By AI agents deployed in enterprise IT operations have privileged access to endpoints, identity systems, and ticketing platforms. When an agent acts autonomously and without enforced boundaries, it can override access controls, trigger patch cycles on mission-critical devices, or execute actions that no human reviewed or can explain after the fact. The gap between a useful autonomous agent and a production incident is the presence or absence of guardrails, audit logging, and explicit access limits.
In this interview on TFiR, Romanus Prabhu Raymond, Director of Technology at ManageEngine, covers how ManageEngine is embedding AI agent governance, human-in-the-loop controls, and audited action chains across its unified endpoint and IT management suite.
Guest: Romanus Prabhu Raymond, Director of Technology at ManageEngine
Show: TFiR
Here is what every platform engineer and IT operations team needs to know.
Technical Deep Dive
Q: How do you govern AI agents in enterprise IT that have access to production systems?
Romanus Prabhu Raymond, Director of Technology at ManageEngine, explains that every autonomous agent action must be logged, every decision must be explainable, and access control lists must be defined so agents cannot exceed the scope they were given. Raymond draws a clear distinction between pointing accountability at a human employee and managing an autonomous agent that has no named responsible individual: the agent’s permissions, boundaries, and action history become the accountability layer. Monitoring must confirm the agent stays within its defined access boundaries at all times.
“Any decisions made by the AI should be explainable. All the actions are logged and all the ACLs are given so it doesn’t do anything overreaching.”
Romanus Prabhu Raymond, Director of Technology, ManageEngine
Q: What does AI readiness actually mean for an enterprise IT team?
Raymond frames readiness around five asset categories that generate operational data: people and identity, endpoints, applications, infrastructure, and the network connecting them. Until each category has visibility, managed configuration, and a reliable data feed, adding an AI layer produces unreliable output. The AI does not replace the management foundation; it operates as an additional layer that correlates data across those five categories more efficiently than a human navigating multiple consoles.
“Unless you have that foundation, that essential basics are not properly done, you will not be able to have an effective outcome with the AI.”
Romanus Prabhu Raymond, Director of Technology, ManageEngine
Q: How does ManageEngine add AI to an existing product architecture without rebuilding it?
ManageEngine treats AI as an acceleration layer placed above the existing data architecture rather than a replacement of it. The underlying structures for data discovery, inventory, unified visualization, and remediation remain unchanged. The AI layer improves correlation speed, for example surfacing who accessed which endpoint using which application without requiring a human analyst to pull data from separate consoles. Raymond is explicit that re-architecting a production product for existing customers is not an option; AI fits into the existing model instead.
“AI is being added as a layer to make sure that your existing product architecture is effectively utilized.”
Romanus Prabhu Raymond, Director of Technology, ManageEngine
Q: What IT management problems does NLP-driven endpoint interaction solve at the agent level?
Raymond describes a workflow where an IT administrator types a natural language query about a slow endpoint and the AI, which has access to inventory, telemetry, and patch data, identifies the cause and proposes a specific remediation such as clearing the recycle bin or prefetch folder. The administrator approves and the action executes. This converts a multi-step process across separate dashboards into a single conversational interface, and it enables the system to move from reactive response to proactive alerting before the end user reports a problem.
“Before the end user says that I have a problem, the system is going to tell you that these machines are performing very low and this could be the reason. What do you want me to do?”
Romanus Prabhu Raymond, Director of Technology, ManageEngine
Q: When should an enterprise use ManageEngine’s built-in Zia AI versus bringing its own model?
Raymond describes two distinct enterprise postures. Teams that have already validated AI value in their business and hold existing model subscriptions can connect those models through ManageEngine’s MCP connectors. Teams that have not yet adopted AI can use Zia at no additional credit cost to experience AI-assisted IT management in the context of endpoint, ITSM, and identity workflows. The decision is not a technical one; it follows whether the enterprise has already committed to a model investment and seen measurable return from it.
