How Agentic AI is transforming businesses

In this episode of the EY India Insights podcast, part of our EY.AI series, we examine one of the most exciting branches of artificial intelligence – Agentic AI.

Key takeaways

Many organizations expect advanced solutions such as Agentic AI to address complex processes where traditional automation methods like RPA have fallen short.
Businesses are moving up from conducting many proofs of concept (POCs) to thinking about unlocking value and driving ROI.
Enterprises view Agentic AI as a 24/7 digital workforce that can be scaled up or down as per requirements.
Organizations are evaluating their processes to deploy agents with the aim to reduce time and cost and improving process quality.
Integration of a human-in-the-loop approach is crucial to provide guidance in some cases and continuously improve AI agents.
The current challenges with AI agents, such as performance consistency and quality, will likely get resolved over time.
To fully leverage the potential of Agentic AI, enterprises must concentrate on overarching processes and problem statements.
Emphasizing process diagnostics provides insights into ROI and delivery timelines, helping manage expectations from AI.

A lot of people are chasing use cases, but the larger question or a better question to focus on, especially given the promise of agents and Agentic AI, is to think about processes and problem statements.” Hari Balaji, Partner, Technology Consulting Generative AI | AI-led Transformation

Hari Balaji

Partner, Technology Consulting, EY India

For your convenience, a full text transcript of this podcast is available on the link below:

Pallavi

Welcome to a new episode of the EY.AI Series on the EY India Insights podcast. I am your host, Pallavi.
In this series, we explore how AI is transforming industries, driving innovation, and redefining business models. In today’s episode, we dive into one of the most exciting frontiers in artificial intelligence - Agentic AI.
Joining us to unpack this fascinating shift is Hari Balaji, Partner, Technology Consulting at EY India, and who leads EY India’s GenAI Center of Excellence. Hari brings deep expertise from global AI projects and insights from the frontlines of enterprise AI adoption.
Hari, it is a pleasure to have you with us today.

Hari Balaji

Hi, Pallavi. Thank you for having me.

Pallavi

Hari, to start off, a very basic question. Could you explain to all our listeners what Agentic AI is?

Hari Balaji

Yeah, sure. Let me actually start by talking about LLMs and then build from there to agents and Agentic AI. LLMs broadly are what made natural language interaction possible. They are large language models. It is what drives the likes of ChatGPT, which everyone is familiar with, is largely limited by its training.
It does not know what it does not know. It has now been augmented by things like web search, but essentially most LLMs are limited by what they have been trained on. Essentially, when we wanted LLMs to work with live data or real data or large volumes of data, then we needed to have a set-up by which we could introduce that data to the LLM and have it make some contextual inferences from that data. From that was born this idea of retrieval augmented generation which you can broadly think of as doing a Google search and then only presenting the first page of Google search results to your LLM as context and then saying, “Hey LLM, use this and come back to me with an answer that is relevant.” Based on just the data that has been supplied to you. Now for a long time, the capabilities of LLMs were largely limited to this, until we had a breakthrough in reasoning models.
This has been very hotly debated - can LLMs plan, can they reason, and so on and so forth. This idea of using a chain of thought, which is, giving an LLM the ability to think, reason, critique their answers, go back and reason again, come up with a better answer, and so on. Then the idea of just getting LLM to think before answering and have a set of intermediate steps.
Last year, especially post the launch of DeepSeek, which brought all of this into the mainstream, we became reasonably convinced that yes, we can get LLMs to reason and to critique, and to work in a certain fashion right where they are more thoughtful. They are able to plan. Now that gave fuel to this idea of agents.
Now, the idea of agents itself has been around for a while. But, essentially, you could now design these entities, which broadly have the ability to plan or reason and therefore to make some decisions and we give them some tools which allow them to act.
And finally, we also want to control them. So, we give them some form of a budget. You have this little box, which has the ability to plan and reason. It has some tools which it can use, which could be a web search or some calculators, and has some kind of budget which tells it that if it is spending too much time thinking or, reasoning and not getting to an answer, it better hurry up.
And that, sort of is the technical way of thinking about agents. This is sort of how let us say someone who is more of a technical background would define or think about agents. But practically this idea is, sort of, taken life in a very different direction and from a business user perspective, agents have become in their mind this idea of an infinite digital workforce.
Something that can be even, let us call it an intern, but it is available 24/7 because it is able to act and actually perform actions and drive processes. It is available 24/7, can be scaled up and scaled down and can drive any sort of process automation or process completion to a level that let us say previous generation technologies like any programmatic automation or RPA or intelligent automation or hyper automation or whatever you want to call it, was not able to do. All of a sudden, it has created interest because that has now become possible. That is sort of how I would define agents, both from a technical standpoint as well as a business style standpoint.
A technical definition of agents would be this little box which has a mix of autonomy, therefore it is able to think and plan and make some decisions. It has a set of tools which it can use to drive some actions, and it is given some kind of budget or constraints within which it needs to operate and it has this sort of, what we call path-finding flexibility, to be able to use the tools and work towards a goal or a problem, solution of a problem and get to an answer. Whereas for a business user, agents have been abstracted away to being this digital workforce which is available 24/7 and can be scaled up or down and therefore can drive various kinds of processes in a manner that the previous generation of automation technologies could not.
I hope that answers your question. I think quite longwinded, but I hope that answers your question.

