than replacing the author. Claude helped with research synthesis and drafting; the ideas, framing, judgement and any errors are mine.