The 5% of Enterprise AI Projects That Succeed Aren't Winning by Accident

Please Note

This article is part of a developing analysis of the MIT NANDA report and what it suggests about successful AI implementation. I’m publishing this first version now because the central theme already feels useful: many of the AI projects that succeed appear to follow long-established project delivery, process improvement, and change management principles.

I’ll continue adding to it as I work through more of the report.

According to the report State of AI in Business 2025, around 95% of enterprise GenAI initiatives have yet to demonstrate meaningful business impact.

I’m interested in the 5% that appear to be generating significant business value.

Having spent years in software development, particularly as a Business Systems Analyst and Solutions Architect, I’ve seen countless initiatives succeed and struggle.

While AI may be new, many of the challenges surrounding its adoption are not. Successful implementations have always been built on a few fundamental principles.

I don’t believe AI has changed these fundamentals. It’s simply made them more important than ever.

Lessons From Project Delivery

The report suggests that organizations achieve the best results when AI enhances existing business processes rather than attempting to reinvent them.

That finding resonated with me immediately. In my experience, the most successful initiatives tend to build upon what’s already working rather than starting from a blank sheet of paper.

Transformational initiatives often carry more uncertainty because their requirements, outcomes, and benefits are harder to validate upfront

Change requests (CRs) are often successful because they address specific, well-understood business problems and opportunities

Automating proven manual processes is often successful because the business value, requirements, and desired outcomes are already well understood

Looking at the report through this lens, the success of many AI initiatives feels less like a new discovery and more like history repeating itself.

Key Findings

The report paints an interesting picture.

AI adoption is widespread, but measurable business value remains surprisingly elusive.

Looking through the findings, several patterns stood out. Many of them felt surprisingly recognizable from a project delivery perspective.

What Works

1. Narrow, High-Value Workflows

According to the report, successful AI implementations often focus on specific workflows rather than broad transformation initiatives.

Something I’ve noticed over the years is that successful initiatives rarely start with the biggest ideas.

They usually start with a clearly defined problem that people genuinely need solved.

The report’s most successful AI implementations appear to follow a common theme:

There’s another advantage to this approach: the risks are usually easier to manage.

If a single workflow improvement fails to deliver the expected results, the impact is often contained. If an organization attempts to transform an entire department and gets it wrong, the consequences can be much harder to unwind.

I’ve spent a lot of time evaluating AI use cases, and it’s clear that some workflows are naturally more forgiving than others.

Most organizations will have plenty of those to choose from.

Lead qualification

Of course, you want your sales team focusing on the most promising opportunities, but even if the AI occasionally ranks leads imperfectly, calls are still being made and opportunities are still being pursued.

The downside is limited. The potential upside can be substantial.

Customer service routing

If an AI router occasionally sends a ticket to the wrong operator, the issue can usually be corrected quickly by a human.

That’s very different from an AI making decisions that directly affect patient care, financial approvals, or legal outcomes.

2. Workflow Fit

Successful AI implementations also tend to fit naturally into existing processes rather than forcing people to adopt entirely new ways of working.

Buyers repeatedly emphasized:

If it doesn’t plug into Salesforce or our internal systems, no one’s going to use it.

This was one of the findings that resonated most strongly with my earlier observations around project delivery.

Organizations rarely need technology for its own sake. They need improvements that fit naturally into how the business already operates.

On a personal note, I’ve built a number of AI-enabled workflow tools to help with everyday tasks.

The ones I return to most aren’t necessarily the most sophisticated. They’re the ones that save the most time while fitting naturally into the way I already work.

3. The Real Problem Is Not Model Quality

One of the report’s illuminating findings is that organizations don’t see model quality as the primary barrier to successful AI adoption.

Instead, the report argues that the biggest obstacle is what it calls the Learning Gap.

Put simply, most AI systems don’t learn from experience.

It appears users are happy to use tools like ChatGPT for:

• Drafting content
• Brainstorming ideas
• Quick analysis
• General assistance

However, confidence drops sharply when the work becomes:

• Long-running
• Process-driven
• Business critical
• Dependent on historical context

The issue appears to be persistence and contextual awareness rather than intelligence.

An Old Problem In A New Era

Personally, I don’t see this as purely a model problem. I see it as a design problem.

The report’s findings resonated with me because they mirror a challenge that’s existed in process design for decades.

Imagine you’re tasked with designing a process for a human worker. It will need:

Regardless of how capable a worker might be, you wouldn’t expect them to succeed without:

• The right information
• Relevant historical context
• Clear feedback mechanisms
• A way to learn from mistakes

The same principle applies to AI systems.

Why This Matters Now

Viewed through that lens, many of the recent developments in AI start to make more sense.

Although these technologies are often discussed separately, they’re all attempting to provide:

• Context
• Continuity
• Feedback
• Adaptation

The technology may be new, but the underlying challenge feels surprisingly traditional.

Start With The Process

In my experience, the best way to understand a process is still to observe somebody performing it in a real-world environment.

That’s where you discover:

• Information requirements
• Decision points
• Exception handling
• Feedback mechanisms

Perhaps that’s why so many AI projects struggle. I keep hearing about solutions that appear to be little more than a ChatGPT wrapper with a custom system prompt, a polished user interface, and the latest model underneath.

If the report is right, the future advantage won’t come from having the smartest model.

To Be Continued

There’s more in the report that I want to come back to, especially around shadow AI, vendor selection, back-office ROI, and what successful buyers seem to do differently.

For now, my main takeaway is simple: the 5% aren’t just succeeding because they picked better AI tools. They appear to be treating AI implementation like real business change, with clear workflows, practical outcomes, user adoption, and feedback loops built in from the start.

Further Reading

• In the AI Regulatory Landscape, Agentic Graphs Can Make a Lot of Sense
• State of AI in Business 2025