Governance Without Pilots: The Race to Regulate AI Agents We Haven’t Taught Anyone to Supervise
A research journal on why the agentic cloud ends the age of the operator, what the world’s governments have actually written down about AI governance, and the one thing almost none of them has yet legislated: the human who signs off.
For thirty-five years, the question every workforce programme asked was simple. Can you operate the technology? Can you configure the server, pass the certification, run the pipeline?
That question is ending. In the last eighteen months the cloud stopped being a place where we store and compute. It became a place where AI agents plan, decide and act on our behalf. When the system can do the work, the scarce skill is no longer doing the work. It is directing it, checking it, and being accountable for it.
This piece makes three claims. First, agentic AI changes the unit of work from the task to the loop, and that quietly rewrites every cloud, security and operations role. Second, the labour market is already showing the early signature of what I call the Frozen Workforce: people certified for tasks being automated, never taught to direct the systems doing the automating. Third, and most uncomfortable, the governance frameworks now published by the world’s leading economies almost all assume a competent, accountable human in the loop, yet very few of them say where that human is supposed to come from.
Governance is no longer a compliance team’s job. It is a workforce skill. And right now, nobody is training for it at scale.
I. Prompts, Loops, Loop Governance: the new anatomy of cloud work
The unit of work has changed from the task to the loop, and that single change rewrites the job description of everyone who touches the cloud.
I describe the journey in three steps. Prompts (2023 to 2024): the frontier skill was writing a good instruction and judging a single answer. Loops (2025 to now): an agent takes a goal, decomposes it, calls tools, observes the result, critiques itself and tries again, often with no human in between. Loop governance (the decade ahead): the scarce capability is designing the loop, bounding it, and owning what comes out of it.
In engineering terms, a production agent on a cloud platform is a control system, not a chatbot. Strip away the marketing and every serious deployment has the same seven moving parts:
- Goal and success criteria. A specification precise enough that a machine can tell when it is done, and a human can tell when it is wrong.
- Planner. The model that decomposes the goal into steps, re-plans on failure, and decides when to stop.
- Tools and scopes. APIs, consoles, infrastructure-as-code, ticketing systems, each with a permission boundary. This is where an agent stops talking and starts acting.
- Identity. A distinct, non-human identity for each agent, tied to a supervising human or department, so that every action is attributable.
- Budgets and brakes. Spend limits, rate limits, step limits, and a kill switch that actually works under load.
- Checkpoints. Defined moments where a human must approve before the loop continues, calibrated to how reversible the next action is.
- Evidence. Tamper-evident logs of what the agent saw, decided and did, sufficient to reconstruct any outcome after the fact.
Notice that only one of those seven is “the AI”. The other six are governance, and they are being built today by cloud engineers, security analysts and operations leads who were trained to do something else entirely.
That is why the roles move together:
| Role today | Role in the agentic cloud | The new core question |
|---|---|---|
| Cloud administrator | Agent supervisor | What is this agent allowed to change, and how will I know it went wrong? |
| Security analyst | Authority designer | Which identities, tools and data can each agent touch, and under whose name? |
| Operations lead | Loop owner | Where are the checkpoints, and who signs off on the result? |
| Developer | Specification writer and verifier | Is the goal precise enough to pursue, and is the output actually correct? |
| Risk and compliance | Evidence architect | Can we reconstruct every consequential action from the logs? |
Regulators have started to encode exactly this anatomy. China’s agent rules, enforceable since 15 July 2026, reportedly sort agent decisions into three tiers: routine actions the agent may take alone, significant but reversible actions that need prior human approval, and high-stakes, hard-to-reverse actions that must escalate to a human outright. Singapore’s agentic framework asks organisations to bound an agent’s powers upfront and place human approval at critical points. Different political systems, same engineering conclusion: govern the authority of the loop, not just the content of the output.
The hard part is that a checkpoint is only as good as the person standing at it. Both Singapore’s framework and the 2026 International AI Safety Report flag the same failure mode: automation bias, where the human approves because the machine is usually right. A human in the loop who cannot tell when the loop is wrong is not oversight. It is theatre with a signature.
II. The Frozen Workforce: what the data already shows
The early signature is visible, and it is not mass unemployment. It is a freeze at the bottom of the ladder, which is harder to see and slower to fix.
A Frozen Workforce is not unemployed. It is stuck: certified for tasks being automated, never taught to direct the systems doing the automating. Four bodies of evidence point the same way.
The canaries are falling silent. Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, using ADP payroll records, found that early-career workers aged 22 to 25 in the most AI-exposed occupations suffered a relative employment decline of 13 per cent in their August 2025 paper (16 per cent in a later revision), even after controlling for firm-level shocks. Experienced workers in the same occupations held steady or grew. The losses concentrated where AI automates rather than augments.
