Published at: 2026-05-18T21:08:31+05:30
Thesis
The next labor shock from AI will not be a clean wave of “jobs disappearing.” It will be something more operational and harder to see at first: the interface to jobs will change. The verbs that make up a role—drafting, searching, triaging, summarizing, routing, estimating, checking, scheduling—will be pulled out of people and into agentic software systems. The job title may remain, and headcount may even stay flat for a while, but the workflow will flip: people will supervise, specify, and audit work that machines execute.
This matters because we tend to protect jobs (titles, teams, org charts) rather than protect worker bargaining power and human agency inside workflows. If the interface changes, the locus of value shifts to whoever controls the agent stack: the prompts, guardrails, data access, evaluation harnesses, and the right to decide what counts as “done.”
If you want a first-principles prediction, it is this: AI will turn many jobs into API surfaces. Some people will become “API owners” and gain leverage. Others will be treated like interchangeable “human fallback,” invoked only when automation breaks.
Context
We have already seen credible evidence that generative AI tools can raise productivity in specific knowledge-work settings, especially for less experienced workers.
In a field experiment with more than 5,000 customer support agents, access to a generative AI assistant increased productivity by about 14% on average, with especially large gains for novices and smaller gains for experienced workers. It also improved customer sentiment and reduced turnover (a useful reminder that productivity is not the only outcome that moves). Generative AI at Work (NBER Working Paper 31161)
In a controlled experiment on professional writing tasks, exposure to ChatGPT reduced time spent and improved output quality, while changing the structure of the work toward idea generation and editing rather than raw drafting. Noy & Zhang (MIT), “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence”
At the macro level, forecasts disagree. Some institutions project large productivity gains, while others argue for more modest effects once you account for task feasibility, costs, and slow diffusion. Goldman Sachs: “How Will AI Affect the Global Workforce?” and Acemoglu on the economics of AI illustrate the spread.
Meanwhile, labor market exposure measurement is getting more concrete. Anthropic has proposed an “observed exposure” approach that combines what models can do with what people actually use them for at work—an attempt to move from speculative capability to real usage. Anthropic: “Labor market impacts of AI: A new measure and early evidence”
And employer sentiment is shifting from “AI is a tool” to “AI is a workforce transformation driver.” The World Economic Forum’s Future of Jobs Report 2025 frames AI and information processing as one of the dominant trends reshaping work through 2030. WEF Future of Jobs Report 2025 (PDF)
So the argument is not that AI will “do nothing.” It is that the mechanism is more specific than job-loss headlines suggest.
Key ideas
1) A job is not a thing; it is a bundle of tasks, and tasks have interfaces
A job title is a compression algorithm. “Product manager,” “analyst,” “accountant,” “customer support,” “recruiter,” “lawyer,” “teacher”—each label hides a bundle of small actions.
Those small actions have interfaces:
Inputs: data, requests, constraints, context.
Transformations: reasoning, drafting, classification, planning.
Outputs: text, decisions, tickets, code, approvals.
Verification: checks, audits, sign-offs.
Historically, the interface has been human. Work moved through Slack messages, meetings, spreadsheets, and individual memory. The key property of a human interface is informal tolerance: ambiguity can be resolved in conversation. Another key property is informal accountability: a person can be blamed even if the system was unclear.
Agentic software flips this: the job interface becomes explicit, machine-readable, and routable. And once an interface becomes explicit, it becomes copyable.
2) AI’s first big win is not replacing expertise; it is replacing “glue work”
Most organizations are held together by glue work: status updates, coordination, follow-ups, meeting notes, formatting, doc hygiene, and the constant translation between one team’s mental model and another’s.
Glue work is rarely anyone’s “job,” which is why it is under-incentivized. But it is also the easiest surface for AI to attack because:
It is text-heavy.
It is repetitive.
It is often low-stakes per action but high-stakes in aggregate.
The output can be “good enough” and still create value.
This is why the earliest productivity evidence often shows largest gains for novices. The model acts like a best-practice diffuser: it gives newer workers a scaffold. That pattern appears in the customer support study where productivity gains were concentrated among the least experienced workers. NBER Working Paper 31161
The deeper implication is not “novices get better.” It is “the organization can hire fewer mid-level glue workers and still function.” That changes the career ladder.
3) Agents do not just automate tasks; they change where decisions live
Classic automation is procedural: you specify steps, the computer follows them.
Agentic automation is goal-directed: you specify an outcome, the system tries actions, calls tools, asks questions, and iterates.
