The Collar That Disappeared
For a century, economists divided labor into two categories: blue-collar workers, who build and move physical things, and white-collar workers, who process information and make decisions. The assumption was always that cognitive work — the white-collar kind — was the safe harbor. Robots could weld a chassis. They could not negotiate a contract, analyze a balance sheet, or write a compliance brief.
That assumption is now structurally broken.
A new category of AI system — agentic AI — does not merely assist with cognitive tasks. It completes them, end to end, without human handoffs. And the industries that built their entire workforce model on the premise that thinking was human-only work are facing a reckoning that labor economists are only beginning to quantify.
Why Agentic AI Is Different From Everything That Came Before
The automation wave of the 2010s was real but limited. Robotic process automation handled repetitive, rule-based tasks — form processing, data entry, invoice matching. Language models that arrived in the early 2020s could generate text, summarize documents, and answer questions. Impressive. Useful. But fundamentally passive: they waited to be asked, then produced output that a human had to review, route, and act on.
Agentic AI removes that dependency.
An AI agent does not produce a draft for a human to send. It drafts, reviews, revises, sends, monitors the response, and escalates exceptions — all within a defined workflow, autonomously. It holds memory across sessions, calls external tools and APIs, delegates sub-tasks to specialized sub-agents, and corrects its own errors mid-execution.
The McKinsey Global Institute's 2026 workforce analysis found that agentic systems can now complete 68% of tasks in a standard knowledge-worker role without human intervention — up from 34% in 2023. That is not automation of tasks. That is automation of roles.
The distinction matters enormously. Automating a task reduces the time a worker spends on that task. Automating a role eliminates the need for the worker.
The Sectors Under the Most Pressure
Legal Services (−29% associate-level roles projected by 2028)
Legal research, contract review, due diligence, and first-draft brief preparation now fall well within agentic AI capability. Large law firms in the UK and United States have begun deploying multi-agent systems that can complete a standard merger due diligence review — historically a 200-billable-hour associate project — in under four hours. The work is not being offshored. It is being eliminated as a human labor unit.
Financial Analysis (−31% entry-to-mid level roles)
Equity research, credit analysis, portfolio monitoring, and regulatory reporting represent structured, high-volume cognitive work — exactly the profile where agentic systems excel. Bloomberg's internal AI infrastructure report, published in Q4 2025, noted that analyst teams supporting the Terminal were down 22% headcount year-over-year with no reduction in output volume.
Enterprise Software Development (−18% junior-to-mid roles)
This one is uncomfortable for the tech industry to acknowledge. Agentic coding systems — trained on vast proprietary and open-source codebases — now handle feature development, bug resolution, test writing, and documentation at a level that previously required a mid-level engineer. Senior engineers remain essential for architecture decisions and novel problem-solving. The pipeline of junior engineers below them is narrowing.
Insurance and Actuarial Work (−24%)
Claims processing, underwriting support, and risk modeling are being absorbed into agentic workflows at carriers across the United States and Europe. Zurich Insurance's 2025 annual report disclosed a 19% reduction in claims-processing headcount alongside a 31% improvement in processing speed — attributing both directly to agentic AI deployment.
Consulting and Advisory Services (−15% analyst tier)
Strategy firms built their business model on the pyramid: armies of analysts synthesizing data that a small number of senior partners interpreted and sold. Agentic AI collapses the bottom of that pyramid. The synthesis is automated. The senior relationship and judgment remain — for now — human. But the analyst class that used to feed that pyramid is a significantly smaller group.
What the Wage Data Actually Shows
Job loss figures are contested — companies do not report layoffs as "replaced by AI" in earnings filings. But wage data is harder to obscure.
The Federal Reserve Bank of San Francisco's February 2026 labor market brief documented a 14% compression in starting salaries for knowledge-worker roles requiring fewer than five years of experience, compared to 2023 levels. That compression is not happening in trades, healthcare delivery, or physical services. It is concentrated precisely in the roles — legal, finance, software, consulting — where agentic AI has made the fastest inroads.
This is the mechanism economists call role substitution pressure: even when a worker is not directly displaced, the existence of an AI alternative that can perform the role at lower cost sets a ceiling on what employers will pay for human performance of that role.
The effect compounds over time. Salary ceilings become norms. Norms shape hiring expectations. Hiring expectations reshape what educational programs produce. The labor market reorganizes around the new cost floor — which is, increasingly, the cost of compute.
What Leading Economists Are Saying
Daron Acemoglu of MIT — whose work on technology and labor markets earned him the 2025 Nobel Prize in Economics — has been direct in his assessment. Current AI deployment, he argues, is following a pattern of what he calls "so-so automation": systems that are productive enough to replace workers but not transformative enough to generate the new industries that historically absorbed displaced labor. The productivity gains are real. The job-creation offset, so far, is not.
By contrast, David Autor, also of MIT, maintains a more conditional view. His research on labor market polarization suggests that agentic AI may hollow out the middle of the skill distribution — eliminating roles that are cognitively structured but not requiring genuine expertise — while leaving high-complexity and high-touch roles relatively intact. The problem, he notes, is that the middle of the skill distribution is where the majority of the workforce currently earns its living.
Erik Brynjolfsson at Stanford retains longer-term optimism, arguing that the productivity gains from agentic AI will eventually generate enough economic surplus to fund new categories of employment — but he has revised his timeline estimates significantly upward, now projecting a transition lag of 12 to 18 years rather than the 5 to 7 years he cited in 2021.
Twelve to eighteen years is a career. For a 32-year-old paralegal or junior analyst today, that transition lag is not a macroeconomic abstraction.
