By Wu Hui

Independent Researcher on Innovation and Talent | Tencent Research Institute AI Fellow

In the spring of 2023, an OpenAI research paper introduced the concept of "occupational exposure" to the mainstream public. By mapping large language model capabilities against nearly 1,000 occupational descriptions from the US Department of Labor's O*NET database, the study produced a stark exposure hierarchy: mathematicians, tax preparers, quantitative analysts, writers, and web designers topped the list, with coding and writing tasks showing near-100% exposure.

Over the following three years, fresh warnings and divergent data points arrived quarterly. Online professionals shared workflows completing a week's labor in three hours, job seekers watched entry-level openings vanish from corporate boards, and internal corporate town halls echoed a single mandate: deliver identical output with reduced headcount.

By mid-2026, the market conversation has shifted. Professionals no longer ask whether AI will eliminate their occupation; instead, they dissect their daily routines—identifying which specific tasks have been automated, which have been augmented, and which remain unaffected.

By SOLOMOAT Editorial Team

This analysis evaluates three core questions:

The Empirical Record (June 2023 – June 2026): What measurable impact has AI had on labor markets, hiring volume, and functional job categories?

Economic and Theoretical Frameworks: Which validated economic laws—from task-based labor models to Skill-Biased Technological Change (SBTC)—explain this trajectory and inform future projections?

The Human Dimension: How have professionals across engineering, product management, finance, HR, marketing, legal, and enterprise sales adapted their roles?

Part I: The Three-Year Impact — Data, Milestones, and Cases (2023.6–2026.6)

Between 2023 and 2026, market perceptions of AI displacement underwent systematic calibration. The cycle moved through four stages: identifying theoretical exposure, tracking real-world augmentation, diagnosing the lag between feasibility and enterprise deployment, and analyzing capital reallocation toward compute infrastructure.

1.1 2023: Theoretical Forecasts and the Exposure Narrative

The March 2023 OpenAI study evaluated 19,265 discrete work activities across 1,016 occupations in the O*NET database. The baseline findings indicated that roughly 80% of US workers had at least 10% of their work tasks exposed to large language models, while 19% saw over 50% exposure. High exposure was defined as a minimum 50% reduction in task completion time without quality degradation—a measure of technical feasibility rather than an immediate forecast of layoffs.

Concurrently, Goldman Sachs estimated that generative automation exposed approximately 300 million full-time jobs globally, identifying automation potential across 44% of legal workloads and 46% of administrative operations. McKinsey Global Institute projected generative AI could add between $2.6 trillion and $4.4 trillion annually to global output, with value concentrated across sales, marketing, software engineering, and customer operations.

These analyses framed the initial debate around two poles: immediate labor displacement versus macroeconomic expansion.

1.2 2024–2025: Augmentation Dominates and Implementation Quickens

By 2024, developer platforms recorded widespread tool adoption:

A GitHub enterprise survey found 92% of US software engineers utilizing AI coding assistants.

The Stack Overflow 2024 Developer Survey showed 81% of respondents acknowledging productivity gains, led by front-end, full-stack, and back-end engineers.

72% of surveyed developers viewed AI assistants as productivity augmentations rather than existential threats to their employment.

In February 2025, Anthropic’s Economic Index indicated that software engineering represented 37.2% of platform usage, followed by marketing and copywriting at 10.3%. Across measured enterprise workflows, 57% represented task augmentation, while 43% represented end-to-end automation.

Controlled trials by Boston Consulting Group and Harvard Business School revealed a "jagged technological frontier". Consultants leveraging GPT-4 completed 12.2% more tasks at 25.1% greater speed, yet suffered performance drops when assignments crossed beyond the model's reliability boundary. Human competence shifted toward verifying model error boundaries rather than baseline execution.

On labor exchanges, Freelancer.com reported surging demand for prompt engineers and AI content editors, alongside sharp volume declines in conventional copywriting, baseline graphic design, and manual data annotation. An Upwork study found 40% of surveyed executives intended to compress fixed full-time headcount in favor of specialized, AI-augmented contractors.

1.3 2026: Reality Calibration — Theoretical vs. Observed Exposure

In March 2026, Anthropic published Labor Market Impacts of AI: A New Measure and Early Evidence, distinguishing theoretical model capabilities from observed operational deployments.

Software engineering recorded the deepest penetration, matching the 2023 prediction: highly structured, closed-loop tasks with clear inputs and outputs automate first.

Conversely, sectors like legal services exhibit high theoretical exposure but low observed adoption due to institutional liability and regulatory friction rather than technical constraints.

1.4 Agent Clusters: The Transition from Copilots to Autonomous Fleets

By early 2026, industry reports from Anthropic and i-Research outlined a shift from interactive "copilots" to autonomous agent clusters. Multiple agents collaborate in parallel across complex assignments, leaving human operators to establish targets, set boundary conditions, and verify final deliverables.

Enterprise IT adoption followed a clear progression: agent architectures first absorbed development, unit testing, and continuous integration workflows before expanding into customer support and business process outsourcing (BPO). The operating framework transitioned from human-in-the-loop assistance to human-on-the-loop oversight.

