The GenAI Divide: Why Most Pilots Stall and How Yours Won’t

MIT NANDA’s State of AI in Business 2025 report has sparked headlines about “95% of AI projects failing.” That number is real but it applies to task-specific enterprise GenAI. Meanwhile, the same research claims general-purpose LLMs (think ChatGPT/Copilot) show far better traction, with ~40% of orgs reporting success.

What the data actually says

- Devil in the detail: General-purpose LLMs are widely piloted and deployed (40%). Task-specific GenAI stalls only ~5% reach production.
- Good signal for Tech & Media: Despite the hype, AI shows up clear structural disruption mainly in Tech and Media so far out of 9 researched industries.
- Start simple: AI has already “won” simple work. 70% prefer AI for drafting emails, 65% for basic analysis. For complex, multi-week work, humans still dominate 9:1.
- Approach matters: The divide isn’t driven by model quality or regulation; it’s driven by workflow fit, learning, and memory.
- Where ROI hides: Early, measurable savings come from reducing BPO and external agency spend especially in back-office operations more than headcount cuts.

Playbooks that work for organizations (from the 5% who succeed)

- Land small, visible wins in narrow workflows; then expand.
- Choose tools with low configuration and immediate visible value. Initially avoid heavy enterprise customization.
- Prioritize learning, memory, and workflow adaptation not flashy UX.
- Partner to build pilots: AI pilots via strategic partnerships are 2x as likely to reach full deployment as internal builds.

Research Report Link: https://lnkd.in/g3UCZ8Nr