As Employees Spin Up AI Agents, CIOs Need An Orchestration Layer To Turn Experiments Into Production
Gopal Parthasarathy, a technology and AI-enablement leader with recent roles at Zscaler, NetApp, and Oracle, described how enterprises move AI from employee experiments into production without flooding operations with unproven agents.

Across most enterprises, the AI story of the past two years has been a story of enablement: give everyone the tools, let a thousand agents bloom, and see what sticks. The experiments work. The challenge is in what happens next, when a prototype that saved a team a few hundred hours has to become something secure, governed, and reliable enough to run in production. The answer taking shape is an orchestration layer, a governed platform that coordinates specialized agents, deterministic workflows, and existing integrations, and decides which experiments graduate and how they're watched once they do.
Gopal Parthasarathy is a technology and AI-enablement leader whose recent work has run through some of enterprise computing's more demanding environments. Most recently he was the Director of Integrations and Automation at Zscaler, where he led a forward-deployed engineering model that shipped agentic solutions directly alongside business teams. Before that he spent more than a decade at NetApp building integration and platform-delivery systems, and he started out as a developer at Oracle, porting and shipping server software. He has, as he put it, seen it all, from the mainframe world to the cloud world to the AI world. That range is what shapes his read on where agentic AI tends to break.
"It's a workflow where one or two agents are part of a deterministic flow, but you use two or three agents in the mix. It doesn't have to be end to end, because 80% of my flow even today is deterministic. The right mix and match is where some of these technologies come and help you," Parthasarathy said. That estimate, that most of the work stays deterministic, is the premise under everything else he described. The job of an enterprise AI platform is less about building more agents and more about deciding where a few of them actually belong. It's a coordination problem before it's a modeling one.
Empower everyone, then pick the winners: "You empower everyone," Parthasarathy said, "so that they can rediscover themselves with AI and deliver value." In his previous organization, any employee could stand up a personal agent, something that read their calendar and email each morning and laid out the day's tasks, with no risk to anyone else if it crashed. The bigger ideas that surface from that sandbox get packaged and taken to the organization's AI leaders, who decide which ones are worth real investment. By the time a use case reaches IT, it arrives with an articulated problem statement and a business case already attached, which is how some leaders keep broad experimentation inside guardrails rather than shutting it down.
Business value over raw ROI: Hours saved is the metric that gets an agent noticed, but Parthasarathy argued it rarely survives contact with a business leader. A support team that cuts mean time to resolution from eight hours to four still has to answer what those freed hours actually buy the company. He drew the sharper line in finance: "If it's speeding up the close process from 10 days to seven days, that's a huge win, because the CFO's nightmare is on the 10th day you find a flaw and you don't have time." Plenty of IT leaders are still working out where the returns land, and his test cuts through it: "ROI is good, but business value, how do you articulate a business value for a problem that would win over others?" he said.
Proving value is the starting line: His framing is that the demo is the beginning, not the finish: "Proving that AI can deliver value is not the end game, that is the starting game." Getting an agent scalable, secure, governed, highly available, and redeployable when a bug appears is still the work of engineers who build and ship production systems, and none of that discipline disappears because the prototype came from a business team. Production also means wiring in the integrations and live data a demo skips. A demo clears the front door; it does not clear the building.
All of that assumes there's something solid underneath to run on. Parthasarathy's view is that the platform has to exist before the agents do, rather than get assembled in a panic once a handful are already live and starting to drift. That reframes governance, security, and architecture as one problem instead of three, because they're really symptoms of a single gap: not knowing how AI is actually operating inside the organization. The foundation comes first, and the agents get added to it on purpose.
The platform comes first: Before any agent ships, Parthasarathy wants the scaffolding in place: an agent or MCP registry, production evaluation, monitoring, and security controls, what he counts as eight or nine foundational pillars. The reason is that agents decay. "A few days it was performing well, and then slowly it starts tanking. Do you have the support system to continuously monitor and fix?" he said. Dropping a second agent in to cover for a failing one just adds one more thing that can go rogue, and those pillars only hold if the controls are built into the architecture rather than written down and hoped for.
Business and IT co-own it: One of his own teams built a resume-screening agent that ranked hundreds of applicants against a job description in minutes, and the early numbers looked close to 80% time saved. The engineers shipped it without the security and compliance teams in the room. "California introduced a new law," Parthasarathy said, "and we were using certain attributes like age, but we didn't put a guardrail." New state rules now make it unlawful to let an automated system discriminate on protected characteristics in hiring, and the tool was doing roughly what those rules prohibit. The fix wasn't technical. It was pulling the business, security, and compliance partners in before launch instead of after the complaints arrived. His takeaway was a single rule: "Partner with business as number one."
Hide the machinery from users: "Whether it is one agent or tens of agents doing the job should not be visible to the end user," Parthasarathy said. For the person on the other side, it should feel like an army of specialists, each trained for a narrow job, with routing and model selection handled underneath. "It's best to have an army of agents to solve a problem," he said, "and you pick the right agent for the right workflow." The platform carries the integrations and the coordination, and IT maintains it so business teams can plug in and build against clear guidance without touching the wiring.
None of this points toward a smaller technology team. Parthasarathy's wager is the opposite: the more employees who can build, the more the organization needs engineers who understand scalability, security, deployment, and the long tail of production support. The business partners bring the problem and the proof of value, and the technology team keeps the result standing and the guardrails honest. Releasing an agent, in his telling, is the moment the real relationship starts, not the moment it ends.
"We enable, and then you need to make sure it's all working, and then we are partners in crime," he said. For enterprises working out how to scale agentic AI without losing control of it, that partnership, running on a platform built to coordinate and govern the whole mix, is the part no model upgrade can supply.
If this caught your attention, that’s not accidental.
The best editorial systems don’t happen by accident. Outlever builds them.









