MIT researchers dropped a bomb this year: despite $40 billion spent on generative AI over two years, only 5% of enterprises can point to measurable business returns. The rest? Stuck in what the study calls "pilot purgatory"—endless experiments that generate enthusiasm in demos but deliver nothing in quarterly results.
This isn't a technology problem. It's a strategy problem. The 5% who succeed aren't smarter or better funded. They're just doing something fundamentally different from everyone else.
The GenAI Divide: Winners vs. The Stuck
The MIT study introduced a concept that should terrify every CIO and CFO: the GenAI Divide. On one side, a narrow band of organizations that turned pilots into infrastructure. On the other, the vast majority trapped in permanent experimentation.
The difference? Successful companies built systems that learn in the wild, improve with each cycle, and align with how people actually work.
Here's what that looks like in practice. Johnson & Johnson eliminated hundreds of scattered GenAI experiments once they discovered that roughly 10-15% of them produced about 80% of the value. They shifted toward giving individual departments—supply chain, research, operations—the freedom to own the AI process and tools that aligned with their actual work.
Not the shiny stuff. The mundane stuff that moves the needle.
Why Your AI Pilot Will Die in Production
The early stages of an AI pilot move fast. Demos generate excitement. Teams imagine how the system might fit into daily workflows. The interface feels responsive. Early outputs look clean.
Then it hits production, and everything falls apart.
The tool behaves differently in real workflows. It forgets recent instructions. It repeats avoidable mistakes. It needs context that people already provided. After a few weeks, teams start working around it. The project still exists on paper, but progress stops.
Omar Shanti, CTO of Hatchworks AI, nailed it: "Generative AI projects are easy to do but hard to do well. It's easy to get to the pilot phase, but getting to production is an elusive goal for most enterprises."
The MIT study captured this pattern in brutal detail. When systems fail to learn, they stop earning attention. Each session starts from scratch. Feedback evaporates. Teams lose confidence because the system doesn't evolve with use.
It behaves like a new product every time it loads.
Inside the 5%: What Winners Do Differently
Companies that escape pilot purgatory share a common trait: they begin with workflows that deliver real value from day one. Instead of chasing trends, they anchor their efforts in tasks that actually matter.
High-Impact Use Cases That Actually Work:
- Invoice processing that shortens vendor disputes. Not sexy. Saves millions.
- Support agents that draft responses and improve with use. Not trendy. Measurably better retention.
- Supply chain optimization that reduces waste by identifying bottlenecks. Not flashy. Clear ROI in 90 days.
These systems rarely impress in demos. Their value appears in metrics the following quarter.
Deloitte's analysis confirms this. Their case studies show that focusing on a limited set of high-impact use cases—especially those layered on existing workflows—accelerates ROI. Centralizing governance helps too, ensuring integration and scalability without overextending resources.
MIT's research paints the same portrait. The groups that produce real value tend to partner with external vendors skilled in context and domain fluency. These partnerships deliver value roughly twice as often as in-house builds.
Why? Because vendors bring production-hardened systems that have already failed, been fixed, and learned from hundreds of deployments. You're not starting from scratch. You're starting from someone else's expensive lessons.
The Real Problem: Shadow AI and Missing Context
Here's the uncomfortable truth: the most widely used generative AI tools in the enterprise today are the same ones workers use on their phones during lunch. ChatGPT, Microsoft Copilot, and Claude see daily use across teams—often without formal approval, training, or integration.
This is what researchers call "Shadow AI." Employees experiment because these tools are easy to adopt and don't require permission. They help summarize documents, clean up emails, brainstorm ideas.
But ease of use creates a dangerous illusion. Organizations believe they're adopting generative AI because their employees are experimenting with it. What they're actually doing is outsourcing small fragments of knowledge work to systems that remain disconnected from core operations.
Celonis executives hit this point hard in a recent interview. "AI has not delivered fully on its promise at the operational level because AI does not understand how a business really works," said Malhar Kamdar, Chief Growth Officer at Celonis. "The business context it needs to operate is scattered across systems, departments, and people."
95% of companies are doing something with AI, but most of it remains at the productivity layer—writing emails, summarizing documents, making small tasks easier. Truly transforming supply chains, finance operations, or enterprise workflows requires AI to understand operational context.
