The Learning Curve Nobody in Government Is Planning For

The data arrives before the plan

In late March 2026, Anthropic published its third Economic Index report — a rigorous, privacy-preserving analysis of one million Claude conversations from a ten-day window in February. The document is careful, methodologically sound, and written in the measured language of economic research.

It is also, if you read it in full, something of a fire alarm.

The key finding is buried in the section on learning curves: users who have spent more than six months working regularly with AI are now 10% more successful in their conversations, tackle tasks requiring nearly a year more education, and use AI in a fundamentally more collaborative way than newer users. The skill gap is not static. It compounds.

Meanwhile, at the level of entire industries, two automation workflows more than doubled in frequency in a single quarter: business sales and outreach (lead qualification, email drafting, data enrichment) and automated trading and market operations. These are not fringe use cases. They are bread-and-butter knowledge-economy workflows. They are also the entry point for a significant portion of the workforce into professional careers.

The report uses the phrase “skill-biased technological change.” I want to be plain about what that means in practice.

What the data shows sector by sector

Not all sectors are exposed equally, and not all exposure is negative. The picture is more granular than either the optimists or the pessimists tend to acknowledge.

High-exposure, high-displacement risk: Business administration, sales, basic legal processing, financial analysis, customer support, data entry, and content production. These are sectors where AI is already operating — not as a tool for human workers, but as a direct replacement workflow. The Anthropic data shows automation patterns in sales and trading have doubled in volume in a quarter. That pace is structural, not experimental.

High-exposure, transformation opportunity: Healthcare, education, engineering, and creative industries. AI is deeply useful in these sectors, but the value it produces is tightly coupled to human judgment, ethics, relationships, and contextual expertise. The workers who thrive here will be those who invest early in AI fluency — the “high-tenure” users the report describes as outperforming their less experienced peers.

Lower-exposure, but not immune: Manual trades, physical care work, complex physical manufacturing. These roles are protected by embodiment and contextual judgment — for now. But supply chains, logistics management, and the administrative layers surrounding these roles are not protected. The disruption propagates upward from the administrative overhead.

The mistake many sector analyses make is treating these categories as fixed. They are not. The rate of AI capability improvement means that exposure bands shift — what was “transformation opportunity” in 2024 becomes “displacement risk” in 2027 if the sector’s workers do not climb the learning curve.

The labor question nobody is answering

Here is the honest version of the question that polite economic reports tend to sidestep:

Who captures the productivity gains?

AI is a productivity-enhancing technology. That is not in dispute. The question is whether those productivity gains flow to workers — in the form of higher wages, shorter hours, better working conditions — or flow exclusively to capital, in the form of higher margins and reduced headcount.

Businesses do not exist to create jobs. They exist to create value for customers in the form of goods and services. Employment, from the perspective of the business, is a resource allocation decision. When a cheaper, more scalable resource (AI) becomes available for a class of tasks, rational business management deploys it. This is not villainy. It is the logic of competition.

But from the perspective of society, employment is how we distribute the gains of economic activity. It is the mechanism by which productivity increases translate into improved living standards across the population — not just for shareholders and senior management.

AI disrupts this distribution mechanism. If we do not replace or supplement it with deliberate policy — redistribution of productivity gains through profit-sharing, retraining investment, social safety net adaptation, taxation of AI-augmented productivity — the result is not a rising tide lifting all boats. It is a rising tide for the boats that own the AI, and a receding shore for everyone else.

This is not a theoretical future. The Anthropic data shows it beginning now.

The developing world’s double exposure

The global concentration finding in the report deserves more attention than it received: the top 20 countries increased their share of per-capita AI usage from 45% to 48% in six months. That means the world outside the top 20 — where the majority of humanity lives — is falling further behind in per-capita adoption, not catching up.

The developing world faces a specific and uncomfortable double exposure.

On one hand, it is often argued that human labour will remain cheaper than AI automation in lower-wage economies for years, and that this provides a buffer. This is partially true. A call-centre worker in a lower-income country costs less per hour than the cloud compute to replace them. That differential provides some protection — temporarily.

On the other hand, the developing world’s integration into the global economy is precisely through the channels most exposed to AI displacement: outsourced services, call centres, manufacturing for export, back-office processing, and free-zone assembly. Globalization gave these economies a path to prosperity through labour arbitrage. AI partially closes that arbitrage — not all at once, but progressively and irreversibly.

The developing economies most at risk are not those fully outside the global economy — they are the ones most deeply integrated into it through outsourced knowledge work and manufacturing.

And the speed of that exposure is not governed by local adoption rates. It is governed by adoption rates in the client countries — the US, Europe, and East Asia — where the cost pressure to automate will be relentless regardless of what happens in the countries doing the work.

Planning the bridge, not just the destination

I am not a pessimist about this. The era on the other side of this transition could be one of genuine shared prosperity — freed human capacity deployed on care, creation, relationship, governance, and the work that machines cannot do.

But between here and there is a bridge. And bridges do not build themselves.

The planning required operates at five simultaneous levels:

Individual: Build AI fluency now, not as a hobby but as a professional survival skill. The learning-curve finding is directional: there is compounding value in starting earlier.

Organizational: Audit which roles are exposed on a three-year horizon, not a ten-year one. Invest in reskilling before displacement, not after. Redesign workflows that capture AI productivity and share the gains with the workers whose roles change.

Sectoral: Industry associations, professional bodies, and sectoral training institutions need transformation roadmaps. Not generic “AI is coming” communications — specific mapping of which tasks within which roles are changing, by when, and what the transition path looks like.

National: Governments need policy frameworks that address both the opportunity and the displacement simultaneously. AI productivity taxation, universal reskilling funds, updated labour protection for the gig and AI-augmented economy, and investment in digital infrastructure that closes the adoption gap rather than widening it.

Regional and international: The geographic divergence in the Anthropic data is a warning that without coordinated multilateral action, AI will deepen the prosperity gap between the developed and developing world rather than narrowing it.

None of this is unprecedented. Every major general-purpose technology — electricity, the internal combustion engine, the internet — required some combination of these interventions. We managed those transitions. We can manage this one. But we have to start.

The value we carry forward

There is a harder question underneath all of this, and I want to name it before closing.

AI transition planning tends to focus on what we preserve economically. It should also focus on what we preserve culturally, socially, and humanly.

Some values and practices we will be glad to leave behind. Repetitive, underpaid, dignity-stripping work is not a human good worth preserving for its own sake. Some of the labour disruption AI causes is the displacement of work that should never have been designed for humans in the first place.

But other things are worth carrying over. The dignity of skilled craft. The social architecture of workplaces as communities. The structure that regular employment provides in people’s lives — not just economically but psychologically and socially. The transmission of knowledge and professional identity across generations, which happens through apprenticeship and mentorship, not through subscription tiers.

We should not plan the AI transition as if we are moving from one version of an economy to a slightly better one. We are deciding what kind of society we want to be — which values the new arrangement will embody, and which ones it will discard. That decision should be made deliberately, not by default.

The learning curve in the Anthropic report is about AI users getting better at their work over time. There is a larger learning curve we need to climb together: getting better at governing a technology that is transforming everything, faster than any previous technology, with higher stakes than any previous transition.

The data is arriving. What comes next depends on whether we plan or merely adapt.