The Agentic Enterprise : Scaling AI, Controlling Costs and Redesigning Work

FULL TRANSCRIPT

This transcript is based on the episode’s English auto-captions and has been formatted for readability. Please allow for occasional transcription errors in names, acronyms and specialised terms.

[00:00:00]

2026 onwards, what we have seen, um that's when uh the year of economic reality has come into the picture. So, AI builds uh build shocks are now coming. Your organization chart wasn't actually designed to make decisions. It was designed to move information from one layer to the other. So, and these layers exist because information was expensive to transport from lower levels to upper layers, and uh AI makes that essentially free. Data is what is known in an organization, a context is what the agent needs to know, and the memory is what the system learns over time. Imagine um this you know, the this as a Formula 1 uh you know, car, where you have the model as the engine, and the harness is the car. So, a Ferrari engine inside a poorly designed car loses races all the time. Teams are measuring, you know, prompts, tokens, uh conversations and tasks. Uh whereas the executives are measuring revenue, retention, margin, you know, risks, growth, right? So, there apparently seems to be no relationship between these two. You may be doing 80% of your work, which is the iceberg model which MIT came up with, um which may be under the belt, um which basically would be, you know, uh you know, working with uh with different interfacing with different people, coming to know about uh different opportunities here and there.

[00:01:32]

AI is not going to do that. Now, those developers are becoming architects. Your analysts are becoming supervisors. Your SMEs are uh are now looking at themselves as knowledge curators. Uh your product management guys, they were they are now becoming capability managers. So, it's not just uh a change in the title, it's the the way that they're functioning. Hello and welcome to the Enterprise Tech Talk podcast. I'm your host, Saumitra Kalikar. Now, over the past few years, organizations have invested and experimented with various forms of AI. But now there is a decisive shift towards agentic systems that can learn, that can reason, act, and coordinate work increasingly autonomously. But the real challenge now that is emerging for enterprise leaders is how to industrialize this new form of AI.

[00:02:42]

That is how to effectively engineer and architect it, how to govern it, um how to derive value from it, and most importantly, how to prepare their workforce for this next phase. To unpack these important questions, I'm joined today by Amit Mukherjee. Amit has deep experience in agentic systems. He's also founder and CEO of ClearBlade AI. Amit, welcome to the podcast. It's good to have you here. >> Thank you, Soumitra. Pleasure to be here. >> Amit, let's get started with your professional journey. Um tell us a little bit about how you got into uh the world of agentic systems, and what led to the foundation of ClearBlade AI. >> For me, it's a very old journey. Um I entered through creating decision systems, you know, the the fuzzy matching, the data cleansing part of it, uh back in the day, the Bayesian risk scoring uh that we used to do, um you know, for and the weighted scoring uh systems, the ensemble classifiers, and holistic evaluations, and so on and so forth. Uh there were also rules engine which we had developed and um the collaborative filtering and retail in retail. So, for me it was uh it was a quite an old journey and uh you know, right now I think uh when people say that AI started is started by GPT systems, I kind of smile a little. We we have uh uh we we were actually injecting uh intelligence into apps before it was fashionable.

[00:04:18]

You know, I I see AI in three waves. The first wave was a symbolic uh and the rules-based AI which was basically the Eliza systems of uh 1960s and later the expert systems. But uh the second waves is when I actually uh came into the picture with the statistical uh machine learning, you know, scoring, classification, recommendation engine, the the fraud uh uh detection engines and predictive analytics. Um this is this was a wave I was directly part of and uh now the third wave of AI which is a transformer-based uh and agentic AI, that's uh you know, this is this is where it's it's the most exciting part uh of uh AI where, you know, we are we can we can see the reason against uh we can see these agents reason against the goals, uh retrieve context, call tools and so on so forth. So, I've always worked in highly regulated spaces like insurance and financial institutions and also health care. Um but there's a lot of scope for challenges. That's why I kind of um you know, uh founded the Clearways AI for enterprises to move beyond the demos and make sure that uh you know, they they're going through a production implementations smoothly.

[00:05:34]

>> Yeah, as you you rightly said, many of us who have been in the industry for last 20 years or so, they have seen at least those last two decades. Um recommendation wave and Netflix was the poster child for for for that case. And now the transformation phase, um it will be engineer phase. Um wait, but when you talk to uh your enterprise customers, right? Uh about the the AI maturity uh they have uh um what on across the AI maturity curve, in general, where do you see most of the enterprises are today? >> Uh see, AI is going through the psycholo- psychological journey right now, uh which which internet went through in the late '90s. So, at that time, you were making uh you know, websites to show your presence, uh your yeah, I created demos, etc. at the same. So, all, you know, demos are magical to look at the first time. You know, this is like uh showing a Victorian person the microwave to cook.

[00:06:37]

But uh 2026 onwards, what we have seen, um that's when uh the year of economic reality has coming to the picture. So, AI bills uh bills shocks are now coming. And um it was put into place uh to fuel adoption initially with the token uh you know, kind of tokens given out. However, now people are consuming less tokens uh than badly designed agents. So, uh when you know, token audits are taking over. Maturity-wise, the real value is hiding uh in in deeply boring things. Like, for example, your your claims intake, your invoice reconciliation, your uh compliance workflows, and so on and so forth. Now, AI's first superpower may be simply by removing, you know, uh organizational block which is existing there. But uh you know, here's here's something which nobody is uh saying. Your organization chart wasn't actually designed to make decisions.

