Back to the dashboard

Sources, models and method behind the AI impact forecast 2027

31 figures from 17 sources, and seven models. Each figure is graded by how it reached us; the coverage column shows which economy it describes.

How AI capability is measured

The capability tile asks one question: how much real work can AI finish with no human fixing it? The main gauge is the Remote Labor Index, where AI agents attempt real paid freelance projects and human judges reject any deliverable a client wouldn't accept. Two supporting gauges sit beside it: how long AI can work unsupervised and still be right 4 times in 5 (METR), and how much AI is speeding up its own research, a sign of recursive improvement. No accepted measure of AI memory at work exists yet.

Reliability grades

Official source · 6A statistics agency, central bank or international body, read from its own release.
Independent research · 9A university, research lab or polling firm with a published method.
Press-reported · 7Official or research data that reached us through coverage, a tracker or a single secondary source. Not checked against the original.
Vendor survey · 6A survey or dataset from a company that sells related products.
Our arithmetic · 3Our arithmetic on cited figures.
Model outputOutput of a stated mathematical model fitted to the readings. Not a published forecast.

Prediction models

AI capability: logistic S-curve
y(t) = K / (1 + e^−(a + b·t))
slope b = 2.910 per year · ceiling K from 50–100% · residual σ = 0.340 · 3 readings · 1,500 Monte Carlo draws

The Remote Labor Index is the closest measure of work done without a human: real paid freelance projects, judged by people against professional work, where any error a client would reject is a fail. Scores are bounded, so they follow an S-curve. The ceiling is uncertain because some client work may never pass review, so K is drawn between 50% and 100%. Only verified readings are used: two later readings (21% in Oct 2026 and 41% in Nov 2026) are left out because each comes from a single unconfirmed source.

Fitted to: Remote Labor Index, three verified readings, Oct 2025 – Jul 2026.

Business adoption: logistic diffusion (three curves)
F(t) = K / (1 + e^−(a + b·t))
All firms: slope b = 0.548 per year · ceiling K from 45–75% · residual σ = 0.080 · 3 readings · 1,500 Monte Carlo draws. Large firms: K 75–95%. Small firms: K 40–70%.

Technology adoption across firms follows an S-curve (Rogers; Bass). Firm adoption has passed the 16% line into Rogers' early majority. Large firms adopt first and have the higher ceiling.

Fitted to: OECD all-firm rates and Eurostat size classes, 2023–2025.

AI at work: logistic diffusion (two curves)
F(t) = K / (1 + e^−(a + b·t))
Any use at work: slope b = 0.732 per year · ceiling K from 55–90% · residual σ = 0.080 · 3 readings · 1,500 Monte Carlo draws, residual σ raised to 0.25 for three readings. Daily use: K 30–60%, same σ floor.

Generative AI spread to workers about twice as fast as the PC or the internet (Bick, Blandin & Deming). Three readings over six months fit almost perfectly, which would give falsely narrow bands, so the noise term is raised to an honest floor. The ceiling starts at 55% because a Federal Reserve survey found worker use roughly flat (40.7% to 40.5%) over the same period, so a plateau is possible. Daily use has a lower ceiling because many jobs have few tasks AI helps with.

Fitted to: Quarterly employee panel, Nov 2025 – May 2026.

AI job cuts: lognormal flow, running total
cuts(month) ~ LogNormal(μ, σ); total(t) = total(t − 1) + cuts(t)
μ = 8.58 (median about 5,336 a month) · σ = 1.01 · 3,000 simulated paths

The tile shows a running total since January 2025: each month's cuts are added on, so the number only grows and never resets. Announced cuts are lumpy, so monthly totals are modelled as random draws around a level that gives more weight to recent months. The level faded through late 2026, so the model projects lower totals than the 2026 peak. A new wave would sit at the top of the band.

Fitted to: Monthly AI-cited job cuts, Jan–Oct 2026.

AI-centric roles: decaying growth
I(t) = I₀ · exp(g₀ · (1 − e^(−λh)) / λ)
starting growth drawn from +35% to +95% a year around the measured +69% (g₀ = 0.525) · λ drawn from 0.3–1.2 (growth halves in 7–28 months) · all ads grow 0.086 per year

Demand for a new skill grows fastest while it's scarce and slows as it becomes normal, as with computer skills in the 1990s. There is only one measured year of growth, so both the starting pace and how fast it slows are uncertain; this is a hypothetical grounded in that pattern, with low confidence.

Fitted to: Global AI Jobs Barometer, 2025 and 2026.

