I recently read MIT’s State of AI in Business 2025 report from Project NANDA (Aditya Challapally, Chris Pease, Ramesh Raskar, and Pradyumna Chari). Thanks to Michael Burnett for sharing it. The full report is here:

👉 MIT Research Paper

The headline is stark: despite $30–40 billion of enterprise investment into generative AI, 95% of organisations are seeing no measurable return. MIT call this the GenAI Divide — a small minority extracting genuine business value while the rest remain stuck in endless pilots.

Here are the parts that really resonated with me:

1. Why Most Pilots Fail

MIT analysed 300 public AI initiatives and interviewed 52 organisations:

• 80% explored LLMs like ChatGPT or Copilot.
• 40% deployed them in some form.
• Yet only 5% of custom enterprise tools made it into full production.

What struck me most is that the barrier isn’t infrastructure or regulation, but a learning gap. Most systems can’t retain feedback, adapt to context, or integrate naturally into workflows. That felt like the key dividing line: tools that “forget” remain gimmicks, while those that learn can actually shift the economics.

2. The Shadow AI Economy

Another powerful insight is that while only 40% of companies purchase official LLM subscriptions, 90% of employees already use AI personally at work. That disconnect leapt out at me.

It reminded me of what we’ve seen in other industries: transformation often starts at the edges. Employees hack together their own solutions long before IT makes it official. This “shadow AI economy” may be where the real adoption curve begins.

3. Where ROI Really Sits

MIT highlight that roughly half of GenAI budgets flow into sales and marketing — because those functions are visible and easy to measure. Yet the more compelling ROI examples came from back-office automation: eliminating BPO contracts, cutting external agency spend, streamlining reconciliations.

That resonated because it’s less glamorous, but far more defensible. The big winners will be those who deliver utility in these overlooked areas.

4. What Successful Startups Actually Do

The report makes clear that the most effective AI startups don’t boil the ocean. They:

• Go narrow, embedding deeply in specific workflows.
• Deliver quick, visible wins.
• Expand outward only once embedded.
• Build trust through networks, not cold outreach.

It resonated because it’s not about who has the flashiest demo — it’s about who solves real, messy problems inside processes that matter.

5. Looking Ahead: Agentic AI

Finally, MIT’s point on Agentic AI stuck with me. Systems with memory and adaptability that can orchestrate workflows autonomously. They describe the emerging Agentic Web, where agents transact and collaborate without human prompts.

For me, this feels like the logical next platform shift. The report underlined that the window is closing quickly — enterprises are already embedding these systems, and switching costs will rise every month from here.

Closing Thought

What I took away most is that the GenAI Divide is very real — but also very crossable. The gap isn’t about hype or capability; it’s about learning, memory, and integration. Those who solve for that will be the ones who break through.

Have a great week!