The freeze is deepening, not thawing. The Stanford and ADP Canaries Dashboard, launched in June 2026 across 4.6 million workers and more than 730 occupations, shows employment for 22 to 25 year olds in the most exposed roles shrinking 3.8 per cent a year as of April 2026, against 2 per cent growth in the least exposed roles. Mid-career workers aged 31 to 34 are now contracting too, by 1.7 per cent. In aggregate, the gap between exposed and unexposed occupations remains modest. That is precisely why the problem is invisible in headline statistics.
The skills half-life is collapsing. The World Economic Forum’s Future of Jobs Report 2025 projects that employers expect 39 per cent of core skills to change by 2030, with 170 million jobs created and 92 million displaced. Net positive, yes. But net figures hide who crosses the bridge and who is left on the wrong bank.
Deployment is outrunning readiness. Gartner’s 2026 CIO survey found only 17 per cent of organisations had deployed AI agents, but more than 60 per cent expect to within two years, the steepest adoption curve it measured. KPMG’s Q2 2026 pulse found 53 per cent of large US companies had deployed agents, and the share orchestrating multiple agents across workflows doubled in a quarter, from 9 to 18 per cent.
The International AI Safety Report 2026, chaired by Yoshua Bengio, strikes the honest balance: no significant aggregate employment effect yet, agents still complement more than replace, but emerging pressure on junior workers in exposed occupations.
Here is the paradox that should keep workforce ministers awake. The judgement we now need from supervisors was built, for generations, by doing the very entry-level work that agents are absorbing. Junior analysts learned what a wrong number looks like by producing hundreds of them. Junior engineers learned what breaks production by breaking it. We are automating the apprenticeship that manufactures the people who are supposed to supervise the automation.
If we do nothing, the Frozen Workforce has two layers. Today’s mid-career operators, certified for yesterday’s tasks. And tomorrow’s missing supervisors, who never got the first job where judgement is learned.
III. From Certification to Judgement: the four-layer capability stack
Technical cloud certifications still matter, but on their own they now describe the part of the job AI is getting best at. What employers across Asia Pacific ask me for sits on top of that foundation, in four layers.
| Layer | What it means in practice | How to assess it (not a multiple-choice exam) | Failure mode if missing |
|---|---|---|---|
| 1. AI fluency | Knowing what these systems can and cannot do, and where they fail, in your own role | Predict where an agent will fail on a live task, then run it and compare | Over-trust or blanket refusal |
| 2. Problem framing | Turning a messy business need into a goal an agent can pursue, with explicit success criteria and stop conditions | Write a specification; a second agent attempts it; score ambiguity and rework | Agents optimise the wrong thing, fast |
| 3. Verification and judgement | Spotting outputs that are wrong, biased or unsafe, and having the standing to stop them | Seeded-fault drills: outputs with planted errors, timed, with the right to halt | Automation bias; the rubber-stamp checkpoint |
| 4. Governance by design | Permissions, identities, audit trails, data boundaries and accountability built into the workflow | Design an agent’s authority map, then reconstruct an incident from its logs | Trust bolted on after the breach |
The third layer deserves emphasis. In an agentic world, the most valuable person in the room is often the one who says “no, not yet”. That is a technical skill and a cultural one. It requires knowing enough to see the error, and enough organisational standing to act on it without being overruled by a dashboard.
The fourth layer is where my own work has landed. When I co-authored Singapore’s agentic AI governance framework with IMDA, the lesson that stayed with me was this: governance is not a compliance team’s job any more. It is a workforce skill. If every agent needs a human who is accountable for it, then accountability has to be taught, practised and assessed like any other technical competency.
I apply a simple test to any curriculum. Does it teach people to do the task, or to direct, check and own the task? We need both. Today we are still heavily overweight on the first.
And I apply a second test to any governance framework, which I call the enforceability test. If a rule assumes a capable human in the loop, ask who trained that human, to what standard, and how anyone would know. A rule whose human cannot be found is not a control. It is a hope.
IV. Nine Governments, Nine Theories of Trust
Nine economies now have a recognisable AI governance posture, and each one encodes a different theory of where trust comes from. Read them side by side, as of 1 October 2026, and the differences are less interesting than the one assumption they share.
1. Singapore: accountable deployment
IMDA’s Model AI Governance Framework for Agentic AI, launched at Davos on 22 January 2026, was the first government framework written specifically for agents. It is voluntary and organised around four dimensions: bound the risks upfront, make humans meaningfully accountable, apply technical controls, and enable end-user responsibility. Version 1.5 followed on 20 May 2026 with guidance on multi-agent systemic risk. It is the only framework in this set that explicitly names training as a control against automation bias, and asks organisations to audit whether their human oversight actually works.
