Loading...
Patterson Consulting Contact Us

The Task Structure of Work

By: Josh Patterson
October 2026

Key Idea:

A job is a bundle of tasks, and a task is a sequence of steps, each with inputs, an output, and a share of the worker's time. Automation usually takes over a few steps at a time inside tasks that stay human. Because the steps in a workflow depend on one another, automating one step changes the value of the rest. To predict what automation will do to a role, map its steps and how they connect.

In 2013, Carl Frey and Michael Osborne assessed whole occupations and estimated that 47 percent of US employment was in occupations at high risk of automation, meaning an estimated probability above 70 percent.[1]

In 2016, Melanie Arntz, Terry Gregory, and Ulrich Zierahn repeated the exercise using the tasks individual workers reported doing. The share of US employment at high risk fell to 9 percent, the same as the average across 21 OECD countries.[2]

The two studies looked at the same economy and the same technology. Frey and Osborne treated each occupation as a single unit. Arntz and her coauthors looked inside occupations, where workers with the same title do different things.

The unit of analysis changed the answer by a factor of five.

Breaking Work Into Smaller Units

Adam Smith opened The Wealth of Nations with a pin factory. The business of making a pin, he wrote, "is, in this manner, divided into about eighteen distinct operations." Ten workers splitting those operations made "upwards of forty-eight thousand pins in a day." A worker untrained in the trade, working alone, "could scarce, perhaps, with his utmost industry, make one pin in a day, and certainly could not make twenty."[3]

Smith drew his eighteen operations from earlier French descriptions of pin-making, including Alexandre Deleyre's article "Épingles" in the Encyclopédie (1755).[4]

Dividing work has a cost. Gary Becker and Kevin Murphy showed that specialization is limited by the cost of coordinating specialists, in addition to the size of the market.[5] Every handoff between two people, or between a person and a machine, has to be coordinated.

Peter Drucker listed six factors that determine knowledge-worker productivity. The first: "Knowledge-worker productivity demands that we ask the question: 'What is the task?'"[6] On a factory line the task is usually defined in advance, while in knowledge work someone has to define it.

Tasks in the Autor Framework

Labor economists made the task their basic unit. David Autor, Frank Levy, and Richard Murnane sorted tasks by whether a computer could carry them out by following explicit rules (routine) or not (nonroutine).[7] Later work defined a task as "a unit of work activity that produces output" and grouped tasks into three categories:[8]

Abstract Tasks

Problem-solving, persuasion, and design. Computers complement these tasks.

Routine Tasks

Tasks that follow explicit rules, such as bookkeeping or repetitive assembly. These are the tasks computers substitute for.

Manual Tasks

Tasks that need situational adaptability, such as driving in city traffic. Until recently, machines did little to replace or assist this work unless the environment could be controlled.

Autor and Michael Handel found that tasks vary within occupations as well as between them. The tasks workers report predict their wages even after controlling for occupation.[9]

The US government's occupational database, O*NET, captures the same idea in a cross-reference structure: each occupation has its own task statements, which are linked in a many-to-many crosswalk to a standardized taxonomy of detailed work activities; those roll up through intermediate work activities to 41 generalized work activities at the most abstract level.[10]

O*NET cross-reference structure: 923 occupations each have 18,838 occupation-specific task statements, which are cross-referenced many-to-many to 2,087 detailed work activities; those roll up to 332 intermediate work activities, which roll up to 41 generalized work activities at the most abstract level. 923 Occupations 18,838 Task statements (total; ~20 per occupation) cross-referenced (many-to-many) 2,087 Detailed work activities roll up to 332 Intermediate work activities roll up to 41 Generalized work activities
The O*NET cross-reference structure, with counts from the O*NET 31.0 database (August 2026). Occupations have occupation-specific task statements; task statements are cross-referenced many-to-many to a standardized taxonomy of detailed, intermediate, and generalized work activities.
923Occupations
18,838Task statements
2,087Detailed work activities
332Intermediate work activities
41Generalized work activities

Below the Task: Workflows and Steps

Autor's examples of tasks, such as analyzing legal documents or managing people, each contain a sequence of smaller actions. Analyzing a legal document means retrieving it, identifying the relevant statutes, comparing precedents, and forming a judgment. A task in this sense is closer to a workflow.

This series uses a finer unit, the step: the smallest part of a workflow that can be assigned to a person or a machine, with defined inputs and an output. A role contains tasks, each task is carried out through a workflow, and each workflow is a sequence of steps.