When systems are goal-directed, the “decision” shifts from the person doing the task to the person defining:
Objective functions (“maximize resolution speed without hurting satisfaction”)
Constraints (“do not mention internal policy wording”)
Tool access (“can it query customer data or only public docs?”)
Evaluation (“what counts as correct, safe, complete?”)
This is governance disguised as productivity. The person who owns evaluation harnesses and permissions effectively owns the job.
4) The real labor impact arrives as recomposition, not simple substitution
The most common mistake in AI labor talk is to imagine substitution at the job-title level. In reality, work is recomposed.
Consider what happens when an AI assistant drafts first passes. The human role becomes:
Prompting and framing
Selecting and rejecting
Editing and tightening
Checking sources and correctness
Making calls under uncertainty
That is consistent with the writing-task experiment showing a shift in time allocation and the nature of effort when ChatGPT is available. Noy & Zhang
Recomposition is why employment might not drop immediately even as output rises. The organization can keep the same number of roles, ship more, and only later realize it no longer needs the same layers.
5) “Observed exposure” is the right framing: capability is not destiny
A model may theoretically perform a task, but that does not mean it will.
The adoption pipeline has friction:
Data access permissions
Risk tolerance and compliance
Integration cost
Measurement (knowing whether it works)
Incentives (who benefits, who loses)
Anthropic’s “observed exposure” idea matters because it tries to incorporate real usage and the direction of use (automation vs augmentation) rather than treating capability as equivalent to displacement. Anthropic labor market impacts
The same principle applies inside companies: the bottleneck is often not whether the model can draft a memo. The bottleneck is whether the organization can trust and govern the memo pipeline.
6) The ladder breaks where the interface gets automated
Every career ladder is secretly a sequence of tasks that train judgment.
Junior roles historically do high-volume, low-judgment work:
First drafts
Basic analysis
Ticket triage
Simple QA
Research summaries
Those tasks are exactly where AI is strongest.
So the ladder breaks in the middle. Organizations will say, “We still need seniors.” But seniors are created by reps. If reps are automated, senior supply shrinks. This produces a paradox:
AI makes juniors more productive.
Firms hire fewer juniors.
Future seniors become scarce.
Firms respond by overpaying for seniors and importing them.
This is not a theoretical concern. Evidence summarized by the IMF suggests AI-vulnerable occupations may see reduced employment in regions with high demand for AI skills, and it highlights pressure on entry-level work. IMF blog (2026)
You can already feel this in software and consulting: fewer junior analyst tasks exist when a model can produce a decent first draft.
7) The new scarce skill is “workflow ownership,” not prompt writing
Prompt writing is easy to teach.
Workflow ownership is hard. It includes:
Domain understanding
Data governance
Threat modeling
Human-in-the-loop design
Metrics, evaluation, red-teaming
Change management
In other words, the scarce capability is building systems that stay trustworthy under pressure.
This is also where value accrues. If a company can safely let an agent touch customer-facing actions, the ROI is huge. If it cannot, it gets a fancy autocomplete.
8) Human agency becomes the core design question
When AI systems take more actions, humans risk becoming “rubber stamps.”
That is dangerous for two reasons:
It erodes skill. People lose the ability to do the work without the tool.
It erodes accountability. If everyone is “just approving what the model suggested,” responsibility diffuses.
The Stanford SALT Lab’s work on “Future of Work with AI Agents” emphasizes measuring not just what can be automated, but what workers want automated or augmented, and it proposes a Human Agency Scale to quantify preferred human involvement. SALT Lab: Future of Work with AI Agents
This is the right lens. The problem is not AI at work. The problem is the design of work where humans still meaningfully steer outcomes.
Counterarguments
Counterargument 1: “AI will replace whole jobs quickly; the interface story is too slow and optimistic.”
There are roles where substitution will be faster than recomposition, especially where:
Outputs are standardized
Errors are cheap
Data is abundant
The task is mostly text transformation
Some customer support, basic content creation, simple compliance drafting, and internal reporting will compress fast.
But “whole job replacement” still requires integrating with messy systems and tolerating liability. Even in the studies showing big productivity effects, the human remains the accountable agent. The tool changes the workflow more than it eliminates the role overnight. NBER Working Paper 31161
So replacement will happen, but as a downstream consequence of interface change: when the workflow becomes clean enough, the organization realizes it no longer needs as many humans in the loop.
Counterargument 2: “This is just another productivity tool. We had email and spreadsheets; jobs kept growing.”
True: many tools amplify output without reducing employment.
The difference is action. Email and spreadsheets did not autonomously decide what to do next and call tools to do it. Agentic systems can.
That means the “unit of automation” is not a single task. It is a chain of tasks.