2. European Union: rights through product safety
The AI Act remains the world’s most comprehensive horizontal law, but 2026 rewrote its clock. The Digital Omnibus on AI, in force since 27 July 2026, moved stand-alone high-risk (Annex III) obligations, including employment and education uses, from 2 August 2026 to 2 December 2027, and product-embedded systems to 2 August 2028. Transparency duties under Article 50 still applied from 2 August 2026. The quieter change matters more for this essay: Article 4 on AI literacy was rewritten. Providers and deployers no longer have to ensure a sufficient level of AI literacy; they must take measures to support it, with no guaranteed level for any individual, while the Commission and Member States pick up a support duty.
3. United States: innovation-first federalism
Washington’s posture is the 2025 AI Action Plan, the December 2025 executive order targeting state AI laws, and the National Policy Framework for AI released on 20 March 2026. That framework is a set of legislative recommendations, not law. Its centre of gravity is pre-empting “unduly burdensome” state rules, alongside child safety, copyright, energy costs and jobs. Analysts expect the workforce and child-safety provisions, which have bipartisan appeal, to move faster than pre-emption. State laws remain operative until Congress acts.
4. China: state-steered control of authority
China has built a layered stack rather than a single AI law: the 2023 generative AI measures, mandatory content labelling from 1 September 2025, and a Cybersecurity Law amendment effective 1 January 2026 that wrote AI into national legislation for the first time. In 2026 it went further than anyone on agents. Implementation Opinions on intelligent agents, issued on 8 May and enforceable from 15 July 2026, reportedly create a three-tier authorisation model and require end users to retain final decision-making authority. Separate measures for anthropomorphic AI services took effect the same day.
5. United Kingdom: regulate by sector, assure by institute
The UK still has no AI-specific statute. Existing regulators apply their own rules, and the AI Security Institute provides frontier testing and the secretariat for the International AI Safety Report. The AI Opportunities Action Plan one-year report (29 January 2026) claimed 38 of 50 commitments met, and set the most explicit numeric skills target in this group: AI skills for 10 million workers by 2030.
6. Japan: promotion-first soft law
The AI Promotion Act of 2025 carries no penalties for private actors. Instead it created an AI Strategy Headquarters under the Prime Minister and a statutory Basic Plan. The first plan (23 December 2025) framed a national reboot through trustworthy AI. The second plan, adopted on 14 July 2026, pushes “AI Transformation”, asking organisations to rethink how decisions are made as agentic AI speeds them up. Business duties sit in the non-binding METI and MIC AI Guidelines for Business.
7. Korea: comprehensive law, light penalties
The AI Basic Act took effect on 22 January 2026, the second comprehensive AI law after the EU’s. Operators of “high-impact” AI in sectors such as healthcare, energy, finance, hiring and biometrics must assess impact, ensure human oversight and notify users. Fines are capped at KRW 30 million, and the ministry has granted at least a one-year grace period before they are imposed.
8. India: techno-legal and inclusion-first
MeitY’s India AI Governance Guidelines, released on 5 November 2025 ahead of the February 2026 AI Impact Summit, rest on seven guiding principles (sutras) adapted from the Reserve Bank’s FREE-AI framework, six governance pillars and a phased action plan. India chose not to legislate a dedicated AI law, preferring amendments to existing statutes and “techno-legal” safeguards built into systems. New institutions include an AI Governance Group and an AI Safety Institute.
9. Australia: existing law plus an institute
The National AI Plan of 2 December 2025 formally shelved the proposed mandatory guardrails for high-risk AI, choosing to rely on technology-neutral laws, the National AI Centre’s Guidance for AI Adoption, and an AI Safety Institute funded at A$29.9 million. One detail stands out: the Australian Public Service plan makes AI literacy training mandatory for public servants and puts a Chief AI Officer in every Commonwealth agency.
The regional overlay: APEC
Above all nine sits APEC’s own work. Leaders endorsed the APEC AI Initiative (2026 to 2030) at Gyeongju, with expanding AI participation through skills as one of three objectives. Under China’s 2026 host year, digital and AI ministers adopted the Chengdu Statement on 23 July 2026, committing economies to raise digital and AI literacy through capacity building and public education. Viet Nam; Hong Kong, China; Chinese Taipei and others are moving too, and ASEAN’s governance guide frames much of the region’s soft law.
