Last week, media outlets exploded with a staggering insight from MIT: “95% of genAI projects fail” and “almost zero companies are seeing return on their AI investments.” Released by MIT’s Project NANDA, the report The GenAI Divide: State of AI in Business 2025 explores the divide between high adoption and low disruption, as well as reasons why disruption and transformative results are low.

The message is sobering, but it doesn’t mean organizations should panic about their own generative AI implementations or hit the brakes on their plans. In this blog, we’ll unpack the study and uncover primary takeaways to ensure you see the right ROI in your generative AI projects.

Wait—only 5 percent of AI projects succeed?

The executive summary of the report certainly catches your attention, indicating that 5 percent of integrated pilots extract “millions in value” and the rest see “zero return.” A closer look under the hood shows that the takeaway is more nuanced. Researchers define success as an AI pilot moving to deployment with measurable KPIs, with ROI impact measured six months after deployment. 

The researchers don’t detail what’s going on with the remaining 95 percent of pilots. The report indicates that individual worker productivity has increased with these remaining pilots, and longer-horizon impact may have yet to surface especially for complex enterprise systems with longer-term implementations. Additionally, pilots that have stalled may have uncovered other productivity wins without P&L movement (or without “millions in value” of P&L movement).

Key takeaway: There is more nuance than the framing of 5% success vs 95% failure of initiatives. The statistic should still be taken seriously—but perhaps with a grain of salt.

Why do genAI pilots stall?

According to the report, 60 percent of organizations looked at task-specific, embedded genAI tools, 20 percent piloted a tool, and only 5 percent say the tool made it to production with measurable P&L impact or marked and sustained productivity (which is where the “5 percent of projects succeed” statistic seems to come from). When understanding why tools aren’t piloted or why pilots weren’t implemented, organizations have four main complaints:

• The tools don’t learn. When asked to rate common barriers on a scale of 1-10 in frequency, “model output quality concerns” averaged nearly an 8, indicating that the tools didn’t have learning or memory capabilities needed for enterprise work.

• Specialized tools sometimes don’t work as well as generic tools. Users who like consumer-grade LLMs for their flexibility and immediate utility are simultaneously skeptical of custom or vendor-pitched tools, calling them overengineered and misaligned with actual workflows. 

• Both generic and specialized tools don’t retain context for specialized work. The same users who praise consumer-grade LLMs to support repetitive, simple work also criticize the tools for repeating mistakes and requiring too much manual context. Both internal tools and consumer LLMs fall short in memory needed for complex, specialized workflows.
• The tools don’t integrate into real workflows. Users want AI systems that plug into current tools and processes. Quotes from interviews show the disconnect between what vendors offer and what enterprises expect: vendors don’t understand how processes and data flows work, genAI solutions need to work with current toolsets and internal systems, and the AI needs to adapt as internal processes change.

Key takeaway: Only 5 percent of interviewed organizations say their tool made it to production with measurable P&L impact or sustained productivity. For those who didn’t, it’s because the tools didn’t meet the expectations or hype to truly deliver business value.

3 insights for your own implementations 

If genAI pilots fail because the tools have a learning gap and don’t integrate well as this study suggests, then it’s critical for AI systems to adapt, remember, and evolve in order to provide true value. Whether you have a pilot you’re currently running or you’re planning one, there are a few things the report recommends you keep in mind:

• Start with a narrow, high-value use case. Narrow use cases mean lower setup burdens and faster time-to-values. It’s easy to expand niche usages to broader scopes once they’re proven, but it’s harder to scale back once you start. 

• Don’t get swept away with visible use cases. A majority of AI budgets go to sales and marketing usages, both because they’re highly visible and the outcomes can be more directly measured. In reality, more subtle efficiencies, like those in back-office functions, can be more transformative (and possibly lead to the highest ROI opportunities).
• Choose the right partner. Organizations prefer trusted partners rather than newer vendors. Externally co-built solutions are twice as likely to reach deployment than internal builds. 

Key takeaway: Start small, and start with the right partner. Ask vendors and partners how their solutions will adapt, remember, and evolve to your workflows.

What organizations should do now

Although the initial statistic that’s caused this study to take off in the media may sound alarmist, it uncovers crucial considerations for organizations as they work towards their own AI implementations. AI systems targeted at specific processes that are capable of learning and evolving along with the processes they’re attached to will deliver the most value.

And, of course, you don’t need to save millions in P&L to consider your implementation a success. Improved customer retention and sales conversion, more delighted and satisfied customers and workers, or more intelligent and faster decision-making—just to name a few—can all be truly valuable outcomes that make investing in AI systems worth it.

Zirous will continue to watch for releases on best practices, guidelines, and recommendations from credible sources to provide insights and takeaways for our clients. If you’re curious about how these learnings may impact your own AI implementations, or want to learn more about how to identify the right business cases for future implementations, reach out to Zirous.