Controlling AI agents is an emerging business: IBM Software India GM

IBM says enterprise AI success hinges on governing autonomous agents with strong security, identity management, cost controls and auditability across hybrid environments

As companies deploy autonomous artificial intelligence (AI) agents, managing them is becoming crucial. Sriram Raghavan, general manager of IBM Software India and Software Innovation Lab, tells Avik Das in an interview in Bengaluru why controlling the agents is important and why enterprise AI adoption is yet to kick off. Edited excerpts:  Why is AI difficult to scale up globally? 

Data continues to remain probably the number one reason but I think it has taken on a slightly different shape. [Second], the environment is causing everybody to ask not just how just to do AI, but do it in a sustainable, cost-effective and sovereign way. The third is the sense that spreading peanut butter over lots and lots of use cases, versus picking a big one that will move the needle for the company. Going through that end to end, not just on the technology front, but the organisational transformation of the process. I think that is still not something we consistently see everybody doing. 

Enterprise data management always tends to go through some big pivot every five to seven years. What is interesting about the pivot is that three things are changing about enterprise data management, albeit simultaneously. People are recognising that the consumers for enterprise data may no longer be directly human beings or developers but agents. So, the entity for which you are building an enterprise data management system is different. The second part is that while we have been talking about structured data, I think we are at a point where we are seeing AI unlock unstructured data. The last bit is the infrastructure, which has also changed. What is sitting inside the data centres now looks very different. Not just because there are graphics processing units, but also high-speed networks and different kinds of storage. And while the motivation to roll out the new infrastructure may have been AI, a lot of this infrastructure can also be used for accelerated data processing. 

The third stumbling block is how you think about an AI project. We were very intentional in going with specific verticals such as human resources, internal information technology support and procurement. It is also important to have executive ownership at the business process level to drive outcomes. 

How do enterprises figure out which use cases to take forward? 

They need to evaluate which business processes or lines of business have done their homework better, [where] the systems are better to grapple with a problem. What we end up doing is we share our experiences with customers and that is why we have a consulting arm. Because the heart of infusing AI into a business process is not to just attach a new widget but plumb it into the system. I think one of the advantages we have as a company, which has both consulting and technology, is that we are able to offer and bring that together to customers as well. 

But building and buying agents is not the story. How do enterprises control them? 

No client is going to only buy agents, nor build all of those.  So what customers are gravitating to today is how to manage this landscape of agents where they have to go through a build, run, manage cycle. Hence this notion of an agentic control plane is starting to emerge. 

Most of our competition offers control planes that only run in their houses. As a hybrid company, our technology always runs on-premises and even third-party Cloud, giving clients a control plane that actually runs across their estate. We are also proposing to our clients to run those agents on our platform so we can give them governance, security, trust and control. People live with fragmentation but you can’t have fragmented governance. So, this control plane is emerging as a very important area for us. We think that is very strong for us because of our hybrid estate, which allows us to offer an agentic control plane. 

What are the most critical aspects to control enterprise agents? 

Governance, security, identity management, cost and auditability. The development life cycle of agents is very different and so is their testing. So when I develop an agent, how do I do security enforcement? How do I make sure that the agent will work with only the privileges it needs to do its job, as opposed to blindly having super user privileges? How do I make sure the agent isn’t going to have a situation where it will blow up cost? Do I have traces over agentic decision-making? To answer a user’s question, if it keeps trying, I am left with a bill that is unjustified. So, these all become part of the evaluation and optimisation of the agent.