That gap between AI enthusiasm and business outcomes is becoming a boardroom concern. CIOs are under pressure to justify AI spending amid limited measurable returns.
What the 5% Know That You Don't
The companies that succeed follow a clear pattern:
1. They launch inside clear processes. Not broad, transformational initiatives. Narrow, repeatable workflows with well-defined success criteria.
2. They prioritize systems that learn. Every interaction makes the system smarter. Feedback doesn't evaporate—it becomes training data that improves future outputs.
3. They measure from day one. Not engagement metrics. Not user satisfaction scores. Real business metrics: cost per transaction, time to resolution, error rates, vendor dispute duration.
4. They give departments ownership. Centralized AI teams set governance and infrastructure. Individual departments own use cases and tools that align with their work.
5. They accept that most experiments will fail. J&J killed hundreds of pilots. The 10-15% that worked delivered 80% of the value. The rest got shut down fast, before they became expensive distractions.
Meredith Broussard, NYU professor and author of "Artificial Unintelligence," puts it this way: "Is email useful? Yes. Has it totally eliminated handwritten materials? No. Use GenAI the same way. Focus on the mundane, not the shiny, and then you'll make better decisions."
The CFO Question: Can You Quantify Value Creation?
Dilipkumar Khandelwal, former CIO at Deutsche Bank and now Chief Customer Officer at Celonis, believes enterprises are moving from "AI experimentation" to "return on AI."
"There are a lot of enterprises investing in AI, but the biggest question remains: can you quantify the value creation?" he said. "Process intelligence provides an X-ray of how a business runs. Once you can identify inefficiencies, automate workflows, and deploy AI agents with the right context, ROI becomes measurable."
Celonis claims its global customers have generated over $25 billion in value creation through process intelligence deployments, with more than $10 billion already quantified by customers themselves.
That's the standard now. Not "we're experimenting with AI." Not "we're piloting in three departments." The question from the CFO is simple: what's the measurable business return?
If you can't answer that in dollars saved or revenue generated, your AI project is already dead.
What to Do Monday Morning
If you're a technical leader (CIO, CTO, VP Engineering):
1. Kill 80% of your AI pilots. Identify the 10-15% delivering real value. Shut down the rest. Fast.
2. Pick one high-volume, repeatable workflow. Invoice processing, support ticket triage, vendor dispute resolution. Something boring that happens hundreds or thousands of times per month.
3. Measure baseline performance today. Time to resolution, cost per transaction, error rate. If you can't measure it, you can't prove ROI.
4. Deploy a system that learns. Not a static model. A system that retains feedback, improves with use, and accumulates context over time.
5. Set a 90-day ROI checkpoint. If the metrics aren't improving by 10-20%, kill it and move to the next use case.
If you're a business leader (CFO, CMO, COO):
1. Demand quantifiable value. Not engagement metrics. Not "users love it." Dollars saved or revenue generated.
2. Give department heads ownership. They know their workflows better than any centralized AI team. Let them own the use cases.
3. Centralize governance, not execution. Security, compliance, vendor management—that's central. Use case selection and tool ownership—that's departmental.
4. Budget for failure. 80-85% of experiments won't work. That's fine. The 15% that work will deliver outsized returns—if you shut down the failures fast enough.
The Bottom Line
$40 billion spent. Only 5% see returns. The GenAI Divide is real, and it's widening.
The companies that succeed aren't chasing the shiny stuff. They're not building transformational AI visions. They're finding boring, high-volume workflows, deploying systems that learn, and measuring ROI from day one.
They're killing pilots that don't work within 90 days. They're giving departments ownership. They're partnering with vendors who've already made the expensive mistakes.
And they're answering the CFO's question with real numbers: dollars saved, revenue generated, time recovered.
That's the difference between the 5% and everyone else.
The question is: which side of the divide are you on?
Continue Reading
Looking for more insights on enterprise AI strategy and implementation? Check out these related articles:
- How to Build an AI Strategy That Actually Delivers ROI
- The Hidden Costs of Shadow AI in Your Organization
- Enterprise AI Governance: What Works in 2026
Rajesh Beri is Head of AI Engineering at a Fortune 500 security company. He writes THE DAILY BRIEF, a twice-weekly newsletter on Enterprise AI for Technical and Business Leaders. Follow him on LinkedIn or Twitter.
