Why AI Projects in Factories Die in Month Nine

The demo just ended. The model answers everything, the dashboard glows, even the most skeptical executive can’t find a flaw. The CEO says: sign it.

Nine months later, the project is dead.

Not with a bang — with a whimper. The system still runs, servers still hum, but no business unit actually uses it. The vendor delivered every contracted feature; the enterprise paid every invoice. The only thing that never arrived: value.

MIT’s NANDA lab tracked this pattern in their 2025 report The GenAI Divide: 52 organizations interviewed, 153 executive surveys, 300+ public AI projects reviewed. Their conclusion — roughly 95% of the billions poured into enterprise GenAI since 2021 produced nothing reportable on a financial statement. (Their “failure” definition — no measurable P&L impact within six months — is debatable, but RAND and S&P Global surveys paint the same picture.)

Nothing on the failure list is technical

The report’s top barriers to scaling AI, by frequency:

1. Employees won’t use the new tools — the same people who use ChatGPT daily on their own;
2. Concerns about output quality;
3. Poor user experience;
4. Lack of executive support;
5. Change management difficulties.

Notice what’s missing: the model isn’t smart enough, compute is too expensive, the tech isn’t ready. None of them. Projects die at problem definition and organizational reality — not in the algorithm.

Three shop floors, one answer

Airbus: A recurring fuel-pump fault at the Toulouse plant stumped Airbus engineers for two years. Palantir engineers plugged sensor data into the platform and cracked it in two weeks — fuel was sloshing away from the pump during climb. (Reported by Challenges.)

US Navy: A $448M contract to modernize submarine shipyards. At Electric Boat, scheduling dropped from 160 person-hours to under 10 minutes. (Reported by USNI News.)

John Deere: OpenAI FDEs flew to Iowa, followed agronomists into fields, and built an evaluation rubric of what a “good” spray recommendation means — before touching the model. Chemical usage down up to 70%. (OpenAI’s published case.)

Three countries, three industries, one conclusion: the model is never the bottleneck. The engineering that connects it to reality is.

What that engineering actually does

• Turn vague wishes into measurable problems (“planner spends 2h every morning reconciling three systems” — not “improve efficiency”);
• Fish the data out of three systems that don’t talk to each other;
• Define what “good” means, numerically, before build starts;
• Make the first users want it — adoption is behavioral, not administrative.

Three questions before you call any vendor

1. Which role spends half its time on repetitive document work?
2. Can the rules of that work be written down?
3. Can you state that role’s fully-loaded annual cost without looking it up?

Pass two of three — worth a serious look.

(Cai Yulin · AI Engineer for Manufacturing ｜ WeChat: caishide ｜ h2-home.com)