The tooling gap

When researchers at OpenAI scored how much of each job a large language model could touch, they reported two numbers. Alpha counts the tasks the model can affect on its own — paste text in, read text out. Gamma counts the tasks it can affect once you also build software around the model — retrieval, tools, agents, integrations. The distance between them is the whole story of the last two years.

Employment-weighted across the U.S. workforce the same split holds: the model alone reaches 14% of tasks, software-on-the-model reaches 60%. Across 947 occupations with both scores.

Where the tooling does the work

These occupations are barely exposed to the chatbot itself, yet nearly every task becomes exposed once software is built around the model. This is where "AI" means the product, not the prompt — and where the agentic build-out, not the base model, moves the line.

exposed to the model alone (α) added once software is built around it (γ − α)

Where the model itself is enough

The mirror image: text-in, text-out work the language model touches directly, with little extra needed from tooling. Here the base model — the thing you can use today in a chat window — already overlaps most of the tasks.

How to read this. Exposure is the share of an occupation's tasks that overlap with what AI can do — a measure of potential reach, not automation, adoption, or jobs lost. Alpha (E1 in Eloundou et al.) counts tasks a large language model affects with direct access alone; gamma (E1 + E2) adds tasks affected once complementary software is built on top of the model. A task being "exposed" means AI could cut the time to do it by at least half — it says nothing about whether anyone has built that software, deployed it, or chosen to use it.

← Back to AI exposure across all of work