V. Synthesis: nine frameworks, one missing person
Every one of the nine frameworks assumes a competent human will oversee the machine. Almost none of them says who trains that human, to what standard, or how anyone would verify it.
| Economy | Core instrument (status, Oct 2026) | Legal force | Agent-specific? | Human oversight expectation | Workforce competence provision |
|---|---|---|---|---|---|
| Singapore | MGF for Agentic AI v1.5 | Voluntary | Yes | Meaningful accountability; approval at critical points | Training against automation bias; audit oversight effectiveness |
| China | Agent Implementation Opinions; amended Cybersecurity Law | Binding | Yes | Tiered authorisation; user keeps final decision | Not specified |
| EU | AI Act as amended by Digital Omnibus | Binding (high-risk deferred to Dec 2027) | No | Human oversight for high-risk systems | Article 4 literacy duty, softened to “take measures to support” |
| Korea | AI Basic Act | Binding, light fines, grace period | No | Human oversight for high-impact AI | Not specified |
| United States | National Policy Framework (recommendations) | Non-binding; state laws persist | No | Federal agency use only | Workforce named as a legislative priority; no standard |
| United Kingdom | Sector regulators; AI Security Institute | Existing law | No | Via sector regulators | Target: AI skills for 10 million workers by 2030 |
| Japan | AI Promotion Act; second AI Basic Plan | Promotional, no penalties | Strategy references agents | Voluntary business guidelines | Organisation-wide “AI Transformation”; no standard |
| India | AI Governance Guidelines | Voluntary, techno-legal | No | Principle level | Capacity building and awareness in action plan |
| Australia | National AI Plan; Guidance for AI Adoption | Existing law plus voluntary guidance | No | Voluntary guidance | Mandatory AI literacy for public servants only |
Where they converge
- Humans stay accountable. No framework allows accountability to transfer to the system. Singapore says it outright; China writes it into the decision tiers; the EU and Korea attach it to high-risk and high-impact uses.
- Risk is graded, not uniform. Whether the label is high-risk, high-impact or Level 3, every regime concentrates its demands where harm is hardest to reverse.
- Transparency to the end user. Disclosure that you are dealing with AI, and labelling of synthetic content, is the closest thing to a global baseline.
- Institutions over statutes. Safety institutes in the UK, Australia and India, AI Verify in Singapore, a Strategy Headquarters in Japan. Governments are building capacity to test before they legislate.
- Delay is the new consensus. The EU deferred, Korea gave a grace year, Australia shelved guardrails, the US is litigating its own states. 2026 was the year most governments chose to watch before they bind.
Where they diverge
The deepest split is not binding versus voluntary. It is what is being governed. Most frameworks still regulate the output: is the content labelled, fair, accurate? Two, Singapore and China, have moved to regulating authority: what an agent is permitted to do, under whose identity, with which approvals. That is the shift from prompts to loops written into policy, and the others will have to follow.
The second split is the theory of who carries the burden. The EU and Korea place it on the provider and deployer. The US and Japan place it largely on the market. China places it on the state’s filing and testing apparatus. India places it on the technology itself, through techno-legal design.
The blind spot: accountability without a supply chain
Here is the finding that should headline every AI governance conference this year. All nine frameworks demand a human who is accountable. None defines a certifiable competence standard for that human. Singapore comes closest by naming training as a control. The EU had the only economy-wide legal duty on AI literacy, and in July 2026 it softened it from “ensure” to “support”. Australia mandates literacy only for its own public servants. The UK has a headcount target, not a capability standard.
We have regulated the cockpit and forgotten the pilot’s licence. Governance frameworks specify the checkpoint, the log, the override and the escalation path. They are silent on whether the person at the checkpoint can tell a plausible error from a correct answer at the speed an agent operates. That is the Frozen Workforce problem seen from the regulator’s side: the rules are being written for a supervisor class that our training systems are not yet producing.
Governing Intents and Thoughts, not Decisions and Actions – Luke Soon
VI. The Panel: eleven voices, three camps, one hard question
I put the argument to my usual panel. The positions below are paraphrased from their public statements and writing, not interviews, and I have tried to let each of them push back where they would.
They split into three camps, and the split is about speed and durability, not direction. Nobody on this panel thinks the operator era survives.
Camp one: the transition will be fast and painful
Dario Amodei used his January 2026 essay The Adolescence of Technology to restate that AI could displace half of entry-level white-collar jobs within one to five years, and to call for government intervention, including redistribution. His challenge to me: if the timeline is that short, “learning as a loop” is necessary but nowhere near sufficient. Skills policy without an income bridge is a lecture delivered to people who cannot pay rent.