1.5 The Broken Learning Curve: Structural Headwinds for Junior Talent

A 2025 Harvard Business School study by Joseph Fuller, analyzing millions of US job postings, introduced a key operational variable: the slope of occupational learning curves.

In steep learning-curve fields, junior tasks consist of explicit, programmatic steps easily handled by models. This breaks the traditional apprenticeship pipeline: experienced engineers experience productivity gains, while university graduates face reduced junior hiring.

The resulting workforce structure shifts from a traditional pyramid toward a dumbbell distribution: high demand for rare senior architects, reduced entry-level hiring, and hollowed-out intermediate tiers.

1.6 Capital Reallocation: Prioritizing Compute Over Headcount

Financial filings from June 2026 show capital flowing toward infrastructure over payroll:

US Tech Sector: Following macroeconomic reductions in late 2022, AI-driven programming efficiency contributed to an additional 100,000+ job reductions across major software providers over late 2025 and early 2026.

Chinese Tech Sector: More than half of major tech platforms reduced overall headcount starting in 2021. By June 2026, only Meituan, JD.com, PDD Holdings, and Tencent maintained staffing levels above 2021 baselines, with PDD and Tencent simultaneously expanding per-capita operating profit.

Capital Expenditures vs. R&D: Among market leaders (including Microsoft, Google, Meta, and Alibaba), capital spending on data centers, networking hardware, and custom silicon increasingly exceeds traditional software engineering R&D payrolls.

Corporate headcount growth has flattened while per-capita revenue climbs, as token budgets replace baseline administrative and junior developer salaries.

1.7 Micro Portraits: Four Cases of Displacement (2025–2026)

A June 2026 Caixin Weekly investigation highlighted how aggregate shifts translate into individual careers across the creative and analytical sectors:

Part II: Economic Laws and Structural Patterns

Evaluating long-term labor dynamics requires looking past cyclical fluctuations toward three established economic models.

2.1 The Task-Based Model: Shifting from Job Titles to Granular Workflows

Originating from David Autor, Frank Levy, and Richard Murnane (2003) and expanded by Daron Acemoglu and Pascual Restrepo, the task model holds that technology substitutes for discrete tasks rather than whole occupations.

Jobs consist of bundled activities; automation reallocates human labor toward higher-friction, non-routine tasks:

Routine Cognitive Tasks (e.g., standard bookkeeping, basic drafting, simple unit tests): Highly susceptible to automation.

Non-Routine Analytical & Interpersonal Tasks (e.g., system architecture tradeoffs, legal negotiation, crisis management): Retained by human operators.

This task division explains why headline legal and educational exposure metrics did not translate into immediate job losses: automation absorbed document discovery and lesson-plan drafting, leaving court advocacy and direct student evaluation under human direction.

2.2 Skill-Biased Technological Change (SBTC) and the Seniority Premium

Skill-Biased Technological Change explains why technology often increases demand for high-skill talent while compressing wages for intermediate and low-skill roles.

In the AI era, this dynamic functions as Task-Biased Change. Because baseline coding and routine writing are structured, they automate more quickly than tactile trade skills like plumbing or high-context management roles.

This dynamic widens the intra-occupational gap:

Senior Specialists: Command an increased premium for critical domain judgment, edge-case remediation, and risk management.

Junior Personnel: Face declining economic value for routine execution.

This produces a "superstar effect," where the top tier of domain experts captures outsized compensation while intermediate tiers face displacement.

Enterprise compensation systems, however, often lag this shift. Organizations frequently underprice emerging capabilities like multi-agent orchestration while continuing to pay legacy premiums for manual tasks that can now be automated.

2.3 Creative Destruction and Compensation Mechanisms: Where New Jobs Emerge

Joseph Schumpeter’s theory of creative destruction (1942) frames technological progress as an internal restructuring of the economy. Acemoglu and Restrepo categorize this into two counterbalancing forces:

The Displacement Effect: Direct reduction in labor demand for automated tasks.

The Compensation Effect: Rebound mechanisms that expand total employment through productivity gains, capital accumulation, and the creation of entirely new, complex tasks.

Across past industrial transitions, displacement effects materialized immediately, while compensation effects required decades to emerge.

For frontier AI to drive net macroeconomic growth rather than just capital-labor substitution, three conditions must align:

2.4 The Enterprise Shift: Moving from Open-Ended Budgets to ROI Scrutiny

By mid-2026, technology companies shifted internal AI policies from open-ended experimentation to ROI-driven allocation. Meta introduced strict token budget caps for engineering divisions, Microsoft restricted unmonitored third-party API access, and Tencent replaced uniform allowances with task-specific allocation models.

This shift reflects real operational constraints: McKinsey’s 2025 State of AI revealed that only 39% of surveyed global enterprises saw measurable EBIT contributions from generative AI initiatives.

Part III: Frontline Operator Profiles (2023–2026)

Field interviews conducted between March and June 2026 highlight how frontline professionals across key business functions are adapting their day-to-day operations.

