Why most AI pilots stall, and how to avoid it

Most AI pilots stall because the data behind them was never organized for the job, and because no one defined what “working” would look like before starting. The research on this is now blunt: MIT’s Project NANDA reported that “95% of organizations are getting zero return” from generative AI, and Gartner has predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. Neither report blames the underlying models.

This is for owners, managing partners and COOs who’ve either tried an AI pilot that fizzled, or want to avoid running one that will.

What the research actually found

MIT’s Project NANDA report, based on 52 executive interviews, a survey of 153 leaders, and analysis of roughly 300 public AI deployments, reported: “Despite $30–40 billion in enterprise investment into GenAI, this report uncovers a surprising result in that 95% of organizations are getting zero return.” Its explanation for why: “most AI tools don’t learn and don’t integrate well into workflows.” (MIT NANDA, “The GenAI Divide: State of AI in Business 2025,” July 2025)

Separately, Gartner predicted at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, unclear business value, inadequate risk controls and escalating costs (Gartner press release, July 29, 2024).

The pattern behind both numbers

Read together, three things show up repeatedly:

• The data wasn’t ready. A pilot pointed at scattered, uncleaned information produces answers no one trusts, so it never leaves the pilot stage.
• No one defined success first. Without a specific outcome agreed in advance, a pilot can run for months without anyone being able to say whether it worked.
• It stayed generic. A tool that doesn’t know your firm’s procedures, clients and standards produces work that looks plausible but isn’t usable without heavy rework — which erases the time savings that justified the pilot.

MIT’s report also found that “external partnerships with learning-capable, customized tools reached deployment ~67% of the time, compared to ~33% for internally built tools.” The authors note these percentages come from their interview sample of 52 organizations and may not represent the broader market.

What a firm can do differently

• Fix the data first. A pipeline that connects and cleans your firm’s and clients’ information is the difference between a demo and something your team will actually rely on.
• Name the outcome before you start. Decide what you’re measuring (hours saved on a specific task, turnaround time, error rate) and how you’ll know in a defined window.
• Scope it narrow. One process, not a department-wide rollout, so a stall is cheap and a success is easy to point to.
• Keep a person in the loop. A result nobody has to trust blindly is a result people will actually use.

Where Precision AI OS fits

Precision AI OS is built around the failure mode the research describes: we build the data pipeline first, so agents work from your firm’s real information instead of guessing, and every output goes to a person on your team to approve before it reaches a client. See how we protect your data or book a discovery call.