AI enthusiasts often critique the technology for its inability to yield substantial business results, frequently citing studies such as the one from MIT NANDA, which indicates a staggering 95% failure rate for enterprise AI solutions. Similarly, a report from IDC reveals that only 9% of organizations in Europe, the Middle East, and Africa have managed to deliver measurable business outcomes from their AI initiatives over the past two years.
However, many skeptics overlook the experimental nature of AI prototypes; not all projects are designed to transition beyond the testing phase. Nonetheless, a mere 5% success rate raises eyebrows.
4 reasons prototyping infrastructure ≠ production infrastructure
When discussing the challenges of moving AI prototypes into production, Phillip Merrick, co-founder, CPO, and chairman of pgEdge, points to data infrastructure as a primary culprit.
Merrick identifies four critical shortcomings of prototype environments that hinder their scalability to production:
- Deployment Flexibility: Prototyping environments often lack the flexibility required for large-scale production deployment. Although vendor-managed cloud platforms may facilitate quick starts, they typically fall short in security, compliance, and governance—especially in regulated sectors like healthcare and finance.
- Data Sovereignty: Transitioning to production can introduce data sovereignty challenges at both enterprise and regional levels. Merrick emphasizes that the data layer is crucial for enforcing sovereignty, but many prototype environments complicate this aspect.
- Reliability: High availability is essential for production environments. Merrick notes that vendor-managed cloud platforms often cannot guarantee seamless upgrades or hardware swaps without downtime.
“You’ve got to be able to choose where that AI prototype is ultimately going to be put into production.”
So why are developers prototyping where they can’t productionize?
Given that data infrastructure choices hinder the transition of AI prototypes to production, one might wonder why developers continue to initiate projects on unsuitable platforms. Merrick attributes this trend to the allure of ease-of-use offered by vendor-managed cloud platforms. While these environments simplify the prototyping process, they often lead to complications when the time comes to scale.
He explains that a disconnect exists between the prototyping phase and the production requirements, leaving developers unaware of what is necessary to transition their prototypes into operational applications.
More pgEdge articles in The New Stack
Historically, tooling decisions were made top-down, but over the last 15-20 years, developers have regained the autonomy to select their own tools. The current issue, according to Merrick, is that developers often choose tools solely for prototyping, neglecting to consider the production implications. This division of responsibilities means that developers build prototypes without a comprehensive understanding of the production requirements.
Merrick argues that this disconnect is where many AI projects falter, as teams struggle to migrate prototypes from inadequate vendor-managed platforms to robust enterprise-grade infrastructures that fulfill deployment flexibility, security, data sovereignty, and high availability needs.
He advocates for selecting the right data infrastructure from the outset to avoid this disconnect, highlighting Postgres as an ideal solution. He describes it as “the Swiss army knife of databases,” due to its extensibility and open-source nature, capable of addressing a wide range of data management challenges.
But picking the right data infrastructure only solves half the problem
Beyond infrastructure, Merrick emphasizes the importance of human resources, particularly database administrators (DBAs), who are increasingly overwhelmed. According to Stack Overflow’s 2025 Developer Survey, 84% of respondents now utilize AI tools, a significant increase from the previous year. However, this surge in productivity comes with a caveat: a shortage of DBAs to manage the growing number of databases.
Merrick insists that while agentic engineering has accelerated development, operational teams must also evolve to keep pace. He proposes that organizations should leverage AI DBA agents to enhance the capabilities of human DBAs, allowing them to manage more databases efficiently.
AI DBA agents can give humans “superpowers”
While Merrick does not advocate for fully autonomous DBA agents just yet, he acknowledges the pressing need for support due to the increasing complexity of AI applications. He introduces pgEdge’s Ellie, an AI agent designed to assist DBAs by performing various tasks, such as running diagnostics and providing actionable SQL code for human review.
“The world is not ready for fully autonomous databases administered by AI DBA agents.”
Ellie aims to alleviate the burden on operations teams, which are often stretched thin. With predictions indicating that 41% of current database professionals may leave the industry in the next decade, the need for efficient management tools becomes even more critical.
“An agent … can actually respond to those alerts far more quickly and productively than a human can.”
Better infrastructure AND people: It takes two to improve AI prototype success rates
In the realm of AI, the balance between quality and quantity remains a pertinent discussion. While agentic engineering enables developers to produce more prototypes, the transition to production is fraught with challenges stemming from both inadequate infrastructure and limited human resources. To enhance the success rates of AI prototypes, Merrick advocates for a dual approach: investing in robust data infrastructure and empowering operational teams with the necessary tools and support to keep pace with the rapid advancements in AI technology.