3.1 Software Engineering: Lin Zhou — The Self-Orchestrating Developer

Lin Zhou entered big-tech front-end development in 2022. By late 2024, model improvements changed her daily workflow: "With modern reasoning models, the system outperformed typical junior interns."

She spent three months mapping her development habits, breaking her work into discrete, auditable skills with clear performance parameters.

Today, routine daily deliverables take two hours to complete, freeing bandwidth for system architecture, industry research, and technical blogging. "Junior front-end roles have compressed," she notes. "The real value has shifted to systems-level harness design and cross-functional project coordination."

3.2 Product Management: Zhou Zheng — Generative Engine Optimization (GEO)

After three years managing digital advertising infrastructure, Zhou observed that enterprise search traffic was moving away from traditional blue links toward conversational summaries like Google AI Overviews.

Recognizing that legacy SEO playbooks were losing effectiveness, he left his corporate role to build an enterprise platform that helps brands monitor and improve how they are cited in AI-generated answers.

"The barrier to building single-point AI writing tools has collapsed," Zhou explains. "The defensible opportunity lies in building the commercial infrastructure that connects enterprise brands to emerging AI search interfaces."

3.3 Marketing & Business Development: Chen Nian — The Trust Boundary

Chen Nian observed a clear split across her marketing and semiconductor BD roles: creative copywriting automated quickly, but high-stakes enterprise sales remained unchanged.

"AI can draft marketing copy in seconds, but it cannot navigate client politics, establish trust, or manage public brand risk," Chen notes.

Routine content generation has become commoditized, while high-touch relationship management and commercial risk judgment command growing premiums.

3.4 Human Resources: He Man — Managing Structural Restructuring

As an HR Business Partner, He Man automated routine administrative tasks like leave tracking and basic compensation spreadsheets using standardized prompts.

However, the interpersonal demands of organizational restructuring—such as exit interviews, team reorganizations, and cross-functional upskilling—remain high-touch.

She notes an emerging organizational friction: "Individual workers are completing tasks faster with AI, but broader corporate revenue and efficiency have not jumped in tandem, because overarching organizational designs have not updated to match these workflows."

3.5 Enterprise Sales: Lao Zhou — Transitioning from Hardware to Tokens

At thirty-eight, Lao Zhou adapted his enterprise hardware sales experience to modern cloud infrastructure—moving from selling physical server clusters to selling compute tokens.

While his product catalog changed, his core sales process relies on multi-month trust cycles, on-site meetings, and personalized account management.

He notes that enterprise demand for public API tokens remains cautious due to internal data privacy concerns, with security-conscious clients continuing to favor private deployments.

3.6 Corporate Finance: Song Yao — Moving from Bookkeeping to Strategic Advisory

As CFO of a Sino-foreign hardware joint venture, Song Yao observed two different deployment paces: the foreign parent automated standard accounting and consolidated regional reporting two years ago, while domestic operations moved more cautiously due to data governance rules.

"Junior financial analysts who only compile spreadsheets face serious displacement," Song observes. "The enduring value lies in strategic tax planning, deal structuring, and working directly alongside operational units to guide commercial decisions."

3.7 Corporate Legal: Shen Mo — Managing Token Costs and Model Hallucinations

In-house counsel Shen Mo uses specialized legal AI platforms for preliminary case law searches and standard contract drafting, saving 20% to 30% of her aggregate weekly review time.

However, adoption is tempered by model reliability and token expenses: unspecialized models continue to hallucinate citations, while specialized platforms generate high usage bills on multi-turn document reviews.

"AI handles standard boilerplate effectively, but resolving disputes, running negotiations, and interpreting business intent require direct human judgment," Shen concludes.

Strategic Takeaways for the Next Cycle

The impact of artificial intelligence on the labor market follows three distinct dimensions:

The Task Level: Highly structured, routine cognitive tasks continue to automate, while ambiguous, high-context judgment roles remain under human direction.

The Organizational Level: Corporate structures are shifting from layered management hierarchies toward lean, highly leveraged specialist nodes running multi-agent toolchains.

The Macroeconomic Level: Capital continues to flow toward compute infrastructure, while long-term employment growth depends on whether emerging AI applications create genuine net-new industries.

For individual professionals, career resilience depends on cultivating portable, non-routine competencies: task decomposition, output verification, risk arbitrage, and cross-disciplinary communication.

For enterprise leaders, the mandate is to move beyond short-term cost cuts and update operational workflows to support agent tools, while maintaining clear talent pipelines to train the next generation of senior domain experts.

For policymakers, the priority is modernizing educational and social safety frameworks to support flexible, modular work structures as the conventional single-employer model continues to fragment.

Frequently Asked Questions

Which work is most exposed to AI change?

Work with repeatable information processing changes first, while roles with judgment, trust, physical context, and accountability are redesigned differently.

What is a practical response?

Map your tasks, preserve the decisions only you can make, and build assets that compound your expertise beyond a job description.