Too many companies confuse buying AI with adopting AI.

A new report from MIT’s NANDA initiative (Networked Agents and Decentralized AI) analyzed 300 public AI deployments, and found that only 5% saw rapid revenue acceleration (mostly at startups).

This issue isn’t technology. It’s adoption.

The majority of AI application budgets are going to sales and marketing tools, but the biggest ROI came from back-office automation–in other words the boring stuff that people hate to do, but is unavoidably important.

Many enterprises prefer to build their own AI for security reasons, but this approach only succeeds 22% of the time, in comparison to 67% for buying a commercial tool or partnering with dedicated AI supplier.

Many enterprises have established central AI labs, but the projects that succeed tend to be led by line managers–the people actually doing the work.

The problem is that far too many leadership teams treat AI as a magic pixie dust. The Board of Directors tells the CEO, “We need to become an AI-first organization.” The CEO tells the CTO and CIO, who then launch a project to select and procure “AI”. It’s a top-down “push” approach driving primarily by fear, and a little bit by greed.

A better approach is to build decentralized AI familiarity within different departments and business units, so that the people actually doing the work can figure out how AI can help, and then “pull” AI into those departments and business units with the help of IT and its centralized AI expertise.

Don’t buy AI unless you have frontline managers and users who were part of the technology selection process who are eager to adopt the new technology they helped pick.