Agentic AI breaks the assumption behind every task-exposure forecast

Labour-market forecasts have assumed that automation removes tasks one at a time and leaves the occupation standing. Agentic AI, which completes whole workflows, invalidates that assumption, and a new composite score, Agentic Task Exposure (ATE), is built to measure what the old frameworks can't.

Executive Summary

• Earlier forecasts of automation, including those built on the Autor-Levy-Murnane task framework, share one assumption: the occupation survives as a coordination shell around the remaining human work.
• Agentic AI systems satisfy four criteria: they take a goal specification, autonomously select and invoke tools, maintain state across inference steps, and produce a completed work product without human intervention at intermediate steps.
• The ATE score combines AI capability scores, workflow coverage factors and logistic adoption velocity. Credit analysts, judges and sustainability specialists reach ATE scores of 0.43-0.47.
• The paper applies ATE to 236 SOC codes in six major occupational groups across five US technology regions, and describes itself as the first multi-regional empirical application of the task exposure framework specifically targeting agentic AI systems.
• Alongside displacement, the paper identifies seventeen emerging occupational categories, concentrated in human-AI collaboration, AI governance and domain-specific AI operations.

The forecasts that overshot #

Industrial robots did displace manufacturing workers. They didn't do it on the timeline or at the scale that the initial forecasts implied. That sentence should make anyone who reads labour-market predictions for a living a little wary, because it describes a pattern: the direction was right and the magnitude was off.

The best-known headline number in this literature comes from Frey and Osborne, who asked experts to score 702 US occupations on overall "computerizability." Their result was that 47% of occupations were at high risk. It's a striking figure, and it travelled far.

Arntz, Gregory and Zierahn then showed what happens when the analysis drops to the task level and accounts for task heterogeneity within nominally exposed occupations. High-risk estimates fall to 9-14% across OECD countries. The distance between 47% and 9-14% isn't a rounding disagreement. It reflects a choice about the unit of analysis, and the choice moves the answer by a factor of three or more.

The source is candid that the debate between occupation-level and task-level estimates is unresolved. What the field did reach is a practical consensus: task-level decomposition is necessary, and occupation-level labels are too coarse. That consensus is the foundation the rest of the argument stands on. It's also the foundation that agentic AI puts under strain.

How the task frameworks work #

To see where the strain comes from, it helps to be precise about what the task tradition claims.

Autor, Levy and Murnane partitioned occupational tasks along two axes: routine versus non-routine, and cognitive versus manual. Their argument was that automation substitutes for the routine while complementing the non-routine. It's a clean idea, and it gave later work a vocabulary for talking about which parts of a job a machine could absorb.

Acemoglu and Restrepo built a formal structure on top of that. They model output as resulting from a continuum of tasks allocated between capital and labour. In their framing, the labour share rises or falls depending on whether new-task creation, which they call the reinstatement channel, outpaces task displacement. Following them, the labour share in value added is written sL = W * L / Y = f(|NL| / |N|), where W is the wage rate, L is employment and Y is output. Two forces pull against each other, and which one wins determines who captures the gains.

Two other measurement efforts matter here because they turned the framework into occupation-level numbers. Felten, Raj and Seamans built the AI Occupational Exposure (AIOE) score by linking published AI capability advances to the O*NET ability taxonomy. Eloundou et al. constructed a parallel measure focused on GPT-4, combining crowd-worker annotations with GPT-4 self-assessments at the task level.

Their headline finding was that 80% of US workers hold jobs where at least 10% of tasks are LLM-exposed. The heaviest exposure was concentrated among high-wage, degree-requiring occupations. That result runs against the older intuition that automation hits routine, lower-wage work first, and it explains why the white-collar question has become urgent.

The blind spot every framework shares #

The source makes a pointed claim about all of this work. Every cited framework shares an architectural blind spot: they model automation as picking off individual tasks while the occupation stays intact. Almost all forecasts, including those built on the Autor-Levy-Murnane task framework, share the assumption that automation picks off tasks one at a time and the occupation survives as a coordination shell around the remaining human work.

That's an assumption about structure, not about technology. It says nothing regarding how capable the machine is. It says that however many tasks the machine takes, someone still has to hold the remainder together.

The paralegal makes the point concrete. A task-exposure analysis might flag 80% of a paralegal's tasks as LLM-exposed and yet conclude the occupation is safe, because coordinating those tasks demands human judgement. The conclusion follows from the shell assumption. The tasks are exposed, but the stitching between them isn't, so the job stays.