Task

A unit of work activity that produces output, such as a decision, a document, or a service.

Workflow

The ordered sequence of steps by which a task is done today.

Step

The smallest part of a workflow that can be assigned to a person or a machine, with defined inputs and an output.

The step formalizes a unit already present in the literature.

Autor (2015) names it directly: "automation or computerization makes some steps in a work process more reliable, cheaper, or faster."[12] Kremer's (1993) O-ring production model supplies the formal version: each link in the chain has inputs, an output, and a quality level, and the chain's total output is the product of each link's quality.[11] The step, as defined here, is the unit that chain operates over. Gans and Goldfarb (2026) apply this logic directly to automation decisions — which links to hand to machines, what happens to the value of the remaining human links, and why automation choices tend to come in bundles rather than one step at a time.[13]

The step level matters because it is where automation actually enters knowledge work. A role can score as predominantly non-routine — and appear safe from automation at Autor's occupation-level measure — while several of its constituent steps quietly become automatable. The job title persists; headcount does not immediately fall; but the hours per task shrink as those steps migrate to machines and the remaining workflow reconstitutes around what is left. Each subsequent article in this series — on cognitive labor, on the value of human expertise, on the labor automation cycle — depends on this step-level view of how technology changes what workers actually do.

For illustration, an insurance claims analyst adjudicating a property claim might work through these steps:

  1. intake and classify the claim
  2. pull the policy and its coverage terms
  3. gather evidence such as photos, adjuster notes, and invoices
  4. check the loss against the coverage
  5. estimate the payout
  6. screen for fraud indicators
  7. decide, document, and communicate the decision

Each step can be described by six attributes that together determine how automation will affect it and what happens to the rest of the workflow when it changes.

Attribute Why it matters for automation analysis
Share of the task's time Sets the labor-hours at stake. A step that consumes a small fraction of the task's time has limited productivity impact whether automated or not; the steps worth targeting are the ones where the hours are concentrated.
Boundedness of output Steps with tightly bounded outputs — a classification, a yes/no, a value from a known range — are the primary candidates for rule-based or model-driven automation. Open-text outputs require generative capability and are harder to verify at scale.
Whether its rules can be written down Autor's (2003) routine/non-routine distinction, applied at the step level. A step whose complete rules can be specified is the classic automation target. Steps requiring tacit judgment — weighing incommensurable risks, reading context not present in the data — resist substitution even when the surrounding workflow does not.
Cost of an error Determines the oversight burden after automation. A cheap, immediately visible error tolerates light review. An expensive or delayed error requires sampling and human sign-off, which can consume a significant share of the time the automation saved — sometimes all of it.
Expertise required From Autor and Thompson (2025): automating inexpert steps raises the remaining bundle's average expertise level and wages. Automating expert steps removes the scarce skill that commands a premium, lowering the role's wage floor. The wage direction turns entirely on which steps are targeted.
How strongly its quality affects the other steps Kremer's (1993) O-ring coupling: a step whose failure degrades the whole task's output is a bottleneck. Making a high-coupling step more reliable raises the value of every remaining human step in the chain — this is the mechanism behind Autor's (2015) observation that automation increases the value of the remaining human links.

The output of a step is work, and the hours spent on it are labor, so the distinction from "What Is Work?" becomes measurable inside a single task.

Why Steps Change the Automation Picture

A job can score as mostly non-routine while several of its steps become automatable. The title persists and headcount does not fall at first.

Each task takes less time, and the job gradually reorganizes around the remaining steps.

The steps are also linked. Michael Kremer's O-ring model treats output as the product of the quality of each step, so one weak step can spoil the rest.[11] Autor applied the idea to automation:

"when automation or computerization makes some steps in a work process more reliable, cheaper, or faster, this increases the value of the remaining human links in the production chain."[12]

Joshua Gans and Avi Goldfarb show that when a step is automated, the worker moves time to the remaining steps and does them better. Automation decisions then tend to come in bundles, and labor income can rise under partial automation. They conclude that what predicts the outcome is the structure of bottlenecks in a workflow, and the average exposure of its tasks predicts much less.[13]

Illustrative step map for adjudicating a property claim: intake, policy lookup, and evidence gathering are automated; the coverage check and the final decision are retained; payout estimation and fraud screening are augmented. 1 Intake and classify AUTOMATE 2 Pull policy and terms AUTOMATE 3 Gather evidence AUTOMATE 4 Check coverage RETAIN 5 Estimate payout AUGMENT 6 Screen for fraud AUGMENT 7 Decide and communicate RETAIN
Automate: a machine performs the step Augment: a person performs it with a tool that speeds it up Retain: a person performs it
An illustrative step map for the claims task. Arcs mark steps whose quality strongly affects a later step. Real decisions depend on the firm's data, rules, and tolerance for error.