Exposure by experience: damped trends
y(h) = L + b · (1 − e^(−λh)) / λ
Fresh grads: L = −17.9%, b = −4.36 pts/yr · Mid-career: L = +1.6%, b = +0.45 · Seniors: L = −0.9%, b = −0.25 · λ = 0.51 (trend halves in about 16 months)

Damped trends beat straight lines in forecasting competitions because shocks fade. The fresh-grad trend uses the measured 4-point widening over the year; mid-career and senior trends use the average pace since 2022.

Fitted to: Payroll study vintages Aug 2025 and Aug 2026.

Junior jobs economy-wide: firm-level elasticity
Δjunior(t) = −δ · A(t), A(t) = 16% × F(t) / F(2025)
δ from 0.08–0.09 (junior employment 8–9% lower at adopting firms) · A = share of employment at adopting firms, scaled by the business-adoption curve

A hypothetical built from science: a firm-level effect measured on 65 million workers, scaled by how much of the economy adopts AI.

Fitted to: Hosseini & Lichtinger (May 2026); business adoption model.

How the outlook was built

  1. Eleven evidence threads

    Global bodies (ILO, OECD), the EU, the UK, the US and Canada, plus a review of forecasting models and their published parameters.

  2. Global first, largest dataset where no global series exists

    Where a worldwide or OECD-wide series exists, the tile uses it. Where none exists (AI at work, AI job cuts, exposure by age), the tile uses the largest dataset available and the coverage column says which economy it is.

  3. Models, then uncertainty

    S-curves are fitted by least squares on the logit scale across a range of ceilings, then Monte Carlo draws vary the ceiling, slope and noise. Bands show the 10th to 90th percentile.

  4. Hypothetical where data run out

    Junior jobs and AI-role growth use models built from published science, labelled low confidence.

