Andreessen Horowitz (a16z) · Article

Where Enterprises Are Actually Adopting AI

Revenue data on where enterprise AI is real: coding first, then legal, customer support and search.

Published
Reading time
About 12 min
Authors
Kimberly Tan
Read the original

What is it about?

Using private revenue data shared by leading AI startups, a16z finds that 29% of the Fortune 500 and about 19% of the Global 2000 are live, paying customers of a leading AI startup. It is a direct counterpoint to the MIT “95% fail” narrative.

The more interesting argument is why adoption lands where it does. Coding leads because code is text, precise and verifiable, and engineers can adopt tools without cross-department sign-off. Support works because it is already outsourced and measured, with a human fallback. Adoption follows verifiability and low coordination cost, not model capability alone.

Key numbers

  • Startup revenue by use case: coding about $3 billion, legal $500 million, support $400 million, health administration $350 million, search $250 million.
  • Models beat human lawyers in only 36% of GDPval legal tasks (down from 46%), yet legal is the second-largest revenue category.
  • The five traits of fast-adopting work: text-based, repetitive, natural human review, limited regulation, verifiable output.

Why read it

  • The five-trait test is a practical filter for choosing your first AI use cases.
  • It explains why customer support is the lowest-risk agent pilot: the worst case is a hand-off to a human.
  • It pairs well with the MIT NANDA report as a debate about what “adoption” actually means.

Who should read it

  • CIOs choosing a first portfolio of AI use cases
  • Customer-support and BPO leaders
  • Strategy teams sizing enterprise AI markets

About the author: Kimberly Tan

Investing partner, Andreessen Horowitz

Tan joined a16z in 2019 and invests in companies using AI to rethink large legacy markets. She introduced Decagon's two founders to each other at an a16z retreat, wrote the first check and sits on its board. Decagon builds AI customer-support agents, one of the use cases this article highlights.

Before a16z she worked at McKinsey, the World Bank and Goldman Sachs and spent time at a journalism nonprofit in Bosnia and Herzegovina. She graduated from Stanford with distinction and also wrote a16z's thesis “Unbundling the BPO”.

Read it carefully: common misquotes

  • “29%” means paying customers of a startup under a top-down contract, not company-wide use or proven ROI.
  • The data comes from startups, many of them in a16z's own portfolio, and the sample size isn't disclosed.
  • It measures something different from the MIT report, so it does not “debunk” it.

How it shapes the CIO's AI Playbook

Notes by Pavel Veselov, last reviewed . Summaries are my own; quotes and figures belong to the original authors.

All 11 sources in the CIO's AI Library