Dr. Sachin Agrawal on Telling a Production-Ready AI System From a Convincing Demo

From research to production

Dr. Sachin Agrawal has spent more than 23 years moving between AI research and enterprise product work, most recently at PepsiCo and before that at Sony Research India and Samsung. Most of that time has gone into the distance between a working model and a working system.

“I have learned that building a good AI model is only the beginning. The real challenge starts when you put that intelligence into a real business environment, where the data is messy, users are unpredictable, systems are interconnected, and the cost of getting an answer wrong can be significant.”

The research and the product work taught him different things.

“My research background taught me to think deeply about algorithms and intelligence. Working on real products taught me something equally important: AI has to work as a system, not just as a model. You need the right data, context, retrieval, orchestration, evaluation, security and governance around it. That is what turns an interesting prototype into something people can actually depend on.”

What he means by discipline is the whole lifecycle, not the build.

“For me, the transition from research to production is ultimately about engineering discipline. You have to think about the full lifecycle: how the system is built, evaluated, deployed, monitored, improved and adopted. That is where AI starts becoming a durable enterprise capability rather than another proof of concept.”

The shift to context

Prompt engineering turned into a job title inside about two years. Agrawal thinks the useful work has already moved on.

“I think we are moving from prompt engineering to context engineering. A better prompt can improve an answer, but in enterprise environments the bigger question is: does the AI have the right information and context to make the decision in the first place?”

The pieces that supply that context usually get sold as separate products.

“That is where RAG, knowledge graphs, memory, tools and agent orchestration become important. I don’t see these as separate technologies. I see them as pieces of a larger intelligence architecture that helps an AI system understand the situation before it tries to answer or act.”

He sketches the shape he expects it to settle into.

“The architecture I see emerging has several layers: a context and knowledge layer, a reasoning layer, an agent and tool layer, and a trust and governance layer. The models will continue to evolve, but the surrounding architecture is what gives those models useful enterprise context.”

“For me, the next competitive advantage in AI will be less about writing better prompts and more about engineering better context.”

What happens when the AI is wrong

Reliability is often sold as the absence of error. Agrawal starts somewhere else.

“I don’t think we should promise that an AI system will never hallucinate. That is not how I approach production AI. The better question is: what happens when the AI is uncertain or wrong?”

For systems that act instead of answering, he wants depth.

“For a system that can take action, I want multiple layers of protection: grounded information, evidence, validation, permissions, guardrails, monitoring and human intervention where the risk warrants it. I also find the idea of AI evaluating AI very powerful, particularly through judge and verification agents.”

How much control applies depends on what the action can reach.

“The important point is that not every action needs the same level of autonomy. A low-risk recommendation can be handled very differently from an action that changes a financial, operational or customer-facing system.”

“My principle is simple: the more consequential the action, the more evidence and verification we should require before allowing the agent to act. Autonomy should be earned through reliability, not assumed because a model is capable.”

Beyond text

Agrawal’s earlier research covered computer vision, signal processing and audio intelligence. That background shapes how he reads the move to multimodal systems.

“The real world doesn’t communicate with us through text alone. We see things, hear things, read documents, work with structured data and increasingly receive information from sensors and real-time systems. AI will have to work in the same way.”

“My background in computer vision, signal processing, audio intelligence and intelligent systems has made me particularly interested in this convergence. The exciting part is not having five different models for five different modalities. It is enabling an intelligent system to bring those signals together and reason over them as one situation.”

The change he is describing is in what the system is doing.

“That takes AI from answering a question to understanding what is happening. An agent could combine what it sees, hears, reads and retrieves from enterprise systems before deciding what to do.”

“I see particularly strong potential in environments where decisions depend on multiple forms of evidence: industrial and field operations, healthcare, manufacturing, supply chain, customer experience and other real-world settings. The long-term opportunity is AI that can perceive, reason and act in the same environment as people do.”

What scale actually costs

A pilot runs on a few thousand interactions. Production runs on millions, and the arithmetic stops behaving the same way.

“When AI moves from a few experiments to thousands or millions of interactions, the economics change completely. I don’t believe every problem should automatically be sent to the biggest and most expensive model.”

“The question I ask is: what is the right level of intelligence for this task? Some tasks can use smaller models, some need deeper reasoning, and some may not need generative AI at all. Intelligent routing, semantic caching, inference optimization and good observability can make a significant difference.”

Cost control on its own does not get an organization there.

“But technology optimization alone is not enough. At scale, you also need a repeatable operating model: common architecture, evaluation standards, security, governance, FinOps, monitoring and a clear way to measure business outcomes.”

“Ultimately, I want AI systems to be intelligent, fast, economical and reliable at the same time. Scaling AI successfully is not about spending more on models; it is about getting more intelligence and business value from every unit of compute.”

The failure he sees most often has little to do with the model.

“One of the biggest lessons I have learned is that a technically impressive AI solution does not automatically create business value. Sometimes the technology works perfectly, but the problem was not important enough, the workflow did not change, or nobody defined how success would be measured.”

His sequence runs from the problem to the measurement, and the technology choice comes after both.

“I prefer to start with the business problem and baseline first: what are we trying to improve, what does it cost today, what changes when AI is introduced, and how will we measure the outcome?”

“Then we decide what technology is appropriate. Sometimes it is an agent, sometimes an LLM, sometimes traditional AI or optimization, and sometimes a combination. I don’t believe in using generative AI simply because it is the current technology trend.”

“The strongest AI initiatives connect business pain to an intelligence capability, then to a change in the workflow and finally to a measurable outcome. Technology creates the possibility; adoption and process redesign create the value.”

“That is also why I think an AI leader has to be comfortable moving between technology, product thinking and business economics.”

Governance and the CEO test

Responsible AI usually arrives as a checklist near the end of a project. He wants it in the architecture.

“I don’t see responsible AI as something that should come at the end of an AI project as a compliance checklist. It needs to be designed into the system from the beginning.”

Not every use case earns the same scrutiny.

“At the same time, I don’t believe every AI use case needs the same level of control. A simple summarization task and an autonomous business transaction obviously have very different risk profiles. Governance should therefore be proportional to risk.”

“That means thinking about identity and access, data protection, policy enforcement, evaluation, monitoring, auditability and human escalation as part of the architecture. It also means being transparent about where the system is confident and where it needs additional verification.”

“The goal is not to prevent AI from becoming autonomous. The goal is controlled autonomy: giving an AI system enough freedom to create value while putting the right boundaries around consequential decisions.”

“If we design trust into the system from the start, governance becomes an enabler of scale rather than a barrier to innovation.”

Asked what a chief executive should put to a vendor, he narrows it to three.

“If I were a CEO evaluating an Agentic AI platform today, I would ask three simple questions.”

“First: what measurable business outcome will this change, and what is the baseline today? A sophisticated demo without a measurable outcome is still only a demonstration.”

“Second: what happens when the AI is wrong? I would want to understand the evidence, evaluation, monitoring, escalation, permissions and recovery mechanisms. A high benchmark score alone does not answer that question.”

“Third: can this capability scale beyond one impressive use case into something the enterprise can actually operate? I would look for reusable architecture, integration with enterprise knowledge and systems, security, governance, observability and a credible cost model.”

“Those questions quickly move the conversation away from the demo and toward reality.”

“For me, the distinction is simple: a demo proves that AI can do something. A production system proves that an organization can trust it, operate it, scale it and create value from it.”