We are a young company, and we will not pretend otherwise. We do not yet have decades of client engagements to draw from or a catalog of case studies to reference. But we have been paying attention. Closely.
Throughout 2025, we watched what practitioners shared on LinkedIn and X. We read the reports, attended the conferences, and listened to the hard-won wisdom of data engineers, ML practitioners, and enterprise leaders who were generous enough to share what they learned in the trenches.
What follows is not a summary of our own client work. It is a synthesis of what the industry taught us, lessons we carry into our practice as we begin serving organizations navigating these same challenges.
We share these observations not as authorities, but as students of a field that rewards humility. If 2025 made anything clear, it is that the organizations succeeding with AI are those willing to learn from failure, their own and others.
The High Failure Rate Problem
The most sobering finding of 2025 came from research on AI in business. Multiple studies, including joint research from BCG and MIT Sloan Management Review, found that the vast majority of generative AI pilots, estimates range from 80% to 95%, fail to deliver measurable impact on the P&L.[1] This finding ricocheted across LinkedIn feeds and conference stages, forcing uncomfortable conversations in boardrooms worldwide.
The instinct was to blame the technology. But practitioners who dug deeper found something more instructive. The problem is not the quality of the AI models, it is the learning gap for both tools and organizations.
This resonated with what we observed in countless posts from data leaders. The technology works. The organizational readiness often does not.
The Data Quality Reckoning
If there was a single theme that echoed through practitioner communities in 2025, it was this: data quality is not a technical detail. It is the foundation upon which everything else depends.
Research has consistently shown that poor data quality is the root cause of failure in the majority of AI projects.[2] We saw this acknowledged repeatedly by engineers and executives alike, often with a tone of hard-won recognition. Organizations had invested heavily in models and platforms only to discover that their underlying data assets were fragmented, inconsistent, and riddled with definitional ambiguities.
A survey from Informatica identified the top obstacles to AI success: data quality and readiness at 43%, lack of technical maturity at 43%, and shortage of skills at 35%.[3] The pattern was clear. The limiting factor was rarely the sophistication of the algorithm. It was the reliability of the inputs.
The principle of garbage in, garbage out has been understood for decades. Yet somewhere in the excitement around generative AI, organizations forgot to apply it.
What struck us was how many practitioners described this as a rediscovery rather than a new insight.
The Pilot Paralysis Pattern
One of the most discussed failure modes in 2025 was what practitioners called pilot paralysis. Organizations launch proof-of-concepts in safe sandboxes but often fail to design a clear path to production. The technology works in isolation, but integration challenges, secure authentication, compliance workflows, real-user training, remain unaddressed until executives request the go-live date.
We observed this pattern described again and again in conference talks and LinkedIn posts. Teams would build impressive demos, generate excitement, and then stall. The gap between a working prototype and a production system proved far wider than anticipated.
A related phenomenon was what some called model fetishism, engineering teams spending quarters optimizing performance metrics while integration tasks sit in the backlog. The result was technically impressive systems that never delivered business value because they were never truly deployed.
Where the ROI Actually Lives
One of the most counterintuitive findings of 2025 challenged assumptions about where AI investments should focus. Over half of enterprise AI budgets went to sales and marketing pilots, yet research showed the biggest measurable ROI is actually in back-office automation, where AI reduces reliance on business process outsourcing, slashes agency costs, and streamlines repetitive workflows.[1]
This disconnect between investment allocation and value realization was a recurring theme in practitioner discussions. Organizations were drawn to visible, customer-facing applications while the most reliable returns came from less glamorous operational improvements.
The data also revealed that specialized vendor-led projects succeed approximately 67% of the time, while internal builds succeed only about 33% of the time.[1] This finding prompted reflection on when to build versus buy, and on the value of domain-specific expertise over general-purpose tools.
The Wisdom of Failure
Perhaps the most valuable lesson we absorbed came not from success stories but from practitioners willing to share their failures. Tim O'Reilly captured this ethos well in his writing about learning technology: "If you've never broken it, you don't really know it."[4]
He described the fake confidence that comes from tutorials and toy examples, and the real learning that only emerges when things go wrong. The failure muscle you build with databases is the same one you need with AI coding tools. You cannot tiptoe in. You have to push until something breaks. Then you figure out how to approach a new technology as a professional.
This perspective reframed how we think about competence. The practitioners who seemed most credible in 2025 were not those who claimed seamless success. They were those who could articulate specifically how and why their projects had failed, and what they learned as a result.
The Enduring Importance of Foundations
Despite all the excitement around AI-assisted development, experienced practitioners consistently emphasized that foundational skills remain essential. AI can help you code, but it is not always correct. You can have AI write a complex project, but how do you know what it does and what data source it uses? Is it privacy compliant? Does it meet your requirements? It is still up to you, the data engineer, to decide.
This was a recurring theme in posts from senior engineers. The tools are changing rapidly, but the judgment required to use them well depends on deep understanding of the underlying systems. Those who skipped fundamentals in favor of AI shortcuts often found themselves unable to debug problems or validate outputs.
The Human-AI Collaboration Imperative
Finally, we observed a significant shift in how practitioners framed the relationship between human workers and AI systems. The early narrative of automation and displacement gave way to something more nuanced.
Research consistently shows that human feedback loops improve model performance significantly.[5] The most effective implementations were not those that removed humans from the loop. They were those that structured collaboration between human judgment and machine capability.
Organizations reporting significant financial returns were far more likely to have redesigned end-to-end workflows before selecting modeling techniques.[1] The focus shifted from replacing people to enhancing what they could accomplish.
What We Carry Forward
We enter 2026 without the battle scars of failed deployments or the confidence that comes from years of successful engagements. But we do carry something valuable: a clear-eyed understanding of what the industry learned the hard way.
We know that data quality is not optional. We know that the path from pilot to production requires deliberate planning from day one. We know that the most reliable value often comes from unglamorous operational improvements rather than flashy customer-facing applications. And we know that the practitioners worth listening to are those who can speak honestly about what went wrong.
These lessons shape how we approach our work. Not as experts who have seen it all, but as practitioners committed to learning from those who came before us.
The industry taught generously in 2025. Our job now is to apply those lessons well.
This is the third in our January series on data and AI strategy for 2026. Subscribe to receive the full series as it publishes throughout the month.
Sources
-
BCG & MIT Sloan Management Review, "Where's the Value in AI?" (2024). Research finding that most organizations struggle to capture value from AI initiatives, with failure rates between 80-95% widely reported. bcg.com
-
Gartner, "Data Quality." Research establishing data quality as a persistent barrier to AI and analytics success, with up to 59% of organizations not measuring data quality at all. gartner.com
-
Informatica CDO Insights Survey (2024-2025). Survey identifying data quality and governance as top obstacles to AI success (~40-50%). informatica.com
-
O'Reilly Radar, "Technology Trends for 2024" (2024). On building real competence through failure and iteration. oreilly.com
-
Human-in-the-loop machine learning research consistently shows that structured human feedback improves model performance. This pattern has been documented across domains including document classification and entity extraction.