Once you automate chains, you compress coordination layers: fewer managers, fewer project coordinators, fewer analysts whose job is to assemble status.
Counterargument 3: “If productivity rises, wages rise, and we all win.”
Sometimes.
But wages rise when workers have bargaining power. If AI changes job interfaces such that the worker becomes replaceable, the gains can flow primarily to capital and to the small set of people who control deployment.
This is why the question is not “Does AI increase productivity?” but “Who owns the interface?”
Takeaways
Jobs are bundles of tasks, and AI is changing the interface through which those tasks are executed.
The early, high-leverage wins are in glue work: coordination, drafting, triage, and summarization.
Agentic systems move decision power upstream to whoever defines objectives, constraints, tool access, and evaluation.
The most important labor effect is recomposition: humans shift toward supervision, specification, and audit.
Entry-level ladders are at risk because the tasks that generate reps are the easiest to automate.
“Observed exposure” is a better framing than raw capability: adoption friction and governance matter.
The new scarce skill is workflow ownership: building trustworthy systems, not clever prompts.
Human agency must be designed, measured, and protected, or workers become rubber stamps.
If you want leverage, learn to own interfaces: data access, evals, guardrails, and system design.
Sources
NBER Working Paper 31161: Generative AI at Work — https://www.nber.org/papers/w31161
Noy & Zhang (MIT): Experimental Evidence on the Productivity Effects of Generative AI — https://economics.mit.edu/sites/default/files/inline-files/Noy_Zhang_1.pdf
Anthropic: Labor market impacts of AI (observed exposure) — https://www.anthropic.com/research/labor-market-impacts
World Economic Forum: Future of Jobs Report 2025 (PDF) — https://reports.weforum.org/docs/WEF_Future_of_Jobs_Report_2025.pdf
Stanford SALT Lab: Future of Work with AI Agents — https://futureofwork.saltlab.stanford.edu/
Goldman Sachs: How Will AI Affect the Global Workforce? — https://www.goldmansachs.com/insights/articles/how-will-ai-affect-the-global-workforce
MIT Economics: Acemoglu on the economics of AI — https://economics.mit.edu/news/daron-acemoglu-what-do-we-know-about-economics-ai
IMF blog: New Skills and AI Are Reshaping the Future of Work — https://www.imf.org/en/blogs/articles/2026/01/14/new-skills-and-ai-are-reshaping-the-future-of-work
The next labor shock from AI will not be a clean wave of “jobs disappearing.” It will be something more operational and harder to see at first: the interface to jobs will change. The verbs that make up a role—drafting, searching, triaging, summarizing, routing, estimating, checking, scheduling—will be pulled out of people and into agentic software systems. The job title may remain, and headcount may even stay flat for a while, but the workflow will flip: people will supervise, specify, and audit work that machines execute.
This matters because we tend to protect jobs (titles, teams, org charts) rather than protect worker bargaining power and human agency inside workflows. If the interface changes, the locus of value shifts to whoever controls the agent stack: the prompts, guardrails, data access, evaluation harnesses, and the right to decide what counts as “done.”
If you want a first-principles prediction, it is this: AI will turn many jobs into API surfaces. Some people will become “API owners” and gain leverage. Others will be treated like interchangeable “human fallback,” invoked only when automation breaks.
Context
We have already seen credible evidence that generative AI tools can raise productivity in specific knowledge-work settings, especially for less experienced workers.
In a field experiment with more than 5,000 customer support agents, access to a generative AI assistant increased productivity by about 14% on average, with especially large gains for novices and smaller gains for experienced workers. It also improved customer sentiment and reduced turnover (a useful reminder that productivity is not the only outcome that moves). Generative AI at Work (NBER Working Paper 31161)
In a controlled experiment on professional writing tasks, exposure to ChatGPT reduced time spent and improved output quality, while changing the structure of the work toward idea generation and editing rather than raw drafting. Noy & Zhang (MIT), “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence”
At the macro level, forecasts disagree. Some institutions project large productivity gains, while others argue for more modest effects once you account for task feasibility, costs, and slow diffusion. Goldman Sachs: “How Will AI Affect the Global Workforce?” and Acemoglu on the economics of AI illustrate the spread.
Meanwhile, labor market exposure measurement is getting more concrete. Anthropic has proposed an “observed exposure” approach that combines what models can do with what people actually use them for at work—an attempt to move from speculative capability to real usage. Anthropic: “Labor market impacts of AI: A new measure and early evidence”
And employer sentiment is shifting from “AI is a tool” to “AI is a workforce transformation driver.” The World Economic Forum’s Future of Jobs Report 2025 frames AI and information processing as one of the dominant trends reshaping work through 2030. WEF Future of Jobs Report 2025 (PDF)
So the argument is not that AI will “do nothing.” It is that the mechanism is more specific than job-loss headlines suggest.
