Can Your AI Deployment Avoid the 95% Failure Rate MIT Warns About?

In early 2025, MIT's Project NANDA published a blunt reality check: only 5% of enterprise GenAI pilots are showing meaningful P&L impact. The other 95% stall—not because the tech doesn't work, but because the business never really learns how to use it.

‍

This "GenAI Divide" isn’t about GPUs or model quality. It’s about what happens after the demo. Most agents fail to move past flashy prototypes because they don't integrate into day-to-day operations, don’t adapt to changing workflows, and don’t retain context across systems.

‍

In short: they don't learn.

‍

‍

The Hidden Risk in GenAI Pilots: Lack of System Understanding

‍

What many AI teams miss is that even the smartest agent will fail inside a poorly understood system. When agents are deployed without a clear view of how they’ll affect upstream or downstream flows, they either:

• Fizzle out from lack of adoption
• Create chaos in adjacent functions
• Or get quietly retired without impact

‍

And because there’s no simulation or discovery process upfront, no one sees it coming.

‍

‍

‍

‍

What the 5% Do Differently

‍

What many AI teams miss is that even the smartest agent will fail inside a poorly understood system. When agents are deployed without a clear view of how they’ll affect upstream or downstream flows, they either:

1. They target operational (not just strategic) workflows with measurable outcomes
2. They integrate agents across functions—not just within them
3. They simulate system behavior before deploying into production
4. They embed memory and policy control to ensure agents adapt
5. They govern and monitor performance beyond the pilot phase

‍

And because there’s no simulation or discovery process upfront, no one sees it coming.

‍

‍

How Agentic Discovery Changes the Game

‍

This is where Agentic Discovery—done right—can help any team punch above its weight.

‍

Agentic Discovery isn’t a checklist. It’s a cycle: map the system, integrate real data, simulate changes, pilot, and iterate. When done properly, it allows teams to:

• Understand how work actually flows (not just how it's documented)
• Catch failure modes and bottlenecks before deploying agents
• Coordinate people, systems, and AI toward the same measurable outcome
• Learn from results and continuously improve the system

‍

‍

‍

‍

Where to Start

‍

Want to beat the 95% odds? Start with a domain that has clear throughput, cost, or risk constraints—like support ops, IT/OT triage, or back-office compliance.

‍

Then:

1. Model the actual system behavior, not an idealized version
2. Define the business outcome (cost per resolution, SLA compliance, MTTD/MTTR, etc.)
3. Simulate agentic options before wiring anything to production
4. Orchestrate across tools you already own
5. Govern and track impact using metrics that tie to outcomes, not vanity dashboards

‍

‍

The Bottom Line

‍

The GenAI divide isn’t technical. It’s organizational. To close it, you don’t need smarter agents. You need a smarter way to deploy them. That’s what Agentic Discovery delivers: a repeatable method for deploying AI with intelligence.

‍

And whether you use PathFwd or not, the pattern remains the same:

‍

Simulate first

Govern intentionally

Learn every time.

‍

That’s how you cross the divide.