Then the source's key move: the coordination bottleneck that protected the paralegal role is now inside the agent's planning module. If that holds, the reasoning that kept the paralegal safe stops working. The protection came from the coordination. The coordination is exactly what an agentic system does.

I think this is the most useful idea in the paper, and it deserves to be taken on its own terms. The older frameworks aren't wrong about the tasks. They're wrong about what remains after the tasks go. A measure that scores tasks one by one can't see a system that handles the sequence as a unit.

What counts as agentic #

If the argument depends on a distinction, the distinction needs a definition. The source supplies one with four criteria. An agentic AI system is one that:

1. receives a goal specification rather than a task specification;
2. autonomously selects and invokes tools to pursue the goal;
3. maintains state across multiple inference steps; and
4. produces a completed work product without human intervention at intermediate steps.

This is a list worth keeping as a list, because the criteria are discrete and each one can be checked. A system that answers a prompt fails the first. A system that needs a person to approve each intermediate step fails the fourth.

In plainer terms, an agent receives a goal, selects tools, executes a multi-step plan, checks its own output and delivers a finished work product. Agentic systems chain tool calls, maintain persistent memory across steps, plan multi-action sequences and self-correct when intermediate outputs fail quality checks.

The self-correction detail matters more than it first appears. A pipeline that breaks when one step fails still needs a human nearby. A system that notices a failed quality check and tries again has absorbed some of the supervision that used to justify the human's presence.

The paper defines its subject this way: autonomous AI agents capable of completing entire occupational workflows rather than discrete tasks. The word "entire" carries the weight. The whole argument is about the gap between a discrete task and a complete workflow.

Why the timing is now #

The source offers two kinds of evidence that the question isn't hypothetical. The first concerns capability. GPT-4-class models reached at or above passing-score thresholds on the bar exam and the US Medical Licensing Exam, and came within a few percentage points of passing CFA Level I. Those are credentialing exams for professions that restrict entry precisely because the work is hard.

The second concerns deployment. Major enterprise technology vendors, including Amazon, Microsoft, Salesforce, ServiceNow and Workday, disclosed production agentic deployments touching white-collar workflows in their 2024 earnings calls and investor presentations. These are production deployments, not research demos, and they were announced to investors rather than to academics.

The source reads those disclosures as the kind companies make only when deployments are already affecting headcount planning. The source frames that as an interpretation, not a measurement, and it should be held as one. Nothing in the record I've been given reports headcount figures directly. But the inference is reasonable: firms don't usually put a deployment in front of investors unless it matters to the cost base.

Put the two together and the case for a new measure is straightforward. Capability alone wouldn't force the issue, and neither would a few pilots. Capability that is already crossing professional thresholds, in systems that are already in production, does.

The ATE score: what it measures and what it doesn't #

The paper's response is the Agentic Task Exposure score. It's a composite metric that extends the Acemoglu-Restrepo task displacement framework to account for end-to-end workflow coverage, current agentic AI capability benchmarks and enterprise adoption velocity. The score measures the degree to which an occupation o is exposed to end-to-end displacement by current agentic AI systems.

Three ingredients go in: AI capability scores, workflow coverage factors and logistic adoption velocity. The workflow coverage factor is the piece that directly addresses the blind spot. The source says what's missing from earlier work is a measure of whether the full workflow composition of an occupation sits inside agentic AI's operational envelope, and it formalises that as the workflow coverage factor. The older scores ask whether a task is exposed. This one asks whether the whole workflow fits inside what an agent can do.

Two further pieces of the construction are specified. The term wo,t is the importance-weighted relevance of task t in occupation o, drawn from O*NET task importance ratings. And CAP(t), the capability term, is estimated from AI benchmark literature and practitioner deployment reports. So task importance comes from a standard occupational database, while capability comes from two sources: benchmarks and field reports.

The paper is explicit about method, and a reader should hold it to that. ATE scores are computed algorithmically from O*NET task data using calibrated adoption parameters; they are not regression estimates. That has practical consequences. A regression estimate comes with the usual apparatus of standard errors and identification arguments. A calibrated algorithmic composite doesn't, and its outputs are model-derived projections, a phrase the paper itself uses for its displacement figures. They describe what the model implies under its parameters. They don't report what happened.

I'd argue that's an honest framing, and it's the right one for a forecasting problem where the thing being forecast hasn't occurred yet. A composite built from stated inputs can be inspected, challenged and recalibrated. If someone disagrees about adoption velocity, they know which dial to move.

Where the model applies it #

The empirical application is broad. The paper applies the ATE framework to 236 Standard Occupational Classification (SOC) codes spanning six major occupational groups, generating model-derived displacement projections across five major United States technology regions for the 2025-2030 period. The five regions are the San Francisco Bay Area, Seattle-Tacoma-Bellevue, Austin-Round Rock, New York-Newark, and Boston-Cambridge.

