Adi Kuruganti, Chief AI and Development Officer, Automation Anywhere, on why enterprises need specialised agents, reusable automation components and structural governance to scale agentic AI reliably.

As enterprises move beyond foundational AI models towards agentic systems, the focus is shifting from what individual models can accomplish to how multiple agents and tools can work together to execute business processes reliably. While the idea of deploying a single, powerful model to handle end-to-end workflows may appear attractive, enterprises must contend with operational complexity, probabilistic errors, security risks and the difficulty of troubleshooting failures.
A more modular approach, built around specialised agents, deterministic automation and reusable components, can help organisations distribute tasks according to the capabilities required at each stage. However, scaling these systems also demands stronger governance, particularly when agents handle sensitive or mission-critical processes. Enterprises must establish checkpoints, define appropriate levels of autonomy and ensure that human intervention is built into decisions where the consequences of failure are significant.
In this conversation with CIO&Leader, Adi Kuruganti, Chief AI and Development Officer, Automation Anywhere, discusses why specialised agents offer a more manageable architecture, how enterprises can prevent cascading errors in multi-agent workflows, and why structural governance, auditability and human oversight will be essential to scaling agentic AI.
CIO&Leader: The industry initially focused on “super models” handling everything. What are the practical friction points enterprises face when they try to stretch a single foundational model, and why is an ensemble of purpose-built, specialised agents a superior design pattern?
Adi Kuruganti: One of the biggest challenges with relying on a single model or agent to handle an entire process is the operational overhead involved in managing and troubleshooting it when something goes wrong. When an agent is responsible for too many tasks, it becomes difficult to isolate where a failure occurred or understand the different behavioural factors that contributed to it.
At Automation Anywhere, our approach is to break a process into the individual steps required to complete it and use the appropriate tool at each stage. That might be a specialised agent, deterministic automation or another tool altogether. This approach offers advantages in cost, accuracy and efficacy, but it also makes troubleshooting considerably easier. You can identify exactly which component failed rather than having to investigate the entire process, reducing the operational work involved in diagnostics and recovery.
It is also important to recognise that more powerful AI models do not necessarily become less probabilistic or more deterministic. They may be capable of doing more, but that does not automatically make them more controllable or reliable, particularly when handling sensitive processes or tasks that demand high accuracy. That is why enterprises need to be thoughtful about how they break down processes and determine which tools are appropriate for each step.
CIO&Leader: If foundational models continue to commoditise over time, where will competitive differentiation belong?
Adi Kuruganti: Competitive differentiation will come from understanding how work actually gets done within an enterprise and knowing how to execute it appropriately.
Competitive differentiation will come from understanding how work actually gets done within an enterprise.
Many models operate based on the theory or information they have been trained on. But there is a significant difference between that knowledge and understanding how business processes function in practice. Real-world execution involves exceptions, variations in how work flows and knowing which tools are best suited to particular tasks.
At Automation Anywhere, we are working from real-world execution patterns that we continue to recognise and update. That includes understanding exceptions, how work moves through a business, and which tools are appropriate at different stages of a process.
We are also bringing in a customer’s own business context through our Context Intelligence Graph. Over time, this creates a flywheel of learning from real-world execution and how work gets done within an organisation. That intelligence can inform not only the models but also the different tools available to execute the work.
This understanding will become even more important as enterprises introduce more AI systems. Those systems will change more frequently, and things will inevitably break. Having a deep understanding of what work needs to happen and how it needs to navigate different systems will become increasingly valuable. That is where we see a meaningful opportunity for differentiation.
CIO&Leader: Building a bespoke agent for every siloed process is neither cost-effective nor maintainable. How can enterprises design agents as reusable, modular services that function across finance, HR and supply chain?
Adi Kuruganti: This is how many organisations have approached robotic process automation (RPA) and traditional automation for years: by building individual, reusable components designed to perform specific tasks that can then be applied across different processes and use cases.
Think of these components as Lego blocks. For example, an automation that extracts a particular piece of customer information could be useful across multiple processes and industries. Rather than building that capability again for every use case, organisations can reuse the same component wherever it makes sense.
The important consideration is identifying the specific skill or task being performed and determining where else it might be applied. Enterprises should also be thoughtful about whether a particular task requires AI or an agent, or whether a more deterministic form of automation would be appropriate.
Reusability becomes particularly important when it comes to ongoing maintenance and operational overhead. If you are trying to build one giant super-agent to handle every use case, the work involved in maintaining it, managing its security and troubleshooting it can quickly become more significant than building the agents themselves.