Geoffrey Hinton has argued that routine intellectual labour is the first to go, and has half-jokingly suggested that physical trades such as plumbing are safer bets. His challenge: is “director of agents” a durable role, or just the next rung to be automated? If verification itself becomes machine work, my four-layer stack has a shelf life.
Mo Gawdat has described the coming decade and a half as a hard passage before an abundant one, and is sceptical that new jobs will appear fast enough to absorb the displaced. His point lands close to my own: short-term turbulence on the way to long-term abundance. Where we differ is that I think the turbulence is a design choice, not a weather system.
Camp two: the transition will be gradual and net positive
Sam Altman, in his 2025 essay The Gentle Singularity, accepted that whole classes of jobs will disappear but argued that wealth and new forms of work will grow faster than we expect, as they have in every prior shift. His challenge: I may be overweighting the freeze. Davos CEOs this year made a similar case, and some, such as Cognizant, said they were hiring more graduates, not fewer.
Andrew Ng, who did more than anyone to popularise the phrase “agentic workflows”, argues that people who use AI will replace people who do not, and that everyone should learn to build with it. His challenge: do not over-credential. Heavy competence regimes could slow adoption, and the best training is shipping real agents.
Eric Schmidt frames AI as a contest of national capability; his 2025 Superintelligence Strategy paper treats speed and security as inseparable. His challenge: any governance that slows domestic adoption is itself a strategic risk. Build the supervisors, but do not stop the build.
Elon Musk has repeatedly forecast that AI and robotics will make work optional, underwritten by what he calls universal high income. His challenge is the most radical: if work becomes optional, why design a workforce at all? My answer is that the path to optional work runs through a long period in which humans must still be accountable for what machines do.
Camp three: the question is control, not jobs
Yoshua Bengio chaired the International AI Safety Report 2026, which warns that greater agent autonomy makes it harder for humans to intervene before failures cause harm. His LawZero initiative is pursuing non-agentic “scientist” AI. His challenge cuts deepest: perhaps the answer is not better loop supervisors, but fewer fully autonomous loops. Supervision may not scale to machine speed.
Max Tegmark argues that AI should face binding safety standards like any other high-consequence industry, and backed the 2025 call to halt superintelligence development until it can be made controllable. His challenge: voluntary frameworks, which describe six of my nine, are not governance. They are aspiration.
Erik Brynjolfsson supplied the canary data, and has long warned of the “Turing Trap”: using AI to imitate and replace humans rather than augment them. His challenge: firms automate because incentives reward it. Training people to direct agents will not help if tax and procurement rules pay companies to remove the people.
Fei-Fei Li insists that AI should be human-centred by design, measured by whether it expands human dignity and agency. Her challenge: do not define the future worker solely as a supervisor of machines. Judgement, care and creativity are not leftovers. They are the point.
My response
The panel is right to attack the comfortable version of my argument. Skills alone will not save a cohort if the displacement is as fast as Amodei fears, or if incentives favour automation as Brynjolfsson warns. Supervision alone will not save us if Bengio is right that loops outrun human reaction time.
So the answer has to be Long-AND, not Short-OR. Build supervisory competence AND bridge incomes AND bind the highest-risk loops AND redesign incentives toward augmentation. Any one of those, on its own, is a slogan. Together, they are a strategy.
VII. Learning as a Loop: closing the supervisor gap
A two-year curriculum cycle cannot keep pace with a technology that changes every few months, so training has to become a loop, the same way the technology now works.
Three design principles make that possible.
- Learn on live work, not case studies. The programmes I have seen succeed put people on real, governed AI use cases inside their own organisation, with a coach, from week one. Capability is built in the flow of work, and the organisation gets value while it trains. This is also the only credible replacement for the vanishing apprenticeship described in Section II.
- Industry sets the signal, education builds the pathway, government removes the friction. Cloud and AI providers know which skills are moving fastest. Universities and polytechnics know how to build rigorous, stackable learning. Governments can fund it, recognise it and make it portable. Singapore’s SkillsFuture is one example of that division of labour; many economies have their own.
- Short, stackable, expiring credentials. Micro-credentials tied to real tasks that lapse unless renewed, so a worker’s record shows what they can do now, not what they could do three years ago.
