Task-Level Decomposition of a Role

COMPEL Specialization — AITE-WCT: AI Workforce Transformation Expert Article 24 of 35

Role redesign begins, analytically, at the task level. A role is not a monolith; it is a set of tasks, each with its own characteristics — cognitive depth, externally observable output, tool dependency, AI exposure, value generated. Designing a new role by starting from the job title and working down is the approach that produces the generic redesigns the field has come to recognise — role descriptions that read plausibly but do not survive contact with the work. The disciplined alternative is to decompose the role into tasks, classify each task on a small number of governance-relevant dimensions, and aggregate from the classification back to the new role.

This article teaches the decomposition. Articles 25 (role specification) and 29 (performance evaluation) depend on it; without a disciplined task inventory, those downstream artefacts carry whatever biases and omissions the initial role understanding contained. The most consequential risk is the systematic under-counting of knowledge, coordination, and contextual work — the work that is harder to observe and harder to instrument, that does not show up in process maps, and that AI-exposure methodologies frequently miss. A decomposition that counts only the visible, documentable, repeatable tasks produces a redesign that erodes the role’s actual value.

What counts as a task

A task, for redesign purposes, is a discrete unit of work with a recognisable beginning, a recognisable end, a specific output or contribution, and a sufficient time duration to be meaningful (practically, 5 minutes or more of focused effort). The following are tasks: “review a purchase order and approve or reject it”; “draft a customer renewal letter from a template”; “meet with a cross-functional stakeholder to resolve a specification disagreement”; “diagnose why the weekly reporting pipeline failed”; “coach a junior team member on a difficult client conversation.”

The following are not tasks: “be responsive” (a quality, not a task); “maintain relationships” (an outcome, not a task — the underlying tasks include specific conversations, meetings, and communications); “provide thought leadership” (an aspiration, not a task). The distinction matters because aspirations, outcomes, and qualities cannot be classified against AI exposure; tasks can.

The elicitation discipline is therefore linguistic as well as procedural. An incumbent’s first description of their role will usually include several outcomes and qualities. The interviewer’s job is to press gently: “you said you’re responsive — when someone says your team is responsive, what specifically are you doing that they are recognising?” The responses are the tasks.

Eliciting the task inventory

A defensible task inventory comes from three sources in combination.

Incumbent interviews. The role’s actual practitioners, interviewed individually, produce the most granular and most realistic task inventory. Each incumbent is asked to walk through a recent week in detail, naming every substantive activity and roughly how long it took. The interviews are 60–90 minutes; two to three incumbents per role cover the range of practice. The common failure at this step is interviewing only the high-performer; the variance in how different incumbents spend their time is frequently larger than the variance between roles.

Observation. Where the work is observable (in a centre-of-operations context, in a customer-facing context), a half-day of observation catches tasks that incumbents under-report. The under-report is not deceptive; it is cognitive — people do not register routine tasks as tasks when they describe their work. Observation catches what interviews miss.

Output analysis. The outputs the role produces — emails, reports, decisions logged in a system, meetings attended — can be mined to reconstruct a task inventory from the trail. Calendar data, email metadata, document-authorship data. This source is weakest on the coordination and knowledge work that does not produce trackable output and strongest on task recurrence and time allocation.

Triangulating across the three sources produces an inventory that is more robust than any single source. The expert cross-references the three views and investigates discrepancies rather than averaging them.

Classifying tasks on three dimensions

The classification dimensions are chosen for governance utility, not for taxonomic elegance. Three dimensions suffice for most AI-redesign work.

AI exposure. How much of the task’s work is exposed to current AI capability? The exposure methodology of Eloundou et al. (2023), adapted for internal use, classifies on a 0–3 scale: 0 = no current AI capability can do meaningful portions of this task; 1 = AI can do portions with significant human adjustment; 2 = AI can do most of the task with light human review; 3 = AI can do the task end-to-end with only exception handling. Exposure is a property of the task and the current AI state together; it moves as AI capability advances.

Augmentation value. If the task were AI-augmented rather than AI-automated, how much value would the augmentation generate? Augmentation value is a property of the task in its context: a task done 500 times per week with high reliability need has high augmentation value; a task done twice per year has low augmentation value regardless of exposure.

Human-centricity. How much of the task’s value comes from specifically human contributions — judgment in ambiguity, relational work, professional accountability, contextual understanding? A task with high exposure but also high human-centricity is a candidate for augmentation, not automation, regardless of how automatable the mechanical steps are. Human-centricity is the dimension most frequently under-attended; the expert actively weights it because AI-exposure methodologies systematically under-score human-centric content.

Tasks score on a matrix of the three. High exposure × low augmentation value × low human-centricity is the automation candidate. High exposure × high augmentation value × high human-centricity is the primary augmentation candidate. Low exposure × high augmentation value × high human-centricity is the role’s irreducibly-human core. Low exposure × low augmentation value × variable human-centricity is work that is not changing in this redesign.

The coordination-work problem

The task inventory, even with triangulated elicitation, systematically under-counts coordination and knowledge work. This under-count is the classical weakness of process-decomposition approaches and is amplified in AI contexts because AI-exposure methodologies are trained on observable task outputs.

Coordination work includes: resolving ambiguity between stakeholders; synthesising partially-conflicting inputs; explaining a decision to someone who needs to act on it without the full context; recognising when a problem has moved outside the expected frame and needs different attention. These activities appear in calendars as meetings with vague titles and in emails as brief exchanges; in task inventories, they often collapse into “stakeholder management” or disappear entirely.