“Those who have not thought about AI in terms of the ways that they manage the enterprise IT, they want to try it. We want to give them the experience through our Zia so they don’t have to spend a lot of credits to experience that.”
Romanus Prabhu Raymond, Director of Technology, ManageEngine
Q: Why does data quality matter more than model quality for enterprise AI in IT operations?
Raymond identifies continuous and valid telemetry as the hardest practical constraint on AI effectiveness in IT management. Inventory data can be fetched periodically, but security use cases such as continuous threat exposure management require data that is current and accurate at the moment of decision. An AI operating on stale or incomplete telemetry produces unreliable threat assessments. The data pipeline, not the model, is the rate-limiting factor.
“Continuous monitoring is something that gives you better threat security perspective.”
Romanus Prabhu Raymond, Director of Technology, ManageEngine
Q: How does ManageEngine handle data sovereignty when agents and models run across global deployments?
ManageEngine operates twenty data centers distributed globally, with separate facilities for the UK, Europe, and the United States. Agents and models run within the data center serving the customer’s region, and data does not leave that jurisdiction for processing. Raymond confirms that local regulatory requirements for auditing and data destruction are applied at the regional data center level. No cross-border processing occurs for regions with strict data residency policies.
“Their data stays in the nation state. It’s not going to go out.”
Romanus Prabhu Raymond, Director of Technology, ManageEngine
Q: What is Autonomous Endpoint Management and how does human-in-the-loop work in practice?
Raymond positions Autonomous Endpoint Management (AEM) as the next stage after Unified Endpoint Management, where the system correlates security signals, makes patching and configuration decisions, and executes them, but keeps a human approval step for actions on mission-critical devices. He uses a healthcare example: patching a machine connected to a pacemaker during an active surgical procedure is catastrophic if done without human review. The agent knows the patch is required; the human decides whether this specific endpoint can be taken offline at this moment. The rule set alone is insufficient because business context and human values determine the actual safe action.
“It can make its own decisions, but depending on the nature of the business and the mission-critical endpoints involved, you have to decide whether this is feasible or not.”
Romanus Prabhu Raymond, Director of Technology, ManageEngine
Q: How are enterprise customers measuring the value of AI in IT management versus tracking token spend?
Raymond observes that the customer conversation has shifted from skepticism about sharing data with AI systems to a clearer cost-benefit framing: is the enterprise paying for credits or gaining measurable efficiency? The practical value case centers on lean teams managing remote workers who need consolidated dashboards and automated suggestions rather than manual correlation across multiple tools. Customers with existing AI subscriptions want that same contextual efficiency applied within their endpoint, ITSM, and IAM workflows.
“Are we paying unnecessary to the credits, or are we gaining out of AI which runs the business? People are clear now.”
Romanus Prabhu Raymond, Director of Technology, ManageEngine
Resources & Documentation
- ManageEngine, unified endpoint management, ITSM, and identity platform with embedded Zia AI and MCP connector support for third-party models
- ManageEngine Endpoint Central, unified endpoint management product covering Windows, Mac, Linux, mobile, and server patching from a single console
***
👇 Click to Read Full Raw Transcript
Swapnil Bhartiya: Every enterprise team is now being told to deploy agents. But an agent that manages tickets, endpoints and incidents needs access to the same system it is trusted to run. So the question that decides whether AI survives in production or not are the same ones. Can you govern what it does, audit what it did, and stop it when it’s wrong? At ManageEngine, that reality is reshaping the entire product architecture from purpose built models to audited guardrail agents actions to dig into how they are building this. We are here at ManageEngine’s UserConf 2026 in Orlando, Florida and I am joined by Romanus Prabhu Raymond, Director of Technology at ManageEngine. Romanus, it’s great to have you on the show.
Romanus Prabhu Raymond: Yeah, thanks for having me, Swapnil. Yeah, a pleasure.
Swapnil Bhartiya: It’s my pleasure. Let’s start with the platform itself. From a technology standpoint, what has changed in how ManageEngine architects its product so AI runs at a scale across the whole suite.