A proposal: license the human to the tier of the loop
The synthesis in Section V points to a concrete fix. Regulators are already grading agent authority into tiers. We should grade human competence to match, so that the person approving a Level 2 action is demonstrably capable of judging it.
| Agent authority tier | Example actions | Human role | Competence evidence | Renewal |
|---|---|---|---|---|
| Tier 1: autonomous, reversible | Scheduling, retrieval, routine tickets | Agent operator | AI fluency; logged supervised hours | Every 24 months |
| Tier 2: approval required | Pricing changes, contract edits, infrastructure changes | Agent approver | Passed seeded-fault verification drills at agent speed | Every 12 to 18 months |
| Tier 3: human-only, hard to reverse | Payments, legal execution, safety-critical control | Accountable owner | Governance-by-design assessment; incident reconstruction from logs | Every 12 months, plus after any material incident |
This is not a new bureaucracy. It is the logic aviation, medicine and finance already apply: the higher the consequence, the more recently you must have proven you can handle it. It also answers Andrew Ng’s worry about over-credentialing, because Tier 1 stays light, and Yoshua Bengio’s worry about speed, because Tier 2 drills are run at the pace agents actually operate.
Design for the most exposed first
SMEs, older workers and workers outside the capital cities are the most exposed to this shift and the least likely to be offered training by default. Two of the nine frameworks, India’s and the APEC Chengdu Statement, put inclusion near the centre; most treat it as an afterthought. If we do not design for these groups first, the gap widens into what I call The Fork: one path leads to broad, shared abundance; the other leads to a narrow group who direct the machines and a much larger group who are directed by them.
VIII. Three Asks, and a Choice Being Made Now
The frameworks have built the checkpoints. The next decade will be decided by whether anyone is qualified to stand at them.
For governments: fund judgement, not just certification. Put AI fluency, verification and governance skills into domestic skills frameworks. Attach a competence standard to every human-oversight requirement you write, and recognise short, renewable credentials. If the EU’s softening of Article 4 tells us anything, it is that a literacy duty without a standard collapses under its own vagueness.
For industry: open up your live environments. Let learners work on real, governed use cases, and tell educators early which skills are moving. Rebuild the entry-level rung on purpose, as supervised apprenticeships alongside agents, rather than removing it by default.
For APEC and every regional forum: do what no single economy can. That means shared competency language for AI-augmented cloud roles, tiered supervisor credentials that map to agent authority levels, mutual recognition of those credentials across economies, and an honest exchange of what is actually working. The APEC AI Initiative already commits to building capacity across economies and workforces. This is how to make that commitment measurable.
I remain an optimist. I believe we are living through short-term turbulence on the way to long-term abundance. But abundance is not automatic. It depends on whether the people in our workforce become the directors of these systems, or merely their operators.
Every government in this survey has written that a human must stay accountable. None has yet written how that human is made. That sentence is the most important unfinished line in AI governance today, and it is being drafted, right now, in the training budgets and curriculum committees most of us never attend.
IX. The 2030 Ledger: what the consultancies, multilaterals and frontier labs actually measured
Put the major future-of-work studies side by side and a pattern appears: the bigger the headline number, the more theoretical the measure. The numbers that describe what is happening today are smaller, sharper, and concentrated on the young.
The common mistake is to compare these figures as if they measured the same thing. They do not. I sort them into four kinds of claim.
| Source | Date | Kind of measure | Headline finding |
|---|---|---|---|
| McKinsey Global Institute, Agents, robots, and us | Nov 2025 | Technical potential | Demonstrated technology could in theory automate about 57% of US work hours; about $2.9 trillion of annual US value by 2030, if workflows are redesigned |
| IMF | 2024 | Exposure | About 40% of global jobs exposed to AI, rising to about 60% in advanced economies; roughly half of those may benefit |
| ILO and NASK | May 2025 | Exposure | One in four workers globally (34% in high-income economies) in occupations with some generative AI exposure; transformation more likely than redundancy; women more exposed |
| Anthropic, Labor Market Impacts of AI | Mar 2026 | Observed use | New “observed exposure” measure: office administration about 40% observed against about 90% theoretical; higher observed exposure tracks lower official job-growth projections to 2034; no systematic rise in unemployment yet |
| OpenAI, GDPval | Sep 2025 | Capability against experts | Across 1,320 tasks in 44 occupations, best models were judged as good as or better than experts in roughly 40 to 49% of comparisons, improving roughly linearly |
| Goldman Sachs Research | Aug 2025 | Displacement forecast | 6 to 7% of the US workforce displaced if AI is widely adopted (range 3 to 14%); a temporary unemployment rise of about half a point; productivity up about 15% |
| WEF, Future of Jobs 2025 | Jan 2025 | Employer expectations | 170 million jobs created and 92 million displaced by 2030; 39% of core skills to change |
| BCG, AI at Work 2026 | Jun 2026 | Worker survey | 74% of frontline staff use AI regularly; agents in workflows rose from 13% to 30%; nearly half now spend more time managing and directing AI than doing the work |
| PwC, Global AI Jobs Barometer 2026 | Jun 2026 | Job-posting data | 62% wage premium for AI skills; AI-exposed entry-level roles seven times more likely to demand traditionally senior skills such as judgement and leadership |
| Anthropic, Economic Index | 2025 to 2026 | Usage data | Augmentation still edges automation on consumer use, but directive, delegate-it-all use keeps creeping up, especially through enterprise APIs |
Four readings of the ledger
1. Potential is not destiny, but the gap is closing. McKinsey’s 57% and Anthropic’s theoretical exposure describe a ceiling. Anthropic’s observed exposure describes the floor. The most important number in future-of-work research for the rest of this decade is the speed at which that gap closes, and agents are the mechanism that closes it.