Knowledge work — the cognitive activity of holding a complex model of a situation in mind, reasoning under uncertainty, identifying what is missing from a set of inputs — is similarly invisible to elicitation. An incumbent describing their week may say “I spent Thursday afternoon on the Acme account situation” without decomposing that afternoon into the specific knowledge-work tasks it contained.

The expert’s corrective: dedicate part of the elicitation to coordination and knowledge work specifically. Ask: “what were you doing when you were not in a meeting and not producing a tangible output?” “What does it look like when you are thinking?” “When you synthesise different inputs into a decision, what does the synthesis feel like; what are you noticing?” The responses expose the coordination and knowledge work; the classification then weights it fairly.

Aggregating to a role view

With the task inventory classified, the aggregation produces a role-level view. The useful aggregations:

• Exposure profile. What fraction of the role’s time is at each exposure level? A role at 40% high-exposure, 30% medium-exposure, 30% low-exposure has different redesign implications from a role at 10% high-exposure, 20% medium-exposure, 70% low-exposure.
• Augmentation-value distribution. Where is the augmentation value concentrated? High-value augmentation in a few tasks can drive substantial role change; dispersed low-value augmentation often does not justify the redesign cost.
• Human-centricity concentration. Where is the irreducibly-human core? The core tasks are the anchor of the redesigned role’s identity.
• Total task time. The sum of task times across the role, compared against working hours, reveals the gap between documented work and actual work (often substantial). The gap is itself diagnostic: a role with 60 hours of task content fitted into a 40-hour week is telling the expert something about hidden work or unsustainable load.

The aggregation feeds directly into Article 25’s role specification.

Avoiding three classification biases

Three biases recur in task classification and distort redesigns.

• Tool-centric bias. The classifier lets the current AI tool define what is exposed. A task whose value would be lost to the specific tool now available is classified as not-exposed; a task whose value would be captured by a better tool is classified as fully-exposed. The bias produces redesigns calibrated to a tool rather than to a capability. The corrective is to classify against plausible capability in the next 12–24 months, not against the specific tool the organisation has licensed.
• Efficiency bias. The classifier over-values speed-and-volume tasks and under-values deliberation tasks. A task that takes 15 minutes and happens 50 times a week looks larger than a task that takes 4 hours and happens once a month, even when the latter produces more value. The corrective is to weight by value as well as by frequency.
• Performance-system bias. The classifier treats the tasks that the current performance system measures as the most important tasks. Tasks that the performance system misses — coaching, cross-team collaboration, mentoring — are under-counted. The corrective is to perform the decomposition independently of the performance instruments and then compare the two.

Two real-world anchors

The Eloundou et al. task-level methodology

Eloundou, Manning, Mishkin and Rock’s 2023 paper GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models (arXiv 2303.10130) developed the task-level AI-exposure methodology that has become the de facto reference. The paper examined task descriptions from the US O*NET occupational database and rated each task on the degree to which a language model, with or without additional tooling, could perform it. Source: https://arxiv.org/abs/2303.10130.

The methodology’s usefulness is in the task-level framing: roles are decomposed into tasks before any exposure claim is made. Its limitations — acknowledged by the authors — include the narrow human rater base, the dependence on O*NET task descriptions which are themselves partially out-of-date, and the inherent difficulty of rating knowledge-work tasks that the task descriptions do not capture well. The expert uses the methodology as a reference for exposure-level semantics and applies it adapted to the specific role with internal elicitation, rather than pulling exposure scores directly from a generic database.

The ILO Working Paper 96 task classification

The ILO Working Paper 96 Generative AI and Jobs: A Global Analysis of Potential Effects on Job Quantity and Quality (Gmyrek, Berg, Bescond, 2023, updated 2024) applied a complementary task-level approach, classifying occupational tasks against GPT capability at multiple levels of augmentation and automation. The ILO analysis emphasises distribution across regions and income levels and provides a publicly available methodology for peer-reviewed replication. Source: https://www.ilo.org/publications/generative-ai-and-jobs-global-analysis-potential-effects-job-quantity-and.

The lesson: the task-level approach has converged from two independent directions (US-academic and ILO-international), reinforcing its status as the operative unit of analysis. Experts comparing internal decompositions against external benchmarks can reference both sources for calibration.

Learning outcomes — confirm

A learner completing this article should be able to:

• Distinguish tasks from outcomes, qualities, and aspirations in an incumbent’s role description.
• Elicit a task inventory triangulated across interview, observation, and output-analysis sources.
• Classify each task on AI exposure, augmentation value, and human-centricity using defensible scales.
• Recognise and correct for the under-counting of coordination and knowledge work.
• Aggregate the classified inventory into a role-level view that drives redesign decisions.
• Identify and correct for tool-centric, efficiency, and performance-system biases in classification.

Cross-references

• EATF-Level-1/M1.6-Art08-Workforce-Redesign-and-Human-AI-Collaboration.md — Core Stream workforce-redesign anchor.
• Article 4 of this credential — role exposure scoring (supplies the exposure methodology).
• Article 5 of this credential — skills adjacency (consumed by the redesigned role specification).
• Article 25 of this credential — redesigned role specification (the downstream artefact).
• Article 29 of this credential — performance evaluation (classifier bias comes partly from performance-system artefacts).

Diagrams

• Matrix — task × AI exposure × augmentation value × human-centricity, with classification cells populated for a sample role.
• HubSpokeDiagram — role at hub; tasks as spokes; each spoke coloured by exposure and weighted by time.

Quality rubric — self-assessment