Romanus Prabhu Raymond: All right, that’s a great question because basically what I hear is are you re-rating the product for the AI era? No, that cannot be done. We have existing products, existing customers who are consuming it. We have a philosophy as our CEO was talking about, AI is a technology which has to be consumed contextually. If I am looking at it from the endpoints, then AI has to be leveraged in such a way that how AI can be used in that context by IT administrators who are managing the endpoints. For those who are managing identity, how can AI be leveraged so it’s more focused on that particular environment. So the underlying architecture of how the data is organized, how do you discover the data and add it and how do you have unified platform for that to be visualized, and from there how do you do remediations or actions on the endpoints or any other assets which you manage remains the same. Now AI layer is actually accelerating or making it more efficient in terms of correlation. In terms of, say for example, you want to understand who has done what in terms of identity, in terms of which endpoint they used, in terms of which application is being accessed or things like that. Instead of going into different consoles and getting the data and then correlating it, you are talking to AI which actually has the access to the structured data from different tools and then gives you better visibility. So I think AI is being added as a layer to make sure that your existing product architecture is effectively utilized. Unless you have that foundation, that essential basics are not properly done, you will not be able to have an effective outcome with the AI. So it’s not much changes, it’s being consumed as a separate layer.
Swapnil Bhartiya: Yeah, you mentioned Rajesh and I remember at a conference he also mentioned that, I mean, I’m kind of paraphrasing him, is that the AI outcome will not come from the smartest model you can buy. It comes from the readiness of your infrastructure and your team. And myself as a consumer business that depends on AI, I fully agree with it. But can you just elaborate a bit because you’re also touching the same point. What does that readiness actually mean? What does it look like for customers? Because you can help customers, you can bring the horse to the lake, but if the lake has no water, you cannot. So readiness. Talk about readiness. What do you mean?
Romanus Prabhu Raymond: Okay, so readiness from the management perspective, there is an enterprise right now, it has to manage its business. Now managing a digital business is depending on the assets that run the business. So basically what are your assets? Your people, identity is an asset and they use the endpoint for work. That’s an asset. And the endpoint has application, that’s an asset. The application is going through the infrastructure, what you have provided for connecting everything. And all these things combined together generate data. These are the basic five elements which generate data for the business and which is basically running the business. Now the basic responsibility is to ensure that you manage those assets, have visibility and then do provisioning, configurations, make changes, change management. It could be a cloud, it could be on prem, whatever that matters. It’s going to be within this five major elements. So you want to have visibility, manage, secure them. Now when you have this clarity, you might have multiple tools. Say for example for patching your endpoints, you might have a heterogeneous network: Windows, Mac, Linux, mobile devices, servers. There is a single process called patching. But when it goes down to the endpoint, depending on the type of endpoint, the process changes. Now this is already available in enterprises. You have maybe different tools. You have one tool for managing Mac machines, you have one tool for managing all your Windows, one for Linux, one for applications, third party application patching. Now all these things are already available, being managed. Now from ManageEngine we are enabling the customers with a tool set where from one UI you should be able to entirely manage everything. Now that is the basis which ensures that you are enabling them to manage this basic stuff. Once that is done, you should be able to do better operations, governance from the upper layer.
Swapnil Bhartiya: Excellent. Now let’s flip the script. We talked about the readiness of the customer. You talked about these five layers. Here at the conference, what announcements did you make? And if you just focus on the technical significance, not something that looks great on the demo or the keynote slide, but which is actually solving some of these customer problems. So what announcements stand out even for you?