Breaking processes into reusable components makes them easier to manage and maintain, while allowing organisations to apply the appropriate capabilities across different business processes. Ultimately, this also supports the innovation enterprises are trying to drive rather than creating additional operational complexity.
CIO&Leader: When Agent A hands off an output that serves as Agent B’s ground truth, a small hallucination early on can compound into an operational failure. What safeguards or deterministic checkpoints prevent these cascading errors in automated pipelines?
Adi Kuruganti: With probabilistic systems, errors compound at every step. Even if an agent performs an individual task with relatively high accuracy, reliability decreases as more steps and decisions are added.
Simply giving an agent access to deterministic tools is not enough, because the decision to use those tools is still probabilistic. Those checks need to be built directly into the workflow as independent steps.
Simply giving an agent access to deterministic tools isn’t enough, because the decision to use those tools is still probabilistic.
It is also important to match the level of autonomy to the task. A goal-based agent works when you know the desired outcome, while an open-ended tool like Enterprise Claw is useful for investigating problems when you do not know the cause or the steps required. In those situations, read-only access can help accelerate troubleshooting without allowing the agent to make changes.
A good example is a healthcare customer using our technology for medical billing and coding, a process that demands 100% accuracy. The AI component alone achieved roughly 75% accuracy. The customer embedded it within a thoughtfully designed workflow, using approximately 47 deterministic tools to validate outputs against established sources of truth. Anything that did not align was surfaced to a human for review.
This is what I call structural governance. Rather than relying solely on prompts and guardrails, the structure of the workflow governs what happens next. AI can reason and generate outputs, while deterministic checks validate them and determine whether they can proceed or require human intervention.
CIO&Leader: As agents take on mission-critical execution, complete autonomy can pose significant risk. How should architects determine the threshold between straight-through agent execution and mandatory human sign-off without creating bottlenecks?
Adi Kuruganti: I think the starting point is understanding the risk of getting something wrong. What is the worst-case scenario, and is the outcome reversible? If an error is relatively easy to reverse and unlikely to cause significant damage, there is more room to allow autonomy. But when a process involves finances, customer interactions or potential harm that is not easily reversible, human oversight becomes much more important. Ultimately, organisations need to determine their own risk tolerance.
It is also important to consider the value AI can deliver. As one healthcare provider put it, “The greatest ROI comes from preventing failure.” They used AI to check, roughly a week in advance, whether all the necessary preparation for scheduled operating room procedures had been completed. If something was missing, the system triggered follow-up actions to address it before the procedure. Even if something was flagged incorrectly, the consequences were relatively minor compared with the value of preventing an unused operating room slot.
We also need to rethink what constitutes a bottleneck. Human review is not automatically a bottleneck if it is focused on genuinely high-risk decisions or recommendations that deliver significant value. The goal is to match the level of autonomy and human oversight to the risk of the task, rather than requiring human intervention at every step.
CIO&Leader: Regulatory frameworks such as the EU AI Act require transparency and traceable accountability. When a multi-agent system takes a business action, how do you provide end-to-end auditability showing which agent made what decision, based on which data, and at what time?
Adi Kuruganti: One of the challenges with auditability in multi-agent environments is having to piece together information from different agents and tools to understand what happened during a process. Every handoff creates another point where information needs to be collected and connected, particularly when different systems operate independently.
Our approach is to bring that visibility together in one place. Rather than requiring organisations to cobble together information from multiple agents and tools, the idea is to provide a single pane of glass through which they can see what is happening across the process.
This makes it easier to understand and examine the execution of a workflow without having to investigate multiple disconnected systems independently. Having that information in one place is an important part of making processes more transparent and manageable.
CIO&Leader: The recent EY India report highlights that nearly 58% of Global Capability Centres (GCCs) have already invested in agentic AI. What distinct challenges are GCCs encountering as they shift from experimental pilots to core global process execution?
Adi Kuruganti: The 58% figure reflects real momentum in agentic AI adoption among GCCs, with EY also finding that another 29% plan to scale within a year. But investing in agentic AI and running a global process on it are very different things.
Investing in agentic AI and running a global process on it are very different things.
There are three challenges that stand out: reliability, accountability and sprawl. Sprawl refers to bespoke agents being built team by team, which in turn multiplies maintenance requirements and the number of points where things can break.
At Automation Anywhere, our approach is to break processes into individual steps and determine where AI agents, deterministic automation or human oversight are most appropriate. By orchestrating native agents, third-party agents and people within a single governed process, supported by open standards such as MCP and A2A, enterprises can maintain visibility, simplify troubleshooting and scale agentic AI while keeping outcomes predictable and auditable.