Key ideas
1) A job is not a thing; it is a bundle of tasks, and tasks have interfaces
A job title is a compression algorithm. “Product manager,” “analyst,” “accountant,” “customer support,” “recruiter,” “lawyer,” “teacher”—each label hides a bundle of small actions.
Those small actions have interfaces:
Inputs: data, requests, constraints, context.
Transformations: reasoning, drafting, classification, planning.
Outputs: text, decisions, tickets, code, approvals.
Verification: checks, audits, sign-offs.
Historically, the interface has been human. Work moved through Slack messages, meetings, spreadsheets, and individual memory. The key property of a human interface is informal tolerance: ambiguity can be resolved in conversation. Another key property is informal accountability: a person can be blamed even if the system was unclear.
Agentic software flips this: the job interface becomes explicit, machine-readable, and routable. And once an interface becomes explicit, it becomes copyable.
2) AI’s first big win is not replacing expertise; it is replacing “glue work”
Most organizations are held together by glue work: status updates, coordination, follow-ups, meeting notes, formatting, doc hygiene, and the constant translation between one team’s mental model and another’s.
Glue work is rarely anyone’s “job,” which is why it is under-incentivized. But it is also the easiest surface for AI to attack because:
It is text-heavy.
It is repetitive.
It is often low-stakes per action but high-stakes in aggregate.
The output can be “good enough” and still create value.
This is why the earliest productivity evidence often shows largest gains for novices. The model acts like a best-practice diffuser: it gives newer workers a scaffold. That pattern appears in the customer support study where productivity gains were concentrated among the least experienced workers. NBER Working Paper 31161
The deeper implication is not “novices get better.” It is “the organization can hire fewer mid-level glue workers and still function.” That changes the career ladder.
3) Agents do not just automate tasks; they change where decisions live
Classic automation is procedural: you specify steps, the computer follows them.
Agentic automation is goal-directed: you specify an outcome, the system tries actions, calls tools, asks questions, and iterates.
When systems are goal-directed, the “decision” shifts from the person doing the task to the person defining:
Objective functions (“maximize resolution speed without hurting satisfaction”)
Constraints (“do not mention internal policy wording”)
Tool access (“can it query customer data or only public docs?”)
Evaluation (“what counts as correct, safe, complete?”)
This is governance disguised as productivity. The person who owns evaluation harnesses and permissions effectively owns the job.
4) The real labor impact arrives as recomposition, not simple substitution
The most common mistake in AI labor talk is to imagine substitution at the job-title level. In reality, work is recomposed.
Consider what happens when an AI assistant drafts first passes. The human role becomes:
Prompting and framing
Selecting and rejecting
Editing and tightening
Checking sources and correctness
Making calls under uncertainty
That is consistent with the writing-task experiment showing a shift in time allocation and the nature of effort when ChatGPT is available. Noy & Zhang
Recomposition is why employment might not drop immediately even as output rises. The organization can keep the same number of roles, ship more, and only later realize it no longer needs the same layers.
5) “Observed exposure” is the right framing: capability is not destiny
A model may theoretically perform a task, but that does not mean it will.
The adoption pipeline has friction:
Data access permissions
Risk tolerance and compliance
Integration cost
Measurement (knowing whether it works)
Incentives (who benefits, who loses)
Anthropic’s “observed exposure” idea matters because it tries to incorporate real usage and the direction of use (automation vs augmentation) rather than treating capability as equivalent to displacement. Anthropic labor market impacts
The same principle applies inside companies: the bottleneck is often not whether the model can draft a memo. The bottleneck is whether the organization can trust and govern the memo pipeline.
6) The ladder breaks where the interface gets automated
Every career ladder is secretly a sequence of tasks that train judgment.
Junior roles historically do high-volume, low-judgment work:
First drafts
Basic analysis
Ticket triage
Simple QA
Research summaries
Those tasks are exactly where AI is strongest.
So the ladder breaks in the middle. Organizations will say, “We still need seniors.” But seniors are created by reps. If reps are automated, senior supply shrinks. This produces a paradox:
AI makes juniors more productive.
Firms hire fewer juniors.
Future seniors become scarce.
Firms respond by overpaying for seniors and importing them.
This is not a theoretical concern. Evidence summarized by the IMF suggests AI-vulnerable occupations may see reduced employment in regions with high demand for AI skills, and it highlights pressure on entry-level work. IMF blog (2026)
You can already feel this in software and consulting: fewer junior analyst tasks exist when a model can produce a decent first draft.