Every figure

FigureWhat it measuresDateCoverageSourceGrade
30.5% · 41.2% · 55.0%Large firms (250+ staff) using AI, 2023 · 2024 · 20252023–2025EU-27EurostatOfficial source
6.4% · 11.2% · 17.0%Small firms (10–49 staff) using AI, 2023 · 2024 · 20252023–2025EU-27EurostatOfficial source
8.7%Firms (10+ staff) using AI, OECD economies2023OECD economiesOECDOfficial source
14.2%Firms (10+ staff) using AI, OECD economies2024OECD economiesOECDOfficial source
20.2%Firms (10+ staff) using AI, OECD economies
The 2026 reading is due around Dec 2026.
2025OECD economiesOECDOfficial source
4% to 6%Firms using AI that say it cut their headcount
4% (Dec 2025), 5% (Mar), 6–7% (Jun), 6% (Sep). Admin and clerical roles named most affected by 38% of AI users, up 9 points. Under 1% say AI added jobs.
Dec 2025 – Sep 2026UKONS · Business Insights and Conditions SurveyOfficial source
2.5%Real freelance projects AI completed end to end at a client-acceptable standard
Human judges compare each AI deliverable with work a paid professional delivered. Any error a client would reject counts as a fail.
Oct 2025 (launch)GlobalCenter for AI Safety & Scale AI · Remote Labor IndexIndependent research
4.17%Real freelance projects AI completed end to endEarly 2026GlobalCenter for AI Safety & Scale AI · Remote Labor IndexIndependent research
15.8%Real freelance projects AI completed end to end
Some coverage quotes 16.1%.
1 Jul 2026GlobalCenter for AI Safety & Scale AI · Remote Labor IndexIndependent research
46%Employees who have used AI at work
Share who used AI at work a few times a year or more. The largest quarterly employee panel on AI at work; no global quarterly series exists, so this panel is the trend line. A Federal Reserve survey with a different question found worker use roughly flat (40.7% Nov 2025, 40.5% Aug 2026).
Fielded 30 Oct – 14 Nov 2025USGallupIndependent research
50%Employees who have used AI at work
Daily use 13%.
Fielded 4–19 Feb 2026USGallupIndependent research
52%Employees who have used AI at work
Daily use 15%. Other economies: UK 55% of employees use AI for work or education (ONS, May–Jun 2026); Canada 35.9% of workers in the past year (Statistics Canada, Mar 2026); EU 15.1% of all adults used generative AI for work in 2025 (Eurostat).
May 2026USGallupIndependent research
1.9×Speed-up in AI research at one lab, mostly attributed to AI doing research work
Range 1.5–2.8×; 1.4× when other labs are included. The lab redacted most of the evidence. This is the closest measure of AI improving itself.
Published 5 Nov 2026One labMETRIndependent research
−17.9% · +1.6% · −0.9%Employment change since Nov 2022 in the most vs least AI-exposed jobs: ages 22–25 · 35–49 · 50+
35–40: +0.1%, 41–49: +3.1% (averaged to +1.6%); 26–30: −4.8% (not significant). ADP payroll data on millions of workers. Descriptive, not causal. The OECD calls cross-country evidence of harm to young workers limited so far.
Data to Jun 2026, published 12 Aug 2026US payroll recordsStanford Digital Economy LabIndependent research
Why compareWhy this tile compares AI-exposed roles with roles AI barely touches
Since late 2022, hiring everywhere moved with interest rates, tariffs and the post-pandemic slowdown. Looking only at AI-exposed roles would mix that general slowdown in with AI. Taking out the change seen in roles AI barely touches leaves the part linked to AI. Nov 2022 is the starting point because that is when generative AI tools became widely available. A separate study of 65 million workers finds the same pattern inside firms: junior jobs 9% lower at AI-adopting firms, senior jobs flat.
MethodUS payroll recordsStanford Digital Economy Lab · Hosseini & LichtingerIndependent research
21.25%Real freelance projects AI completed end to end
From one blog only. Not used on the dashboard.
18 Oct 2026GlobalCenter for AI Safety & Scale AI · Remote Labor IndexPress-reported
41.25%Real freelance projects AI completed end to end
Headline on the benchmark's site; the Center for AI Safety dashboard shows 41.3%. The leaderboard itself would not load, so this is unverified and not used on the dashboard.
About 9 Nov 2026GlobalCenter for AI Safety & Scale AI · Remote Labor IndexPress-reported
54,836Announced job cuts citing AI, all of 2025
Job cuts where the employer named AI as a reason. A stated reason, not a measured cause; some firms may cite AI to please investors.
2025USChallenger, Gray & ChristmasPress-reported
38,579Announced job cuts citing AI, May 2026 (peak month)
Job cuts where the employer named AI as a reason. A stated reason, not a measured cause; some firms may cite AI to please investors.
May 2026USChallenger, Gray & ChristmasPress-reported
121,401Announced job cuts citing AI, Jan–Oct 2026
Job cuts where the employer named AI as a reason. A stated reason, not a measured cause; some firms may cite AI to please investors. Monthly values other than May, July and October are subtractions from cumulative totals.
Jan–Oct 2026USChallenger, Gray & ChristmasPress-reported
12%Employees who use AI at work daily
Reported in press coverage of the Gallup release, not read on Gallup's page.
Fielded 30 Oct – 14 Nov 2025USGallupPress-reported
~3 hLength of task AI completes unsupervised and gets right 4 times in 5
About 1 h in Dec 2025. METR said tasks above 16 h can't be measured reliably and froze its page on 8 Sep 2026.
Tested Mar 2026GlobalMETRPress-reported
739Job listings titled 'AI orchestrator' or similar
Up 1,294% from near zero a year earlier, about 1% of agent-related postings.
Feb 2026USLightcast via IT BrewVendor survey
#1AI engineer, fastest-rising job title, second year runningJan 2026GlobalLinkedIn Jobs on the RiseVendor survey
~300 to 5,000+Forward-deployed engineer job listings2025–2026USLinkUpVendor survey
14% / 39%Firms saying AI reduced their workforce last year / expect AI-related cuts next year
1,719 respondents, mostly large firms.
May–Jun 202697 countriesMcKinsey · State of AI 2026Vendor survey
18.8%Working-age people worldwide using AI (any purpose)
16.3% in H2 2025, 17.8% in Q1 2026.
Q2 2026GlobalMicrosoft AI Economy Institute · AI Diffusion ReportVendor survey
+69% / +9%Growth in job ads requiring AI skills vs all job ads, year on year
Wage premium for AI skills 62%, up from 57%.
Published 15 Jun 202627 markets, 1bn+ job adsPwC · Global AI Jobs BarometerVendor survey
Running totalAnnounced job cuts citing AI, added month by month since January 2025
Our running sum: 54,836 in 2025, plus each 2026 month. 176,237 by October 2026.
Jan 2025 – Oct 2026USChallenger, Gray & ChristmasOur arithmetic
36–55%Workers using AI at work in other economies, on different survey questions
UK 55% of employees use AI for work or education (ONS, May–Jun 2026). Canada 35.9% of workers used generative AI at work in the past year (Statistics Canada, Mar 2026). EU 15.1% of all adults used generative AI for work in 2025 (Eurostat; counts all adults, not just workers, so it is lower). The questions differ, so these are not plotted on the line.
Mar – Jun 2026UK, Canada, EUONS · Statistics Canada · EurostatOur arithmetic
19–30%Share of firms using AI in national official surveys, mid-2026
UK 30% of all firms (Sep 2026, ONS, official); US 23.8% (Oct–Nov 2026, Census data via a tracker); Canada 19.2% (spring 2026, Statistics Canada, official). Each asks the question differently, so we show the range, not an average.
Q2–Q3 2026UK, US, CanadaONS · US Census Bureau · Statistics CanadaOur arithmetic