How Is AI Changing Jobs Right Now?

1. We pull each occupation's core tasks from O*NET

The O*NET database defines 8–15 tasks per occupation, ranked by importance and frequency. For example, a cashier's tasks include “receive payment” and “issue receipts.” We select the tasks that cover 80% of the role's importance weight.

2. An AI search scores each task against shipped products

We send the task list to Perplexity with a structured rubric. Perplexity searches the web for commercially deployed AI systems — not research papers or demos — and scores each task based on what it finds. Every non-zero score must cite a named product, a named company, and a source URL. Scores are then adjusted for evidence quality and deployment maturity (a pilot-stage tool counts less than an industry-standard one).

3. The scoring rubric

4. Task scores are classified into a tier

The adjusted task scores are combined with workforce signals (evidence of actual job displacement, role consolidation, or headcount changes) to classify each occupation into a tier: Replacement, Reshaping, Augmentation, or Minimal Change. A high task score alone does not trigger “Replacement” — there must also be evidence that the AI product is deployed broadly and matches the core work output. Inspired by Andrej Karpathy's framework.

Task selection

For each occupation, we query the O*NET Web Services API for task statements, importance ratings (IM scale), and frequency ratings (FT scale). We select tasks covering 80% of cumulative importance weight, capped at 8–15 tasks. If importance/frequency data is unavailable, we fall back to all “Core” tasks.

AI research via Perplexity

Each occupation's task list is sent to Perplexity's sonar-pro model with pro search enabled and high search context. The prompt includes the 7-point rubric (−3 to +3), disambiguation hints (alternate job titles, workflow-specific search terms), and domain exclusions (O*NET, BLS, job boards). Perplexity returns a score, evidence quality rating, deployment scope, and core-output match for each task, plus a list of products and workforce signals.

Quality-adjusted scoring

Raw task scores are not used directly. Each task's contribution is adjusted by three factors: core output match (does the AI product perform the task's core output, or just adjacent admin?), deployment scope (pilot = 0.45× weight, industry standard = 1.0×), and evidence quality (weak = 0.55×, strong = 1.0×). This prevents pilot-stage or loosely-matched tools from inflating scores.

Tier classification

The quality-adjusted task scores are aggregated into workflow shares: penetration share (any AI involvement), assistive share (positive scores), AI-first share (negative scores with human oversight), and autonomous share (full replacement). These shares, combined with workforce displacement signals (e.g., “headcount reduction,” “cashier-free stores”), determine the final tier through a threshold-based classifier.

Validation

Research results that fail validation — hollow retrieval (no meaningful citations beyond O*NET/BLS), schema failures, or missing task coverage — are quarantined and excluded from the treemap. Occupations without valid research appear as “Score pending.”