Pallavi

Yes, thank you Hari. Now that we have set the context, we have seen so many new terminologies come and go. So, is Agentic AI really important? And also, do you think will it truly stand the test of time?

Hari Balaji

That is a very interesting question. While no one can say with any certainty, based on everything we know today, I do think that Agentic AI will stand the test of time. There are some good reasons for that. So, today, if you ask an average person, what do LLMs do, they may not have a good answer. One understands the average person probably does not understand what is a large language model, what is a transformer, etc., but if you ask them what does ChatGPT do, they will probably have a good answer. To draw an analogy, today if you ask people what is Agentic AI, they probably may not be able to give you a clear and concrete answer, except if they are very technical and working in the area of AI. But I think what an agent can do or what an agent is something that people have largely understood and accepted. Because it sort of aligns closely with this mental model of a human performing a task and agent is essentially a digital equivalent who is trying to do the same thing: manage a process, perform a task, solve a problem.
That mental model will stand the test of time. How does that agent perform and what is really the technology driving the agent? What is the definition of that Agentic AI? That could probably change over time, but I think the idea of agents are here to stay. Now, that leaves us with sort of one very large and open question, which is: what is that agent good for?
Today, probably agents are operating at the level of interns. There is a challenge with consistency of performance and quality of performance. But over time I think that will get resolved. The larger idea of having a digital workforce, which is available 24/7 and which can be scaled up and scaled down those three properties definitely captured a lot of people's imagination. And it is definitely an ideal that you will see people striving for. I think it is something that people want, business leaders want that. We will see that in action, and it is not an idea that is going to go away anytime soon.

Pallavi

Thank you, Hari. Pivoting towards the applicability of Agentic AI, where is Agentic AI being leveraged today and what are its core applications?

Hari Balaji

So practical implementation wise we are seeing a lot of our clients experiment with Agentic AI. But most importantly, I think with the whole idea of AI as such we have kind of moved away from doing lots of POCs to sitting down and thinking deeply about where can AI unlock value and drive ROI.
We are seeing a lot of our clients slow down now that the idea of agents is here, sit down and deeply evaluate the processes and workflows that they have across various parts of their firm, and think about where agents can play a role in either shortening completion times, improving process quality, or reducing a process cost, or simply in a field like audit, increase the amount of coverage for a process. So you do not have to sample as much for it, for example.
The other sort of tenet or idea is there have been situations where people have tried RPA or traditional automation and have been limited in solving problems and there I think the idea of using Agentic AI is becoming interesting because if you take any process that has some level of complexity and where, traditional automation has faced some challenges, like, you know, I could name, pretty large list like invoice processing or claims processing, approving a loan, doing quality control or process, doing root cause analysis, automated resolution of tickets, in an industrial setting, looking at various telemetry information and then coming up with, a conclusion or a decision, in an HR setting, doing a talent assessment or doing KYC or customer support, or in a software development lifecycle, writing test cases or doing code review, there is a large list of situations where a traditional approach to automation or AI has sort of fallen short or has had its limitations. People are revisiting those ideas with Agentic AI simply because with agents there is at least promise of better state through processing rates, with just some SOP and process docs and some kind of natural language guidance without extensive engineering or explicit instructions. The idea of just having a human in the loop, to guide the agent when there is an edge case and then using that to reframe or improve the agent has become quite prevalent now.
This approach or this way of using and adopting Agentic AI, I think, will sort of stay and we will only see the use cases and the applications grow from here.