7) The new scarce skill is “workflow ownership,” not prompt writing
Prompt writing is easy to teach.
Workflow ownership is hard. It includes:
Domain understanding
Data governance
Threat modeling
Human-in-the-loop design
Metrics, evaluation, red-teaming
Change management
In other words, the scarce capability is building systems that stay trustworthy under pressure.
This is also where value accrues. If a company can safely let an agent touch customer-facing actions, the ROI is huge. If it cannot, it gets a fancy autocomplete.
8) Human agency becomes the core design question
When AI systems take more actions, humans risk becoming “rubber stamps.”
That is dangerous for two reasons:
It erodes skill. People lose the ability to do the work without the tool.
It erodes accountability. If everyone is “just approving what the model suggested,” responsibility diffuses.
The Stanford SALT Lab’s work on “Future of Work with AI Agents” emphasizes measuring not just what can be automated, but what workers want automated or augmented, and it proposes a Human Agency Scale to quantify preferred human involvement. SALT Lab: Future of Work with AI Agents
This is the right lens. The problem is not AI at work. The problem is the design of work where humans still meaningfully steer outcomes.
Counterarguments
Counterargument 1: “AI will replace whole jobs quickly; the interface story is too slow and optimistic.”
There are roles where substitution will be faster than recomposition, especially where:
Outputs are standardized
Errors are cheap
Data is abundant
The task is mostly text transformation
Some customer support, basic content creation, simple compliance drafting, and internal reporting will compress fast.
But “whole job replacement” still requires integrating with messy systems and tolerating liability. Even in the studies showing big productivity effects, the human remains the accountable agent. The tool changes the workflow more than it eliminates the role overnight. NBER Working Paper 31161
So replacement will happen, but as a downstream consequence of interface change: when the workflow becomes clean enough, the organization realizes it no longer needs as many humans in the loop.
Counterargument 2: “This is just another productivity tool. We had email and spreadsheets; jobs kept growing.”
True: many tools amplify output without reducing employment.
The difference is action. Email and spreadsheets did not autonomously decide what to do next and call tools to do it. Agentic systems can.
That means the “unit of automation” is not a single task. It is a chain of tasks.
Once you automate chains, you compress coordination layers: fewer managers, fewer project coordinators, fewer analysts whose job is to assemble status.
Counterargument 3: “If productivity rises, wages rise, and we all win.”
Sometimes.
But wages rise when workers have bargaining power. If AI changes job interfaces such that the worker becomes replaceable, the gains can flow primarily to capital and to the small set of people who control deployment.
This is why the question is not “Does AI increase productivity?” but “Who owns the interface?”
Takeaways
Jobs are bundles of tasks, and AI is changing the interface through which those tasks are executed.
The early, high-leverage wins are in glue work: coordination, drafting, triage, and summarization.
Agentic systems move decision power upstream to whoever defines objectives, constraints, tool access, and evaluation.
The most important labor effect is recomposition: humans shift toward supervision, specification, and audit.
Entry-level ladders are at risk because the tasks that generate reps are the easiest to automate.
“Observed exposure” is a better framing than raw capability: adoption friction and governance matter.
The new scarce skill is workflow ownership: building trustworthy systems, not clever prompts.
Human agency must be designed, measured, and protected, or workers become rubber stamps.
If you want leverage, learn to own interfaces: data access, evals, guardrails, and system design.
Sources
NBER Working Paper 31161: Generative AI at Work — https://www.nber.org/papers/w31161
Noy & Zhang (MIT): Experimental Evidence on the Productivity Effects of Generative AI — https://economics.mit.edu/sites/default/files/inline-files/Noy_Zhang_1.pdf
Anthropic: Labor market impacts of AI (observed exposure) — https://www.anthropic.com/research/labor-market-impacts
World Economic Forum: Future of Jobs Report 2025 (PDF) — https://reports.weforum.org/docs/WEF_Future_of_Jobs_Report_2025.pdf
Stanford SALT Lab: Future of Work with AI Agents — https://futureofwork.saltlab.stanford.edu/
Goldman Sachs: How Will AI Affect the Global Workforce? — https://www.goldmansachs.com/insights/articles/how-will-ai-affect-the-global-workforce
MIT Economics: Acemoglu on the economics of AI — https://economics.mit.edu/news/daron-acemoglu-what-do-we-know-about-economics-ai
IMF blog: New Skills and AI Are Reshaping the Future of Work — https://www.imf.org/en/blogs/articles/2026/01/14/new-skills-and-ai-are-reshaping-the-future-of-work