---
id: "framework-task-categorization-scoring"
type: "framework"
source_timestamps: ["¶9"]
tags: ["research-methodology", "task-analysis", "llm-as-a-judge"]
related: ["concept-augmentation-score", "entity-chatgpt"]
steps: ["\\\"Compile a comprehensive list of job tasks across occupations (over 19", "000 tasks across more than 900 occupations).\\\"", "\\\"Use OpenAI's ChatGPT to categorize each task", "assessing its specific potential for automation through generative AI.\\\"", "Determine the share of 'exposed' (automatable) versus 'unexposed' (requiring human involvement) tasks for each occupation.", "Construct an 'augmentation score' for each occupation based on this ratio of exposed to unexposed tasks."]
sources: ["reskilling"]
sourceVaultSlug: "hbr-seg-reskilling"
originDay: 10
articleStem: "hbr-edu-35-ai-changing-labor-market"
sourceUrl: "https://hbr.org/2026/03/research-how-ai-is-changing-the-labor-market"
sourceTitle: "Research: How AI Is Changing the Labor Market"
---
# AI Task Categorization and Augmentation Scoring Methodology

The research team used a structured methodology to evaluate the labor-market impact of generative AI by breaking occupations into component tasks and using an LLM to assess automation potential. It is an **LLM-as-a-judge** design and presupposes the [[prereq-task-based-labor-model|task-based model of labor]].

**Steps:**
1. **Compile tasks.** Assemble a comprehensive list of job tasks across occupations — **over 19,000 tasks across more than 900 occupations**.
2. **Classify with ChatGPT.** Use OpenAI's [[entity-chatgpt-d35|ChatGPT]] to categorize each task, assessing its specific potential for automation via generative AI. (Note the dual role of ChatGPT here: it is both the *treatment event* — see [[claim-post-chatgpt-demand-shift]] — and the *measurement instrument*.)
3. **Compute exposure share.** For each occupation, determine the share of **'exposed'** (automatable/assistable) versus **'unexposed'** (strictly human) tasks.
4. **Build the score.** Construct an [[concept-augmentation-score]] for each occupation from the ratio of exposed to unexposed tasks — the number that sorts occupations into [[concept-ai-automation-displacement|displacement-prone]] vs. [[concept-ai-augmentation-complementarity|augmentation-prone]].

**Enrichment / confidence note:** Supported. The working paper ([[entity-displacement-or-complementarity-paper]]) uses a task-based model and an LLM (ChatGPT) to classify tasks by generative-AI exposure across ~900 occupations and ~19k tasks, building occupation-level exposure/augmentation metrics from the exposed/unexposed share. This belongs to a validated methodological family — Anthropic ([[evidence-anthropic-labor-study]]) and the World Bank ([[evidence-world-bank-labor-demand]]) construct comparable LLM-feasibility exposure indices.