MariaScales/ai-job-duties

What AI Jobs Ask People to Do: Duty Labels for 1,273 Job Postings

Full page: https://mariascales.com/research/ai-jobs/data/

1,273 job postings, each labelled by what the posting says the hire will do: classic work of the job family, building AI, rolling it out, training people, setting AI strategy, measuring impact, picking vendors, AI governance, working on an AI product, working with clients, managing a team. 64% of the sample are roles with AI in the title, the rest are ordinary roles from matching job families. Word counts tell you a posting mentions AI. These labels tell you what the AI part of the job is.

What is inside

file	what it is
labels.csv	1,273 postings, one row each, twelve duty flags
summary_duties_by_employer.csv	duty shares by employer type, AI roles next to comparison roles
summary_duty_parts.csv	the parts of the duty to bring AI into a company, AI roles against comparison roles
codebook.md	the instructions used for labelling
label_check.csv	150 postings labelled a second time, blind, by the model that made the main labels
label_audit.csv	120 other postings labelled again in an independent blind audit (version 1.0)
label_verify_v11.csv	100 postings labelled blind by a separate labeller under version 1.1

The two rounds

• Round 1 (593 postings). 50 AI roles and 50 comparison roles in each of six job families (seven postings were drawn twice and are kept once): marketing, product, transformation, AI operations, sales and customer success, governance. From company career pages and the 12-country postings (US).
• Round 2 (680 postings). Up to 40 non-engineering AI roles for each region (US; UK; Western Europe; Poland, Serbia, Cyprus and Turkey together) and employer type (tech, consulting, non-tech, AI companies), plus 40 comparison roles per region. From the 12-country postings of October 2026.

1,253 postings had real duties in the saved text. Use only text_quality = duties_found for shares.

How the labels were made

Version 1.0. A language model read the duties section of each posting with the codebook in codebook.md and returned twelve flags, a short main_duty line and ai_share.

Two checks of version 1.0. A blind second labelling of 150 random postings agreed with the main labels on 89% to 99% of postings per code (Cohen's kappa 0.57 to 0.98; the lowest is the broad code for classic work of the job family, and every AI code is at 0.79 or higher); both label sets are in label_check.csv. An independent blind audit of 120 other postings, run in a separate session with no access to the labels, agreed on 87% to 99% of postings per code (kappa 0.59 to 0.98; the lowest is measuring the impact of AI); both label sets are in label_audit.csv. Of the audit's 34 disagreements, 4 were errors in the main labels, 4 were errors in the audit and 26 were cases where the codebook allowed two readings: AI built by a consultancy for one client (product or not), a builder who also ships the tool (rollout or not), customer interviews (client work or not), and product metrics of an AI feature (impact measurement or not).

Version 1.1 (this release). Seven clarifications were added to the codebook to close those readings (end of codebook.md). Five codes were labelled again for all 1,253 postings under version 1.1, by ten labellers working in parallel on separate batches with no access to the earlier labels: duty_ai_product, duty_ai_rollout, duty_customer, duty_ai_measure, duty_ai_strategy. The other seven codes come from version 1.0, with the 2 clear errors the audit found in them corrected. ai_share was aligned with the new codes for 21 postings.

Check of version 1.1. A separate labeller labelled 100 random postings blind under version 1.1 (25 AI roles each at tech companies, consultancies and other employers, plus 25 comparison roles). Agreement on these 100 postings:

code	version 1.1 vs independent labeller	version 1.0 vs independent labeller
works on an AI product	98%, kappa 0.94	91%, kappa 0.71
rolls AI out	98%, kappa 0.96	88%, kappa 0.74
works with clients	97%, kappa 0.93	94%, kappa 0.86
measures the impact of AI	97%, kappa 0.92	94%, kappa 0.85
AI strategy and use cases	98%, kappa 0.96	95%, kappa 0.89

Both label sets are in label_verify_v11.csv. The biggest change from version 1.0 is AI product work at consultancies: under the stricter rule, AI a consultancy builds for one client is client work, and the share fell to 9%.

Three numbers from this data

• An AI title means one of two jobs, and the employer tells you which. At non-tech employers 86% of AI roles bring AI into the company's own work and 11% work on an AI product (236 roles). At tech companies 52% work on an AI product (359 roles). At consultancies 56% work with clients and 9% on an AI product (197 roles).
• The rollout duty. 41% of AI roles roll AI out across teams or processes, against 8% of comparison roles (801 and 452 postings).
• A pure label is rare. In 4% of the labelled AI roles, 35 of 801, the duties hold nothing about AI.

Limits

• Samples built for comparison, with quotas per cell. Shares across the whole sample depend on those quotas; compare inside a cell or across cells.
• Labels come from language models with a written codebook. All three checks are model-made; there is no human gold standard.
• Seven codes (duty_core, duty_ai_build, duty_ai_use, duty_ai_train, duty_ai_vendor, duty_ai_govern, duty_lead) were not labelled again in version 1.1.
• The employer type is wrong for about one employer in five (see the 12-country data set for the check).

