← Back to index
Methodology
Version 0.5 · Last updated June 2026

Every week, Pace AI reads hundreds of news items about AI and work. It sorts them into five categories, removes duplicates, classifies each story by type and severity, and calculates a score using fixed rules — no human judgment, no editorial picks.

The result is a single number from 0 to 10 that tells you how much actually happened this week.

Each component tracks a different dimension of AI's impact on work. They're scored independently, then combined.

Labor 30% OF INDEX

The heaviest weight, because it's what most people care about.

Tracks layoffs citing AI, hiring freezes, role eliminations, reskilling announcements, and workforce restructuring. Only stories with explicit AI attribution are counted — a company laying off staff without mentioning AI doesn't register.

Evidence quality matters. A single anonymous claim scores very differently from an official company filing or an independent research report. Multiple verified companies in the same week push the score higher.

Deployment 20% OF INDEX

Tracks real organisations rolling out AI tools to real users. Not announcements. Not pilots. Not "we're exploring AI." Actual deployments with named organisations and measurable scale.

A vague press release about "AI transformation" scores near zero. A named company deploying a specific tool to a specific number of employees scores meaningfully. The system tracks organisations across weeks to avoid counting the same rollout twice.

Capability 20% OF INDEX

This one works differently from the others.

The score comes primarily from benchmark data — standardised tests that measure what AI models can actually do. When a new model enters the top rankings or existing models improve significantly on established benchmarks, the score reflects that.

Separately, the system tracks news about features that shipped to real users this week. These "new capabilities" appear on the site alongside the score, showing what's newly possible in practical terms.

Regulation 15% OF INDEX

Tracks laws passed, executive orders signed, court rulings issued, and enforcement actions taken. Not rhetoric. Not op-eds about what regulation should look like. Actions with legal weight.

A senator giving a speech about AI scores very differently from a court ruling that changes what companies can do. The system also distinguishes between narrow, jurisdiction-specific actions and broad international frameworks. Like Deployment, it tracks regulatory entities across weeks so the same law doesn't get re-scored every time it's mentioned.

Markets 15% OF INDEX

Tracks macro-level market signals — analyst forecasts, aggregate investment data, sector-wide spending reports, and ROI studies. The filter is strict: individual company funding rounds don't qualify. Only aggregate or sector-wide data from named research sources.

This component answers: where is the money going, and what are the people who move money saying about AI's trajectory?

The system uses a two-stage process:

1 Classification. An AI model reads each news item and tags it — what type of story is this, how strong is the evidence, what's the scale, who's involved. The AI does classification only. It doesn't decide what's important.
2 Scoring. A fixed, rules-based system converts those tags into a score. Same tags always produce the same score. There's no AI judgment in this step, no randomness, and no editorial discretion.

This separation is deliberate. The AI is good at reading and categorising. It's not good at deciding what matters. The scoring rules are written by a human and versioned like code.

The five component scores combine into a single index:

ComponentWeight
Labor30%
Deployment20%
Capability20%
Regulation15%
Markets15%

Labor carries the most weight because workforce impact is what working professionals care about most. If a component has no data for a given week, its weight is redistributed proportionally among the others.

The composite is divided by 10 to produce the 0–10 pace reading shown on the site.

The pace reading maps to one of four zones:

Walking Quiet week. Business as usual.
Jogging Things are moving. Worth a glance.
Running Something notable happened. Pay attention.
Sprinting Rare, significant shift.

Zone thresholds are calibrated against a year of historical data to ensure the distribution feels right — Sprinting should be rare, Walking should be common. If every week is Running, the thresholds need adjusting.

Consistency. The same inputs always produce the same outputs. No mood, no bias, no slow-news-day editorial decisions.
Speed. Hundreds of items processed in under an hour, every week.
Transparency. Every story that drives a score is shown on the site with source links. You can check our work.
Separation of concerns. AI classifies. Rules score. Humans calibrate. Each does what it's best at.
We measure attribution, not causation. When a company says "we're laying off staff because of AI," we record that. Whether AI was the real reason or a convenient excuse is beyond what this system can determine. Every score on the site comes with this caveat.
We depend on what gets reported. The system reads news feeds, which means it can only see what journalists choose to cover. Quiet deployments, internal workforce shifts, and regulatory discussions behind closed doors are invisible to us.
Keywords shape what we see. Our news sources are filtered by topic-specific search terms. If AI's impact on work starts showing up in language we're not tracking, we'll miss it until the keywords are updated.
English-language bias. Current feeds are predominantly English-language sources. Significant AI developments in non-English-speaking regions may be underrepresented until picked up by international outlets.
AI reading AI news. Yes, we use an AI model to classify news about AI. We've designed the system to minimise this (the AI classifies, it doesn't score), but the model occasionally misreads tone, conflates similar stories, or misses nuance. We catch what we can through calibration testing.
Some source links are behind paywalls. Stories are sourced from major publications, some of which require a subscription to read in full. The summary shown on the site captures the key facts, so you're not left in the dark — but if you want the original article, a subscription may be needed.

The methodology is versioned. When we change scoring rules, thresholds, or source feeds, we document it and note the version. Historical scores are not retroactively changed — they reflect what the system saw at the time.

We run calibration checks against the full historical dataset whenever we make changes, ensuring scores still make sense across known events.

If a week's score doesn't match what you observed in your industry, we want to hear about it. The system is only as good as its inputs and rules, and both can be improved.

Scores reflect news attribution, not verified causation · paceai.work