Romanus Prabhu Raymond: You could see a lot of announcements in the AI segment apart from other areas. An AI segment which is not a completely new feature. Now we call our AI Zia. Right now, instead of people looking at the dashboard, you should be able to talk to it or type in a quick NLP query. For example: this computer, I’m getting requests from this computer saying it is exhibiting a lot of slowness. Give me some insights. Just type that. Now it has access to all the data, inventory, and latest patches and the telemetry data which is collected. Based on that it could say that there is no disk space. Do you want me to clear that? Now when you say clear that, whatever is predefined, things like cleaning the recycle bin or cleaning the prefetch, all these things can be done from there. So now you see that these are all basically configurations which can be individually done based on some discovery of inventory data or telemetry data. Now everything is done with human natural language processing and you should be able to execute that. So these simple things which had to be done as a configuration are now becoming very NLP-driven. I would say from reactive to moving to proactive. Before the end user says that I have a problem, the system is going to tell you that these machines are performing very low and this could be the reason. What do you want me to do? So that’s where it moves from reactive to proactive IT.
Swapnil Bhartiya: And as we all know that you have Zia, your agent runs on your own Zia. But do you support other LLMs as well? So customers can bring their own models. And if you do enable that, where do you draw the line between building in house of course and versus integrating third party models, and what pushes that decision about which model to use? First of all, you don’t charge for your AI. So there is no reason for them to need a subscription, but customer may want flexibility. But where do you draw the line that it’s better to use Zia, but you can bring your own AI?
Romanus Prabhu Raymond: Now that’s kind of an option or flexibility for the enterprises right now. I mean two or three years back people were really skeptical about giving data to the AI. Right now it’s not the case. People are trusting AI as long as AI brings better outcome for the business. They have started understanding the cost benefits. Are we paying unnecessary credits or are we gaining out of AI which runs the business? People are clear now. Now if you are already having an AI subscription and you are using it and you feel that you want to make use of that, we have MCP connectors where it can be consumed. At the same time, those who have not thought about AI in terms of the ways that they manage enterprise IT, they want to try it. We want to give them the experience through our Zia so they don’t have to spend a lot of credits to experience that. But still they should be able to have that. So that’s where the strategy is.
Swapnil Bhartiya: Now when you embed AI into enterprise IT management tools, what are the technical challenges that bite hardest when it comes to practical day to day life? Is it data quality, latency, or security?
Romanus Prabhu Raymond: Definitely the quality of data and how frequently you are getting your telemetry. Apart from inventory, which can be fetched on a time to time basis, when you talk about security, the data you have is more than has to be very current. So that’s why we talk about continuous threat exposure management. How exposed is your asset to a threat attacker outside? So you have to have a view of a threat attacker so that you should be able to protect your network. So continuous is something that is required, and at the same time it has to be valid and correct. Integrity has to be there. So everything matters. But I would say continuous monitoring is something that gives you better threat security perspective.
Swapnil Bhartiya: While here we are talking about external threats, I feel sometimes, and I talk to a lot of customers, their biggest threats come from inside. Because now agents can take actions and they are running in production environments. Sometimes there is no human in the loop. So how does ManageEngine handle the whole governance, guardrails, audit trail? And when it also opens the door for a human to step in when the agent is going rogue?
Romanus Prabhu Raymond: So this was being talked about today. You have so many autonomous agents doing things for you. Now if you have named a person, a people identity, you can point at him, he’s responsible, he’s the DRI for that. But when it comes to autonomous tasks being done, who’s going to be responsible? So that’s altogether another area of scope of management for the enterprises. Now they have to keep track of all these agents, and they are not overreaching, they are not going over and beyond your allowed levels. Access controls: it’s not doing more than what it is requested to do. Now all these things are guardrails. According to the instructions that you have given, it should not cross that limit and should be within that. And so it’s being monitored. Any decisions made by the AI should be explainable. So that’s how I think the guardrails are instructed in the system so that it is actually monitored and all the actions are logged and all the ACLs are given so it doesn’t do anything overreaching.
Swapnil Bhartiya: Another aspect is that many of your customers use a global audience. They run hybrid on-prem environments. Where do the agents and the model actually run, and what does that mean for latency? And these days data sovereignty is becoming a big topic as well. So if you can touch on these two topics because this is becoming a really important discussion.