The paper describes this as the first multi-regional empirical application of the task exposure framework specifically targeting agentic AI systems. That's a claim of priority made by the authors, and I can't check it from the material in front of me. What I can say is that the regional design is a sensible response to a problem the national-level studies share: exposure is a property of occupations, but adoption is a property of places and firms.

On the results the source gives, credit analysts, judges and sustainability specialists reach ATE scores of 0.43-0.47. The grouping is worth a second look. A credit analyst, a judge and a sustainability specialist don't resemble each other in training, pay or day-to-day work. What they have in common, on the paper's account, is a score in the same range. I'd read that as a mark against any lazy sorting of jobs into "safe" and "at risk" by sector or by perceived creativity.

I'd also resist reading more into the figure than the source supports. The material I have gives the range for these three occupations and doesn't say how it compares with the rest of the 236 codes. So the number shows these roles are substantially exposed on the paper's scale. It doesn't show they're the most exposed.

The regional findings come with a hedge worth repeating. The paper states that Seattle's workforce effects in 2026-2027 preview what the broader national non-technical workforce will experience in 2028-2030. That's a prediction with a stated lag, and it's hedged. It reads as a hypothesis the model produces, and it shouldn't be taken as an established fact. But it's a useful kind of claim, because it's specific enough to be wrong. If Seattle's early pattern doesn't carry over to the national non-technical workforce by 2028, the model has a problem it can't explain away.

The reinstatement side of the ledger #

A displacement score on its own tells half of the Acemoglu-Restrepo story. In their framework, the labour share depends on whether new-task creation outpaces displacement. A paper that extended the framework and measured only displacement would have abandoned the part of it that gives the framework its tension.

The paper doesn't. It identifies occupational categories where agentic AI creates reinstatement effects: new roles that did not exist before agentic deployment creates demand for human oversight, governance and collaboration at scale. In total it identifies seventeen emerging occupational categories benefiting from reinstatement effects. They are concentrated in human-AI collaboration, AI governance and domain-specific AI operations roles.

The mechanism is easy to follow. If agents carry out workflows end to end, someone has to oversee them, set the rules they run under, and work alongside them. Those needs hadn't existed as jobs because there was nothing yet to oversee.

The source doesn't tell me how many people these seventeen categories would employ, or whether their growth would offset the displacement the ATE scores imply. So the data I can use leaves the central question open: whether reinstatement outpaces displacement. Anyone who writes confidently that it does, or that it doesn't, is going past the verified claims. The paper's contribution is to put both sides of the ledger into the same framework and to name where the new roles appear.

What to take from it #

There's a methodological lesson and a practical one.

The methodological lesson is that the unit of analysis decides the answer. The jump from 47% to 9-14% came from moving from occupations to tasks. The paper argues that a further move is now needed, from tasks to workflows, because the thing being automated is no longer a task. Each time the unit shifted, the estimates changed, and the field's history suggests that the choice of unit deserves as much scrutiny as the choice of data.

The practical lesson is for anyone planning headcount, training or products around the next five years. A score that tells you 80% of a role's tasks are exposed isn't enough, because it doesn't say whether the remaining 20% is a durable human core or just the glue that an agent's planning module can now supply. The workflow coverage factor exists to ask that question directly.

The ATE score is a composite with calibrated parameters, and its projections are model-derived. They're a structured argument about 2025-2030, not a record of it. That's a reason to test them rather than ignore them. The Seattle hypothesis offers an early test: if the 2026-2027 effects there look the way the model implies, the lag to the national non-technical workforce in 2028-2030 becomes the next thing to watch. And the seventeen reinstatement categories pose the question the claims can't settle: whether the roles created to oversee agents will arrive in numbers that matter to the labour share.

Questions? Answers.

References #

• Ravish Gupta and Saket Kumar, (2026). Agentic AI and Occupational Displacement: A Multi-Regional Task Exposure Analysis of Emerging Labor Market Disruption ⧉. arXiv.
• Tyna Eloundou, Sam Manning, Pamela Mishkin, and Daniel Rock, (2023). GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models ⧉. arXiv.
• Daron Acemoglu and Pascual Restrepo, (2019). Automation and New Tasks: How Technology Displaces and Reinstates Labor ⧉. Journal of Economic Perspectives.
• National Center for O*NET Development, (2025). O*NET Resource Center ⧉. US Department of Labor.

Last reviewed 2026-11-01.