Where these numbers are used

• Is AI creating entirely new jobs? I read 17,513 postings. It adds one duty.

Cite this

Atamanova, M. (2026). What AI Jobs Ask People to Do: Duty Labels for 1,273 Job Postings [Data set]. https://mariascales.com/research/ai-jobs/data/

License

CC BY 4.0. Titles and employer names come from public postings. Posting texts are left out of the files; main_duty is a short summary written during labelling.

Columns

labels.csv

column	description
posting_id	Posting id: J plus five digits for the 12-country postings, an internal id for career page rows.
label_round	1: 600 postings sampled by job family (career pages and the 12-country postings, US). 2: 680 postings sampled by region and employer type (the 12-country postings).
source	career_page or postings_12_countries.
country	Country the posting was collected for (ISO code; UK for the United Kingdom). Empty for career page rows.
region	US, UK, Western Europe (DE, FR, NL, ES, IE, PT), Central and Eastern Europe (PL, RS), Cyprus, Turkey.
kind	ai_role (AI in the title) or comparison_role (matching job family, no AI in the title).
employer_group	Employer type: tech, consulting, non_tech, unknown. Career page rows also use ai_vendor (companies whose product is AI). See README.
company	Employer name as published.
title	Job title as published.
title_cluster	Job family assigned by rules on title words. On a check of 150 roles one in ten landed in the wrong family and two more in ten were borderline.
text_quality	duties_found, or no_duties_in_text when the saved text held only navigation, banners or a company blurb. Use duties_found rows for shares.
ai_share	How much of the listed duties are about AI: none, low (under a quarter), mid (a quarter to three quarters), high (over three quarters).
main_duty	What the person mainly does, in 3 to 8 words, written by the labeller.
duty_core	Does the classic work of the job family (campaigns, roadmaps, accounts, audits, process work).
duty_ai_build	Builds AI or automation: agents, prompts, LLM workflows, automations, internal AI tools, models.
duty_ai_use	Uses AI tools in their own daily work.
duty_ai_rollout	Rolls AI out across teams or processes: implementation, change management, integration into how the company works.
duty_ai_train	Trains, coaches or enables other people on AI.
duty_ai_strategy	Sets AI strategy or roadmap, finds and ranks use cases.
duty_ai_measure	Defines metrics or measures the impact or return of AI work.
duty_ai_vendor	Evaluates, selects or manages AI tools, vendors or partners.
duty_ai_govern	AI policy, AI risk and compliance, responsible AI, AI regulation.
duty_ai_product	The work is about an AI product the company sells or builds for customers.
duty_customer	Works directly with external customers or clients.
duty_lead	Manages people or builds a team.
any_internal_ai_duty	1 if any of duty_ai_build, duty_ai_use, duty_ai_rollout, duty_ai_train, duty_ai_strategy, duty_ai_measure, duty_ai_vendor is 1: the job brings AI into the company's own work.

summary_duties_by_employer.csv

column	description
label_round	1: 600 postings sampled by job family (career pages and the 12-country postings, US). 2: 680 postings sampled by region and employer type (the 12-country postings).
kind	ai_role (AI in the title) or comparison_role (matching job family, no AI in the title).
employer_group	Employer type: tech, consulting, non_tech, unknown. Career page rows also use ai_vendor (companies whose product is AI). See README.
n	Number of postings in the cell.
share_ai_product	Share with duty_ai_product, percent.
share_any_internal_ai_duty	Share with any_internal_ai_duty, percent.
share_customer	Share with duty_customer, percent.
share_ai_rollout	Share with duty_ai_rollout, percent.
share_ai_govern	Share with duty_ai_govern, percent.
share_ai_build	Share with duty_ai_build, percent.
share_ai_strategy	Share with duty_ai_strategy, percent.
share_ai_train	Share with duty_ai_train, percent.
share_core	Share with duty_core, percent.
share_ai_share_none	Share where ai_share is none, percent.
small_n	1 if a base in the row is under 15. Read such rows as anecdotes.

summary_duty_parts.csv

column	description
duty_code	Column name of the duty flag in labels.csv.
duty	The duty in plain words.
ai_roles_with_duty	AI roles whose posting assigns this duty (text_quality = duties_found).
ai_roles_total	AI roles with duties found: the base of ai_roles_pct.
ai_roles_pct	ai_roles_with_duty / ai_roles_total, percent.
ordinary_roles_with_duty	Comparison roles whose posting assigns this duty.
ordinary_roles_total	Comparison roles with duties found: the base of ordinary_roles_pct.
ordinary_roles_pct	ordinary_roles_with_duty / ordinary_roles_total, percent.
ai_roles_tech_pct	Share of AI roles at tech companies (AI vendors included) with this duty, percent.
ai_roles_consulting_pct	Share of AI roles at consultancies with this duty, percent.
ai_roles_non_tech_pct	Share of AI roles at other employers with this duty, percent.