2. The frontier labs are the most honest about uncertainty, and the most alarming about direction. Anthropic’s own economists found no systematic unemployment effect yet, while the company’s chief executive warns of a white-collar entry-level shock. Both can be true. One describes today’s data; the other describes the slope.
3. The work is already becoming supervision. BCG’s finding that nearly half of workers spend more time directing AI than doing the task is the clearest empirical confirmation of the shift from operator to director in Section I. It happened before most organisations wrote a single job description for it.
4. The ladder is being rebuilt with the bottom rungs missing. The wage-premium data and the canary data are two sides of one coin. Entry-level roles that survive are being “seniorised”: juniors are now asked for judgement, leadership and creativity on day one. That is the apprenticeship paradox from Section II, showing up in job adverts. Employers want judgement from people who have not yet been allowed to earn it.
X. The Workforce Thinkers: economists, HR strategists and the people who study work itself
My usual panel are builders and scientists. The people who study work for a living reach a sharper conclusion: the danger is not a shortage of work, but a failure of people to reach the work that exists.
Daniel Susskind: the friction, not the void
Susskind, author of A World Without Work, Growth: A Reckoning and the new What Should My Children Do? (September 2026 UK, November US), sits on the UK Government’s Expert Panel on AI and the Future of Work. His central distinction is between a world with too little work and a world where work exists but people cannot move into it, because they lack the skills, live in the wrong place, or cannot let go of an identity tied to the old job. In his April 2026 Gresham lecture he framed the near-term worry as exactly that kind of frictional mismatch. His education advice is pragmatic: do not ban the tools, teach people to use them effectively and critically, and think in tasks rather than jobs.
Where he sharpens my argument: the Frozen Workforce is Susskind’s friction made visible. Supervisory work is being created; the question is who can get there.
Daron Acemoglu, David Autor and Simon Johnson: pro-worker AI
In Building Pro-Worker Artificial Intelligence (NBER and the Hamilton Project, February 2026), two Nobel laureates and one of the world’s leading labour economists sort technology into five kinds: labour-augmenting, capital-augmenting, automating, expertise-levelling and new-task-creating. Only the last is unambiguously good for workers. They argue the market under-invests in pro-worker AI because of misaligned incentives, path dependence and a pervasive pro-automation ideology, and they propose nine policy levers, from tax reform to protecting worker expertise from being harvested to train its replacement.
Where they sharpen my argument: “agent supervisor” is a new-task-creating category. That is precisely the kind of work their framework says markets will under-produce unless policy tilts the field.
Carl Benedikt Frey: the political economy of displacement
Frey, the Oxford economist whose 2013 study with Michael Osborne launched the modern automation debate, argued in The Technology Trap that when technology replaces workers rather than enabling them, the losers resist, and progress itself becomes politically fragile. His warning for policymakers: a Frozen Workforce is not only an economic cost. It is a constituency for backlash.
Ethan Mollick: the jagged frontier
Mollick, at Wharton and author of Co-Intelligence, co-led the 2023 field experiment with BCG consultants that found AI made people faster and better on tasks inside its capability frontier, and measurably worse on tasks just outside it, because they trusted it where they should not have. This is the empirical foundation for my verification layer: the most valuable skill is knowing where the frontier is today, and it moves every few months.
Ravin Jesuthasan: work without jobs
Jesuthasan, co-author of Work Without Jobs and The Skills-Powered Organization, argues that the job is the wrong unit for managing work in an AI economy. Organisations should deconstruct roles into tasks and match skills to them dynamically. His contribution: tiered, expiring supervisor credentials only work if HR systems track skills at task level, not job titles.
Josh Bersin: the superworker
The HR industry analyst has popularised the idea of the AI-enabled “superworker”, one person whose output is multiplied by agents, and argues HR must redesign roles and operating models around it rather than bolt AI onto old structures. His challenge to me: if superworkers emerge, organisations may need far fewer people per outcome, which makes the inclusion agenda in Section VII more urgent, not less.
