WAIBS — World AI Business Survey

Concept

Can we do business surveys with LLMs?

Much of the research in economics and business studies needs data on firms, and gathering that data can be difficult. But a great deal of information about firms has been collected by the builders of large language models, and much of it is now encoded in the models themselves. To make it accessible for research, we need to prompt LLMs systematically.

Surveys

Global AI Innovation Survey GAIIS

Click to read more

Prashant Garg, Peter John Lambert, Ralf Martin & Denis Medvedev

Read the CEP Discussion Paper → Explore the innovation atlas (GAIISia) →

Patents capture only a small fraction of innovation, and firm surveys are costly, so only small samples of firms are ever asked. Yet information on firms' new products and processes is everywhere: company websites, news reports, online forums, social media. Frontier large language models have already absorbed much of it in training, so GAIIS runs an innovation survey through LLM prompts.

~600kfirms surveyed
40countries
~6Mreported innovations, each with a text description

Is it just AI slop? No. Using cross-fitting (no firm is scored by a model that saw its own data), the innovation text predicts productivity beyond inputs, patents and country×sector effects: +1 SD in product innovation is associated with about 5% higher TFP (about 2.5% for process innovation).

  • ~40% of firms show product and over 25% process innovation, against fewer than 4% that patent.
  • 14% have green product and 12% green process innovation, against under 1% with a green patent.
  • Labor-augmenting innovation is about 3× as common as labor-saving for products, 2× for processes.
  • Countries with greener innovation also tilt more towards labor-augmenting innovation.
Map of 5.9 million innovations, coloured by labor effect: blue labor-augmenting, red labor-saving
All 5.9M innovations placed by the meaning of their description and coloured by labor effect: blue = labor-augmenting (3.2M), red = labor-saving (1.1M), grey = unclear. Labor-augmenting innovation dominates everywhere; labor-saving innovation clusters in a few pockets such as CNC/CAM fabrication, payments automation and ERP. Explore in GAIISia →

Paper: CEP Discussion Paper · Data for researchers: GAIIS innobase on Hugging Face (gated; request access with your name and affiliation).

Tools

Explore the data

Click to see the tools
GAIISia Global AI Innovation Survey innovation atlas Explore the Global Innovation Space →

Team

Who we are

Click to see the team

Drag the boxes around · click a box to unfold it