According to the latest MIT NANDA report on the State of AI in Business, enterprises have poured between $30 and $40 billion into generative AI initiatives. The result? A staggering 95% of these projects never make it out of the pilot phase. In other words, the ROI is missing, the transformation is stalled, and someone’s probably regretting that enthusiastic LinkedIn post.

But why is the failure rate so high? It turns out that most of these pilots don’t fail because of bad models or regulatory roadblocks. They fail because they’re glorified science fair projects: one-off proofs of concept that never get embedded into real business workflows. The divide between early experimentation and actual operational change has become so wide that the report gives it a name: the GenAI Divide.

From pilot to purgatory

For most companies, the journey looks like this: explore a flashy new tool, give it a whirl in one department, then realize it doesn’t integrate well with existing systems or processes. Or, worse, it needs to be fed the same context over and over like a forgetful intern.

Generic tools like ChatGPT are adopted widely because they’re flexible and easy to try. But when companies attempt to scale with custom enterprise-grade solutions, the story changes. Only 5% of these tools ever reach production. The reason? Lack of memory, poor contextual awareness, and workflows that break the moment someone takes a PTO day.

And yet, employees are using AI all the time – just not the way leadership intended. The report highlights a robust "shadow AI economy" where workers are using personal ChatGPT accounts to get actual work done, while official company deployments languish in eternal pilot limbo. Shadow tools may lack enterprise security, but they do one thing very well: they work.

The real MVPs: systems that learn

Here’s where things get interesting. The few companies that are crossing the GenAI Divide are doing so not with brute force or bigger budgets, but with smarter approaches. They focus on embedding AI in narrow, high-ROI workflows and demand tools that learn from feedback, remember preferences, and improve over time.

In other words, the winners aren’t building AI to show off. They’re building AI to get stuff done. These organizations treat AI providers more like strategic partners than shiny software vendors. They empower line managers to vet tools and drive adoption, rather than waiting for central labs to issue a memo. And when it works? The results are measurable: improved retention, streamlined operations, and in many cases, a healthy trim to outsourced service costs.

A lesson from visual workflows

All of this should sound familiar to anyone working in visual media. Image workflows are notoriously complex. They involve multiple stakeholders, constant iteration, and the perpetual balancing act of quality, speed, and performance. Plug-and-play solutions rarely cut it. The tools that succeed are the ones that have their role in a bigger system.

So perhaps the future of successful enterprise AI lies not in building the biggest model or chasing the most viral demo. It lies in solving for specificity. In finding the quiet corners of our workflows where AI can do the most good.

That’s the story we’re interested in at Imgix: not just where AI is going, but where it’s genuinely helping. And especially in the world of visuals, where art meets infrastructure and every second counts.

Want to see what happens when AI does more than generate?

Join CNET Editor at Large Connie Guglielmo and Imgix CEO Chris Zacharias in a candid, forward-looking conversation about how AI is transforming design, marketing, and media production. From ethical best practices to unleashing your inner visual storyteller, the webinar digs into what it really means to create in an AI-powered world.

Watch the recording of How AI Is Shaping Storytelling and Visual Media.