Lynda Gratton: redesigning work as a leadership act
Gratton, at London Business School, has long argued that work redesign has to be intentional and human-centred, built around energy, productivity and fairness, rather than left to whatever the technology defaults to. Her addition: governance of the loop is also governance of the employee experience. HX = CX + EX applies inside the firm as much as outside it.
Aneesh Raman: the broken bottom rung
LinkedIn’s chief economic opportunity officer warned in 2025 that AI is breaking the bottom rung of the career ladder, the junior tasks through which careers have always begun. His point is my apprenticeship paradox, stated from the hiring side, and PwC’s finding that entry-level roles are being “seniorised” is the data behind it.
What the thinkers agree on
| Thinker | Core idea | What it adds to this essay |
|---|---|---|
| Daniel Susskind | Frictional, not absolute, technological unemployment | The Frozen Workforce is a mobility problem, solvable by design |
| Acemoglu, Autor, Johnson | Pro-worker AI; markets under-invest in new-task creation | Supervisory roles need policy tilt, not just training budgets |
| Carl Benedikt Frey | Displacement breeds backlash | Inclusion is a stability issue, not charity |
| Ethan Mollick | The jagged frontier | Verification is the core skill, and it decays fast |
| Ravin Jesuthasan | Work without jobs; task-level skills | Credentials must attach to tasks and tiers, not titles |
| Josh Bersin | The superworker | Fewer people per outcome raises the stakes on access |
| Lynda Gratton | Intentional, human-centred work design | Loop governance is also employee-experience design |
| Aneesh Raman | The broken bottom rung | Rebuild apprenticeships alongside agents, on purpose |
The consultancies measured the potential. The labs measured the usage. The thinkers explain the mechanism: work is not disappearing, it is relocating upwards, into judgement and supervision, faster than people can climb. Which brings us back to the one line nine governments have yet to write: how the accountable human is made.
Sources and notes
Regulatory status is as of 1 October 2026. Panel positions are paraphrased from public statements; secondary reporting on China’s agent rules should be checked against the official texts.
Labour market and adoption
- Brynjolfsson, Chandar and Chen, Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of AI (Stanford Digital Economy Lab)
- Fortune, Stanford/ADP Canaries Dashboard coverage, June 2026
- Stanford Digital Economy Lab, AI Economic Indicators, June 2026 update
- World Economic Forum, Future of Jobs Report 2025 summary
- Gartner, 2026 Hype Cycle for Agentic AI
- KPMG Q2 2026 AI Pulse, via AI agent statistics roundup
- International AI Safety Report 2026; summary of findings
Government frameworks
- Singapore: Baker McKenzie on IMDA’s agentic framework; Version 1.5 update
- EU: Orrick on the Digital Omnibus; Morgan Lewis on the deferral; Article 4 as amended
- United States: Morgan Lewis on the National Policy Framework; EDUCAUSE summary
- China: Cloud Security Alliance research note on agent rules; Reed Smith on the Cybersecurity Law amendment
- United Kingdom: CMS on the Action Plan one-year report
- Japan: Second AI Basic Plan (provisional translation); White & Case on the AI Promotion Act
- Korea: Cooley on the AI Basic Act; US ITA market note
- India: MeitY guidelines launch coverage; PIB explainer, February 2026
- Australia: MinterEllison on the National AI Plan
- APEC: APEC AI Initiative 2026 to 2030; Chengdu Statement
Panel
- Amodei: Fortune on The Adolescence of Technology
- Other panel positions are drawn from widely reported public statements and essays, paraphrased rather than quoted.
Future-of-work research (Section IX)
- McKinsey Global Institute, Agents, robots, and us, November 2025
- Anthropic, Economic Index report, 2026; Massenkoff and McCrory, Labor Market Impacts of AI, via Fortune
- OpenAI, GDPval paper; TechCrunch coverage
- ILO and NASK, Generative AI and jobs: a 2025 update
- Goldman Sachs Research, How will AI affect the global workforce?, August 2025
- IMF 2024 exposure estimates, via AI and unemployment statistics roundup
- BCG, AI at Work 2026 press release
- PwC, 2026 Global AI Jobs Barometer press release
Workforce thinkers (Section X)
- Daniel Susskind: What Should My Children Do? (Oxford Martin School event); Gresham lecture, April 2026; Catalonia interview, July 2026
- Acemoglu, Autor and Johnson, Building Pro-Worker Artificial Intelligence, NBER Working Paper 34854, February 2026
- Frey, Mollick, Jesuthasan, Bersin, Gratton and Raman: positions drawn from their published books, studies and widely reported statements, paraphrased rather than quoted.
Leave a comment