A July 2025 MIT study shook the enterprise AI world: 95% of AI pilots deliver no measurable ROI. Despite $30–40B invested, most corporate projects stall in proof-of-concept limbo, leaving boards and executives questioning the real value of GenAI.
At Firemind, we see this “GenAI Divide” every day. Enterprises are eager to experiment, but too many build isolated pilots that never scale. MIT’s findings mirror what we hear from executives: AI feels exciting in the lab, but nothing shifts in operations.
Why pilots fail
The MIT report identifies structural reasons:
- Limited disruption: Only Tech and Media see real transformation.
- Enterprise paradox: Big firms run the most pilots, but struggle to scale.
- Investment bias: Budgets chase marketing hype, while real ROI hides in operations and finance.
- Implementation advantage: Vendor-led builds succeed twice as often as in-house attempts.
Most strikingly, the study documents a “shadow AI economy.” Employees turn to personal ChatGPT or Claude accounts because corporate tools are clunky, underpowered, or not trusted. This disconnect creates risk and lost opportunity.
Firemind’s perspective: Run focused experiments
The gap isn’t inevitable. We believe organisations can escape the pilot trap by running focused, value-driven experiments that scale.
To do that, we use the IFD framework:
- Intensity: How painful or costly is the problem?
- Frequency: How often does it occur?
- Density: How many people or processes are affected?
Experiments should target challenges that score high on IFD. They should be low-cost to start, value-driven from day one, and designed to scale if successful.
Proof in practice
Our work demonstrates how this approach turns pilots into real value:
- In the legal sector, a generative AI assistant reduced routine casework by 20%—equivalent to giving every lawyer a full day back each week.
- In the insurance industry, document review times dropped by 95%—from hours to minutes, unlocking both scale and accuracy.
- In financial services, mortgage document verification now happens in 5–7 seconds per file, balancing compliance with customer experience.
These aren’t experiments stuck in labs. They are delivering measurable ROI today.
The path forward: Agentic AI
MIT concludes the next phase of AI will be defined by Agentic AI, systems that don’t just respond, but remember, adapt, and act. At Firemind, we’re already testing agentic models through AWS SageMaker and Bedrock, building tools that act as co-workers rather than static assistants.
For enterprises, the message is clear: AI success won’t come from dabbling in pilots. It comes from focused, IFD-driven experiments, trusted vendor partnerships, and systems built to scale.
The GenAI Divide is real but it can be bridged. Our partners are already on the other side.
Find the original report here.
Do you want to talk more about successful GenAI projects? Connect with us!