Pallavi

Thank you Hari. Now that we have discovered endless possibilities of Agentic AI, for interested listeners, how can they get started with Agentic AI? And what are some common Agentic use cases that you have come across so far?

Hari Balaji

I sort of covered some of these before, but let me take a step back and say there is an over obsession with this idea of use cases today, a lot of people are chasing use cases, but the larger question or a better question to focus on, especially given the promise of agents and Agentic AI is to think about processes and problem statements.
The first step is to run a deep process diagnostic. What is the current way of doing things? What are the inefficiencies of that? Has traditional automation been tried? Has it worked? Why did it not work? And typically, running this process diagnostic knowing that Agentic AI exists, typically the solution is some kind of a mix of traditional automation, programmatic tooling, probably even adding some machine learning in the mix and kind of putting all of this under the control of an agentic orchestrator, to offer some reasoning and decision making capabilities along with a human in the loop as required. Everything is going to, you know, kind of be decomposed into this mix or template. And you will find that processes are getting constantly, made more efficient and reinvented, in light of Agentic AI.
As EY, of course, we are obsessed with bringing all of our firm to clients. This is especially true of AI, and we can not run this kind of process diagnostic without, for example, bringing together people, all sorts of people into the mix. The users, people who understand the process, domain experts, people like me who are familiar with AI or Agentic AI, and then also people who understand data very deeply.
So this is kind of a long winded approach for a CXO, who is just looking to focus on the lowest hanging fruit and quickly taking a few use cases to production. But we are in the third or fourth year of AI now, I cannot underestimate the importance of just focusing on processes and running these diagnostics.
Because while this approach may take its own time, following this will allow us to get a very good estimate of ROI also time to delivery and manage expectations much better and sort of deliver AI transformations successfully. I know this is not like a straightforward answer to your question, but, I hope that this concept or this idea of this tenet is helpful to your listeners, right.
To sort of understand how to think about getting started with Agentic AI, it is always process first and then breaking that down into how the template of having an AI as an orchestrator, with oversight and reasoning, and below that, having a set of tools, or even traditional AI or ML along with the human in the loop.
So just that template, figuring out how that fits into your problem statement of process and then designing accordingly and doing this at scale, I think, is what will drive success.

Pallavi

Thank you. Hari, thanks for sparing time and walking us through the power and potential of Agentic AI. This conversation has certainly given us a lot to think about and to all our listeners as well. And also what we can look forward to in the future of intelligent systems.

Hari Balaji

Thank you for this opportunity. I hope this was helpful and we are always available if you have questions about AI and Agentic AI.

Pallavi

Thank you. And to all our listeners, thanks for tuning in. Do not forget to subscribe to the EY India Insights Podcast and follow our EY.ai series for more episodes like this. Until next time, this is Pallavi signing off.

Presenters

Hari Balaji

Partner, Technology Consulting, EY India

If you would like to listen to our podcasts on the go:

Amazon music

Spotify

SoundCloud

Podcast

Episode 25

Duration

16m 21s

In this series

EY refers to the global organization, and may refer to one or more, of the member firms of Ernst & Young Global Limited, each of which is a separate legal entity. Ernst & Young Global Limited, a UK company limited by guarantee, does not provide services to clients.

Insights

Highlights

Services

Spotlight

Industries

Case studies

Careers

Spotlight

About us

Top news

How Agentic AI is transforming businesses

Key takeaways

Pallavi

Hari Balaji

Pallavi

Hari Balaji

Pallavi

Hari Balaji

Pallavi

Hari Balaji

Pallavi

Hari Balaji

Pallavi

Hari Balaji

Pallavi

Presenters

Hari Balaji

In this series

Presenters