Romanus Prabhu Raymond: So we have strategically placed our data centers across the globe. Twenty data centers. Especially wherever we have very strict policies, we have data centers. For example, UK we have a separate data center. For Europe, we have separate data centers. For the US, we have separate data centers. So their data stays in the nation state. It’s not going to go out.
Swapnil Bhartiya: I’m not asking about data. I’m talking more about agents as well as models. Where do they run?
Romanus Prabhu Raymond: Actually yes, it’s in the data center. It’s within the allowed premises, allowed areas, within the reach. So we abide with the local rules. If data has to be shown for auditing or data has to be destroyed, all these policies and guidance are considered and it’s going to be available locally. So there’s no matter of taking everything to the US and processing it and then giving it back.
Swapnil Bhartiya: We are here at the user conference and you work closely with customers. Everybody’s excited about AI these days, but sometimes the whole focus shifts to how many tokens I have consumed, how many tokens I burned, versus ROI, how much value I got out of it. So when you talk to customers, what are they talking about? Are they talking about the value agents are actually delivering today versus what is still promised but not seen in practice?
Romanus Prabhu Raymond: Okay, so that’s a great question. As I initially said, that’s the reason we have options to choose Zia or bring your own AI. So if you have already started seeing and tested the AI outcome in your business, you definitely want to use that in the products contextually for endpoint, for assets, for ITSM, and things like that. Right now, if you have not experienced that, we want to introduce that, making sure how AI in the context of a particular asset management, ITSM, or IAM can give better efficiency and better outcomes. In terms of now we see a lot of lean teams, we see a lot of remote workers. So all these things, you are trying to automate. Right now if you have to automate, you don’t have to juggle with multiple dashboards. Things are there, you should be able to consolidate and get suggestions and you can make a decision. Now that level, your AI which is within the product should be able to do that. Now again it depends on whether you want to experience that. Zoho gives you AI. If you already have an AI which you have started seeing value in your business and want that same context to be available here, you bring your AI. That’s where strategically we think we have positioned the products.
Swapnil Bhartiya: As AI is maturing, of course it is in production now. I don’t even know which company is not doing it. Some are actually overdoing it, they should tone it down. And I’m not going to ask you how do you think AI will look like next year, because we don’t know. Something will happen next week and we don’t know. But looking at the current trends, looking at the current customer pain points, looking at the patterns of how AI is reshaping ManageEngine, just give a glimpse of your roadmap. Not one year from now, but the recent roadmap. This is what is shaping ManageEngine and this is what customers can expect from us.
Romanus Prabhu Raymond: Okay, so I will reduce the scope to endpoint, because it’s too much to discuss about everything. So endpoint management was called Systems Management in the era before SCCM. It was called SMS Server, then it was called Systems SCCM. Then they introduced MDM, Mobile Device Management, then they said Enterprise Mobility Management, then they called it Unified Endpoint Management. Now the future is going to be AEM, Autonomous Endpoint Management, which means you have the same level of inventory, visibility, policies, things like that. But from a security perspective you have better correlation and make quick decisions, and the decisions are keeping human in the loop so that you also see that it’s not only based on the rules, it’s also based on human value. I’ll give you an example. In a healthcare network, patching is the same process for healthcare, for banking, for an organization which runs a retail business. But from the healthcare perspective, if you are going to patch a machine and reboot it, and that machine is connected to a pacemaker during an operation, it’s going to be catastrophic. Now that’s where human in the loop comes in. It can make its decision but depending on the nature of business, depending on the mission critical endpoints, you have to make a decision whether this can be done or not. So that’s where I think we are investing more in AI. That’s the future: autonomous but ensuring human in the loop so that business grows and human value is also maintained.
Swapnil Bhartiya: Romanus, thank you so much for joining me and sharing some of these great insights about what customers are going through and how ManageEngine is actually solving some of those problems and meeting them where they are in their own journey. Thanks for your time and I would love to have you back on the show.
Romanus Prabhu Raymond: Thank you. Thanks for having me. Those were wonderful questions from the customer’s perspective. Wonderful. Thank you so much.
Swapnil Bhartiya: Thank you.