Bundling also sets who can do the job. Autor and Neil Thompson point out that a worker has to be able to perform every task in an occupation that has not been automated, so the most expert remaining task sets the bar for entry. Automating one step can raise or lower that bar.[14]

The capability frontier of AI also runs between steps. Fabrizio Dell'Acqua and colleagues gave 758 BCG consultants realistic tasks with and without GPT-4. On tasks inside the frontier, consultants using AI completed 12.2 percent more tasks, 25.1 percent faster, at more than 40 percent higher quality. On a task just outside it, they were 19 percentage points less likely to reach a correct answer.[15] Tasks that looked equally hard fell on different sides of the line.

For a CTO evaluating an automation proposal, the practical starting point is a step map of the role, with time shares and dependencies, drawn before anyone estimates savings.

References

  1. Frey, C. B., & Osborne, M. A. (2017). The future of employment: How susceptible are jobs to computerisation? Technological Forecasting and Social Change, 114, 254–280. (Oxford Martin School working paper, 2013.) doi:10.1016/j.techfore.2016.08.019
  2. Arntz, M., Gregory, T., & Zierahn, U. (2016). The Risk of Automation for Jobs in OECD Countries: A Comparative Analysis. OECD Social, Employment and Migration Working Papers, No. 189. OECD Publishing. doi:10.1787/5jlz9h56dvq7-en
  3. Smith, A. (1776). An Inquiry into the Nature and Causes of the Wealth of Nations. Book I, ch. 1.
  4. Peaucelle, J.-L. (2006). Adam Smith's use of multiple references for his pin making example. European Journal of the History of Economic Thought, 13(4), 489–512. doi:10.1080/09672560601025829
  5. Becker, G. S., & Murphy, K. M. (1992). The division of labor, coordination costs, and knowledge. Quarterly Journal of Economics, 107(4), 1137–1160. doi:10.2307/2118383
  6. Drucker, P. F. (1999). Knowledge-worker productivity: The biggest challenge. California Management Review, 41(2), 79–94, at pp. 83–84. doi:10.2307/41165987
  7. Autor, D. H., Levy, F., & Murnane, R. J. (2003). The skill content of recent technological change: An empirical exploration. Quarterly Journal of Economics, 118(4), 1279–1333. doi:10.1162/003355303322552801
  8. Acemoglu, D., & Autor, D. (2011). Skills, tasks and technologies: Implications for employment and earnings. In O. Ashenfelter & D. Card (Eds.), Handbook of Labor Economics (Vol. 4B, pp. 1043–1171). Elsevier. doi:10.1016/S0169-7218(11)02410-5
  9. Autor, D. H., & Handel, M. J. (2013). Putting tasks to the test: Human capital, job tasks, and wages. Journal of Labor Economics, 31(2, Pt. 2), S59–S96. doi:10.1086/669332
  10. Peterson, N. G., et al. (2001). Understanding work using the Occupational Information Network (O*NET): Implications for practice and research. Personnel Psychology, 54(2), 451–492. doi:10.1111/j.1744-6570.2001.tb00100.x Counts from the O*NET 31.0 database (August 2026): 41 generalized work activities, 332 intermediate work activities, 2,087 detailed work activities, and 18,838 task statements across 923 data-level occupations. National Center for O*NET Development.
  11. Kremer, M. (1993). The O-ring theory of economic development. Quarterly Journal of Economics, 108(3), 551–575. doi:10.2307/2118400
  12. Autor, D. H. (2015). Why are there still so many jobs? The history and future of workplace automation. Journal of Economic Perspectives, 29(3), 3–30, at p. 6. doi:10.1257/jep.29.3.3
  13. Gans, J. S., & Goldfarb, A. (2026). O-Ring Automation. NBER Working Paper 34639. doi:10.3386/w34639
  14. Autor, D., & Thompson, N. (2025). Expertise. Journal of the European Economic Association, 23(4), 1203–1271. doi:10.1093/jeea/jvaf023
  15. Dell'Acqua, F., McFowland, E., Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2023). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. Harvard Business School Working Paper 24-013. doi:10.2139/ssrn.4573321 Published in revised form in Organization Science, 37(2), 403–423 (2026), doi:10.1287/orsc.2025.21838. The figures cited here are from the working paper.