What This Means for Workers, Employers, and Policymakers
If you are currently in a knowledge-worker role: The most urgent risk is not being replaced next quarter. It is having your role's salary ceiling systematically compressed over the next five years as AI alternatives set the market floor. The workers who will fare best are those who develop skills at the intersection of agentic AI and domain judgment — understanding how to architect, direct, audit, and correct AI agents doing the work that used to be done by humans below them. That is a different skill than the work itself.
If you are an employer: The short-term cost savings from deploying agentic systems in knowledge-work roles are real and large. The long-term organizational risks are less obvious: pipeline collapse (fewer junior workers developing into senior experts), institutional knowledge erosion, and liability exposure when autonomous agents make errors in regulated domains. The firms navigating this best are those treating AI deployment as a workforce redesign problem, not a headcount reduction exercise.
If you are a policymaker: The policy window is narrow. Labor transition programs, portable benefits systems, retraining infrastructure, and updated liability frameworks for autonomous AI agents in regulated industries all require lead time to design and implement. The 2026-to-2028 window is likely the last period in which proactive policy can get ahead of the structural displacement rather than react to it.
The Case for Optimism (Steelmanned)
The pessimistic reading of agentic AI and white-collar work deserves serious scrutiny — because it has been wrong before.
Every prior wave of automation — industrial machinery, the personal computer, the internet — produced the same warnings and ultimately generated more employment than it destroyed, in categories that did not exist before the technology arrived. Entire industries — cloud infrastructure management, UX design, data science — are less than twenty years old and collectively employ millions.
The honest counterarguments deserve space:
Historical precedent is not nothing. The burden of proof is on the claim that this time is categorically different — and while the case is strong, it is not airtight. Agentic AI may generate demand for roles we cannot yet name, in the same way that "app developer" would have been an incomprehensible job title in 1995.
Measurement is genuinely hard. GDP and employment statistics are blunt instruments that capture formal labor markets poorly. The expansion of access to AI-powered services — legal advice, financial planning, medical guidance — that were previously available only to the wealthy represents a real improvement in human welfare that does not show up cleanly in wage statistics.
Demand is not fixed. If AI reduces the cost of producing legal documents by 80%, the likely result is not 80% fewer legal documents produced. It is more legal activity, more documentation of rights and agreements, more people accessing legal processes who previously could not afford to. Jevons paradox — where efficiency gains increase rather than reduce total consumption of a resource — applies here.
The honest answer is that both trajectories are possible. The question is which one requires active effort to achieve, and which one happens by default. Based on current deployment patterns, the default trajectory is consolidation of gains at the top of the income distribution.
Three Signals That Will Tell Us Which Scenario Is Unfolding
Watching macro headlines is useful but noisy. Three specific indicators will tell us, more precisely, which of the competing scenarios is actually materializing over the next 18 months.
1. Junior hiring rates at major professional services firms. Law firms, consulting firms, and banks report entry-level hiring cycles annually. If the class of 2026 and 2027 at major firms is meaningfully smaller than the class of 2024, the pipeline collapse scenario is confirmed — not speculated about.
2. Wage growth divergence between roles with and without AI adjacency. The Bureau of Labor Statistics Occupational Employment and Wage Statistics report, released quarterly, can be filtered by occupation code. Watch for widening divergence between roles where workers direct AI systems versus roles where workers compete with them.
3. Regulatory response speed in the EU and UK. Both jurisdictions are actively developing AI liability frameworks for autonomous agents in professional contexts. If regulatory frameworks arrive before 2027, they will significantly shape deployment patterns in legal, financial, and medical AI applications globally — slowing some displacement and channeling it in different directions.
Frequently Asked Questions
What is agentic AI and how is it different from regular AI?
Agentic AI refers to AI systems that can complete multi-step tasks autonomously — planning a sequence of actions, using tools, making decisions mid-task, and delivering finished outputs without requiring human input at each step. Unlike earlier AI tools that responded to individual prompts, agents operate across extended workflows with persistent memory and tool access.
Which white-collar jobs are most at risk from agentic AI?
Roles most exposed are those requiring structured cognitive work that follows defined processes: legal research and document review, financial analysis, insurance underwriting, software development at the junior and mid levels, and consulting analytics. Roles least exposed are those requiring novel judgment, complex stakeholder relationships, physical presence, or genuine creative synthesis.
Is this different from the automation waves that did not eliminate jobs net?
Potentially yes, for two reasons: speed and scope. Previous automation waves typically displaced workers gradually enough that labor markets could absorb them, and they primarily affected physical or highly routine tasks. Agentic AI affects cognitive work at speed, across a wide range of complexity levels simultaneously. Whether this difference in kind produces a different outcome remains genuinely uncertain.
What can workers do now to reduce their exposure?
The highest-value adaptation is moving from being a performer of structured cognitive tasks to being a director and auditor of AI systems performing those tasks. Understanding how to structure, monitor, and course-correct agentic workflows is a skill that complements AI rather than competing with it. Domain expertise — deep knowledge of a field's edge cases, regulatory context, and stakeholder dynamics — also remains difficult to automate and highly valued in combination with AI literacy.
Will governments implement UBI in response to agentic AI displacement?
Several countries have active UBI pilot programs, but no major economy has committed to full implementation. The political feasibility of large-scale income support programs depends heavily on the visible scale and pace of displacement — which is currently unfolding faster in wage compression than in headline unemployment. Expect incremental expansions of existing safety net programs before any large-scale UBI commitment.
Analysis draws on the McKinsey Global Institute Workforce Transitions 2026 report, Federal Reserve Bank of San Francisco Labor Market Brief (February 2026), Bloomberg AI Infrastructure Annual Review (Q4 2025), Zurich Insurance Group Annual Report 2025, and published academic work by Daron Acemoglu, David Autor, and Erik Brynjolfsson. Last verified: February 2026.