Many AI pilots stall in production because companies deploy models without clear ownership, updated processes, or built-in governance, which erodes trust and limits real business impact.
Sai Santosh Bandari, Senior Engineer at MetLife, explains that sustainable AI requires human-in-the-lead oversight, strong data guardrails, and disciplined problem framing before any code is written.
He outlines a practical framework that embeds accountability, continuous feedback, and production safeguards so AI systems deliver reliable, long-term ROI.
Sustainable AI ROI starts with accountability, not ambition. Companies racing to deploy new models often focus on capability, but long-term value comes from embedding human judgment, clear ownership, and governance into the system from day one. When accountability is engineered into the foundation, AI moves from experiment to durable business asset.
Sai Santosh Bandari is a Senior Software Engineer at MetLife who specializes in the design and delivery of end-to-end AI solutions. His deep expertise in production-grade NLP and Retrieval-Augmented Generation pipelines, alongside his contributions as an IEEE Member and guest speaker, positions him as a practitioner addressing these challenges in production environments. Many AI proofs of concept stall, Bandari says, because they fail to solve for foundational human factors, even when the underlying technology is sound.
“AI fails long-term when companies treat it like a tool upgrade instead of a behavioral and process transformation. Without human oversight, governance, and continuous improvement, it's just short-term hype,” says Bandari. The problem directly addresses the AI readiness gap many companies face when moving from exploration to production.
The trust deficit: Bandari confirms this diagnosis from his own experience. "People don't trust the output, and the processes aren't updated, which is ultimately an issue of ownership." Rebuilding that trust begins by augmenting human teams, allowing organizations to integrate AI in a way that provides immediate, reliable value.
Let the bot begin: Bandari's team puts this into practice with a production-based agentic system. "We are building a production-based agentic system where, instead of a human-in-the-loop for every production issue, we have a machine-in-the-loop," he explains. "The machine does the initial debugging, triages the issue, and alerts the right people. This helps ensure we don't breach our service-level agreements with the business."
From that practical foundation, a set of "golden rules" can create a framework for reliable and scalable AI. Before any code is written, a strategic foundation must be laid. "Problem framing is a must-have before building," Bandari says. "We need to define which user it's for, what decisions it can make, and why AI must be the one to do it."
With the strategy set, the focus turns to data. That process involves identifying sensitive information and implementing strict guardrails such as enforcing least-privilege access, blocking unknown documents, and encrypting data, all cornerstones of the AI data governance needed to scale securely with platforms. The same philosophy extends to the technical guardrails for the model itself.
Groundedness beats smartness: To achieve factual reliability, Bandari argues for a simple but powerful mandate: “Groundedness beats smartness; it's better to use verified sources. Bad retrieval causes bad answers, so you have to fix the search and chunking strategies before changing methods.” Perpetual oversight involves continuous monitoring for drift and hallucinations, but just as important is the human signal. "We must always take feedback in loops, using even a simple thumbs-up or thumbs-down signal to understand where the system needs to improve." These rules can help separate a fleeting experiment from a lasting solution. "AI delivers long-term value only when it becomes part of the process, not just a demo."
The HEART framework organizes these tactical rules into a complete leadership philosophy, guiding a hype-free AI culture built on the idea that accountability remains fundamentally human, a concept often described as human-in-the-lead. Bandari's framework demands "H" for Human safety and accountability in AI; "E" for Explainability and transparency; "A" for Alignment with business goals; "R" for continuous Review; and "T" for Trackability in production. It's an approach that aligns with the principles of a human-centric AI and the general industry push for a common AI code of practice.
The buck stops here: According to Bandari, the most effective leaders understand how to leverage human judgment in AI-heavy processes. "Leaders are ensuring that AI supports human judgment, not replacing it. While an AI can speed up analysis and recommendations, humans must stay responsible for the final decision, especially in high-impact areas like production incident management, customer impact, company compliance, and risk." That responsibility translates directly into a key tactical choice. For every AI agent, Bandari says leaders must ask: "Should it suggest an action, or should it take the action?" Supporting that choice requires engineering discipline, with built-in fail-safes and fallbacks like timeouts and graceful degradation.
Bandari points to an example of how his team at MetLife processes claims. "For processing insurance claims, an LLM might be 93% to 95% accurate. But in business-critical applications, we need 100%, because millions of policies are moving through our application, and people won't trust the product otherwise. We get 99.9% accuracy with OCR, which is why we choose the technology that meets the business need."
That choice is about prioritizing a "production-ready application" over short-term hype. "AI decision-making is not yet perfected, but it is transforming rapidly." Bandari concludes that effective leadership means grounding AI strategy in the proven capabilities of today. "In the future, we will get to 100% accuracy and use these advanced LLM techniques instead of OCR."
