How Rolesauce measures AI skills in job postings

The dataset

• 9,026 open job postings from 154 companies, collected on October 17, 2026.
• Source: companies' public job boards on Greenhouse (6,939 postings), Ashby (1,561) and Lever (526).
• Duplicate postings (same company, title and description in several locations) are counted once. Boards we couldn't confidently match to the right company were excluded (7 boards).
• Postings were open on the collection date; their posted/updated dates range from 2017 to 2026, and most are recent.

How we count

We split each posting into sentences, skip hiring-process, legal and benefits boilerplate (for example, notes about using AI to screen applications), and skip sentences that a company repeats across many of its postings (company boilerplate). Then we apply three levels:

1. Mentions AI: at least one remaining sentence names AI, an AI tool or an AI technique.
2. AI expectation: that sentence frames AI as something you should know or do (“experience with…”, “comfortable using…”, “you will…”), not as company marketing.
3. Hands-on use: the expectation is about using AI in your own work (“use AI tools to…”, “leverage AI in your workflow…”), not only selling or building AI products.

Skills and tools are tagged with keyword rules (for example, “ChatGPT”, “Claude”, “Zapier”, “prompt engineering”) and counted only inside AI-expectation sentences. Each role's skill % is the share of all that role's postings. Roles are assigned from the job title (then department) with keyword rules.

Limits (read these)

• Tech-forward sample. Our 154 companies are mostly software, fintech, health tech, marketing and e-commerce companies that use these job-board systems. Everyday employers (hospitals, banks, retailers, local businesses) are under-represented, so read our numbers as “where employers are heading”, not a national rate.
• Keyword rules aren't perfect. In a manual check of 30 postings flagged for hands-on use, 22 were clearly correct and 4 ambiguous (73–87%). That check ran before our latest rule fixes and hasn't been repeated yet. We haven't measured how many we miss.
• Tools are counted conservatively. Company names (Anthropic, OpenAI) aren't tools; a vendor's own postings don't count for its product; and AI assistants named only as a place to rank content (SEO/AEO) don't count.
• One snapshot. This is a single collection date; trend lines come once we refresh the data over time.
• Role buckets are rule-based. A role comes from keywords in the job title. Titles whose keyword says marketing, finance or HR but whose job is something else are moved: engineers and engineering managers to Engineering, designers to Design, product managers to Product, data scientists to Data, lawyers to Legal, workplace/facilities to Operations, and executive assistants to their own group. HR keeps recruiters, HR business partners and compensation roles even when the title names another team.
• Postings say what employers ask for, not how much AI people really use on the job.

Quotes and company names

Role pages show short verbatim sentences from postings without employer names or links. We don't name individual employers on role pages.

Download

Role-level summary (CSV): aggregated numbers only, one row per role.

Changelog

• October 17, 2026: first snapshot (9,026 postings, 154 companies).