The market for AI in customer service is moving from $12.06 billion in 2024 toward $47.82 billion by 2030, but the part customers feel is not the model, it is the moment the system actually resolves the issue instead of escalating it. Rohith Narasimhamurthy, a Sr. Software Developer at a global cloud services provider, has spent the last few years building the hard middle where that resolution becomes possible, with the judgment and rigor you would expect from an IEEE senior member. To understand how teams are turning brittle support playbooks into secure, executable AI workflows at scale, we spoke with Narasimhamurthy about what is working, what breaks in production, and what comes next.
“An SOP is only useful if it can run,” Narasimhamurthy says. “If the best answer still ends in a handoff, the customer experiences delay, not help.” He is describing a shift many teams talk about but few implement cleanly: taking human troubleshooting steps and turning them into agent workflows that can execute safely, repeatedly, and fast.
That shift is showing up in the economics. The generative AI market is projected to climb from $63 billion in 2025 to $220 billion by 2030, and enterprises are not paying for demos, they are paying for outcomes that remove friction from daily work. In support, the practical outcome is fewer cases created in the first place.
Narasimhamurthy’s GenAI Contribution Platform was built for that exact conversion: standardizing how SOPs become AI agents and automated workflows so subject matter experts can contribute knowledge without waiting on bespoke engineering per use case. It was designed and launched in phases starting in mid 2024, with milestones through mid 2025, and it is targeted at automating 350+ SOPs with a projection of 500,000 hours of case work time saved annually by 2025. This is the part that sounds unglamorous until you have lived it. One missed step in an SOP becomes a loop in production.
Support automation gets romanticized as a reasoning problem. In practice, the first question is simpler: what is the agent allowed to touch. If you cannot answer that precisely, you do not have automation, you have risk. This is where teams get hurt.
Privileged access management is expanding from $3.82 billion in 2025 to about $9.35 billion by 2030, and that growth reflects a broader reality: modern systems have too many actions that can change state, too many identities, and too little tolerance for vague access rules. You can build an agent that knows what to do; you still have to prove it can do it safely.
In the Contribution Platform, Narasimhamurthy helped architect security first execution using permission boundaries designed to protect customer environments, including the Forward Access Session token approach for scoped access. The platform also went through comprehensive application security reviews and validation processes, because the only acceptable “fix” in a support setting is one that is both correct and authorized. “We treated access like part of the workflow, not a side constraint,” he notes. “If the boundary is unclear, the agent should not act.”
Even good automation fails when the underlying knowledge is messy. Support organizations accumulate years of tribal knowledge, edge cases, and half documented steps, and throwing that into a model does not make it reliable. It makes it louder. Nobody wants that.
The knowledge management software market is expected to expand by $28.33 billion from 2024 to 2029, a reminder that enterprises are still paying for the basics: structure, retrieval, and reuse. The “agent era” does not eliminate that need; it raises the bar, because a workflow that executes must have cleaner inputs than a wiki page ever required.
This is where Narasimhamurthy’s platform design matters. The system was built as a contribution mechanism with standardized SOP structure, vector storage, tool registry, evaluation services, and an execution framework, so knowledge arrives in a shape the system can use. The result is not theoretical. The automation of 117,000 support engineering cases is expected to save 368,000 hours of case work time, a 5.3% reduction that shifts human effort away from repetitive troubleshooting and toward the cases where judgment still matters. Midway through his career, Narasimhamurthy also took on the role of a CODiE Awards judge, which is the kind of evaluation posture that fits this work: you learn to ask what is proven, what is repeatable, and what fails under load.
Enterprises are learning that “agent quality” is not one number. A workflow can be helpful and still unsafe. It can be safe and still useless. The only sustainable path is to build evaluation into delivery, so changes do not become regressions that silently inflate escalations.
MLOps is growing from $2.33 billion in 2025 toward $19.55 billion by 2032, and that trajectory reflects something practical: teams are spending money on the boring work of release discipline, testing, monitoring, and continuous validation. In agentic systems, that discipline is not optional. It is the product.
Narasimhamurthy built the platform with modular separation between SOP authoring, storage, orchestration, tools, evaluation, and execution, and the delivery plan itself was phased so each domain could be validated before expanding. He also had to define integration points across the support assistant surfaces and workflow automation services, which is the part most teams underestimate. “Shipping an agent is not a launch day,” he says. “It is a promise that the workflow stays correct next week.”
The future is not a world where humans disappear from support. It is a world where humans stop doing the steps that should never have required a human in the first place. What rises in value is the ability to draw clean boundaries: what can be automated, what must be escalated, and what evidence is required before either decision is made.
Agentic AI is projected to grow from $6.96 billion in 2025 to $42.56 billion by 2030, and that growth will reward systems that can act without becoming reckless. Narasimhamurthy’s Contribution Platform was built explicitly to programmatize tribal knowledge, with a stated target of driving a $1.2 billion impact on support’s bottom line over three years by scaling automation without scaling headcount at the same rate.
Looking ahead, he expects the differentiator to be evidence. The teams that win will be the ones who can show, step by step, why an agent took an action and why it was allowed to. That outlook also explains why he has continued taking on external technical evaluation roles, including an invitation to participate as a keynote speaker and peer reviewer for International Conference on Research Trends in Artificial Intelligence and Data Science International Conference on Research Trends in Artificial Intelligence and Data Science (ICIRAIDS 2025). In a market that will be crowded with agent demos, credibility will come from systems that can defend their behavior in the moments that matter.