[00:07:37]

It was designed to move information from one layer to the other. So, and these layers exist because information was expensive to transport from lower levels to upper layers, and AI makes that essentially free. So, which makes the org chart itself a bottleneck. Now, I've seen organization in various stages of maturity. It's now left in the board room. It has now left the board room. And you know, it's it's penetrating into workspaces. You know, where it is if it is implemented badly, it's like a like like as a feature, it's showing marginal marginal and no benefits. But in case it's embedded into the DNA of the company, then these boring processes, you know, and in all these boring processes, it actually shows real revenue. >> That's right. Now, I do want to unpack with you the implications of operating model. There are quite potentially significant operating model implications in the future as adoption of AGTK matures more and more.

[00:08:47]

But also want to understand what where you see enterprises focusing in terms of use cases and benefits at this stage. Um Is it more of a back back office kind of use cases that they are safely playing around or now they are starting to also start rolling out the agentic use cases more for customer experience and new revenue generation opportunities? >> So, I think we can differentiate that between exploitation and exploration parts. We have seen that the the start is actually from the boring spaces, like claims processing, like doing you know, boring work like what you may call just getting in all the data, cleaning the data, etc. That's all exploitation work. This was quite quite redundant work. Now, um you know, we would like to actually >> Productivity and productivity and >> Correct. Correct. Correct. That's where we should start from and that takes away in a job function, that takes away like 80% of some people's job. And it's only 20% where they are really thinking about doing something more productive and more, uh you know, kind of uh innovative. That's That's the places where we would like to explore more and that's where, uh you know, agentic framework, not agentic, but at least the uh the the the GNAI work initially and then it has been taken over in the exploitation phase by the agentic frame framework. >> Yeah. Yeah.

[00:10:25]

And how do you see uh ex- most of the organizations um doing experiments versus actually starting to do the productionization of use cases because you and I know, uh may there have been technology bills in the past and uh we know that ex- experimentation, piloting something does not necessarily lead to deriving value to production. Right? And then there have been some reports even from Gartner if you are released that most of the organizations um have impressive uh experimental experimentation level use cases and pilots, but they still keep struggling to um across the board with productionize those, right? So, where do you see organizations on that phase and importantly, get to want to get your perspective as to why organizations struggle there? Um what is it that they miss in terms of productionizing use cases?

[00:11:24]

>> Right. Right. Right. So, in productionization, uh what I what I'm seeing towards is, But, know, like when when we implement a pilot, it just proves that AI is possible in an organization. But, if you go for uh actual implementation, the enterprise deployment, that shows that the uh deployment is actually AI is reliable, it's governable, and it's economically valuable. Now, everything looks great in pilot. Uh the data is clean, the workflow is controlled, and honestly, nobody from compliance has joined uh the meeting yet. But, production is really where it shows up. Now, the systems has to deal with messy enterprise data in production. There are conflicting systems, other permissions, latency, audit requirements, there are there edge cases which come up in production. Uh and you know, there's there's actually uh business accountability uh involved in all of this. And with agentic AI, this becomes even more important because the system is no longer just generating content, it's taking actions on your behalf inside workflows. So, you know, pilots can be impressive, but uh enterprise deployment actually prove whether it's a survivable survivable thing in production or not. So, that's how I see it, uh at least the difference between the pilots and the enterprise-scale deployments today.

[00:12:48]

>> No, that's true. And as you said, there's a lot of explanation going on. Um and for last 2 years, I think most of the organizations spent just to get comfortable with these new new forms of AI. But, now uh they're being asked they're being audited for for the the benefits, and we'll be paying that as well. Uh but, let's let's move uh towards uh a little bit of techni- technical aspects of agentic systems. I want to unpack at least at a high level what does it mean from the engineering perspective of agentic systems, and how to architect those, right? Uh when we talk about agentic systems, properly productionizing agentic systems I've talked about. Um There are a few things coming to mind, right? Uh there are a few capabilities you need to have uh in place, right? Uh Uh there is of course focus on the the the the traditional capabilities like having good data backbone, data governance, strong security posture, etc. Um but I want to unpack what it means here specifically for the agentic context. But then there are few new capabilities that are also very much specifically relevant to agentic, uh particularly multi-agent systems when we talk about the context management. Uh The memory management, state management aspect, right? So, would you mind unpacking that or painting a picture at a high level as to what kind of capabilities organizations should think uh to to have a more effective, successful multi-agent to run.

[00:14:27]

>> Right, right. So, um you know, if you if you're talking about um you know, uh implementations, there are certain foundations which you which you have to take into uh consideration, you know, um so that's that's something which goes uh without saying. Um let's let's actually unpack that to uh you know, you have to have in an agent you have to have a strong memory strategy. Uh you know, if if you go with that, uh you have to have a knowledge memory, an episodic memory, a procedural memory, which is basically your uh you know, uh which goes which is the first aspect of it. What agents are basically agents are nothing but if you see these are orchestrators which uh do work for you know, they have they are specialized specialists in systems and they do specialist uh specialized work for an organization.

[00:15:28]

And the way that they are arranged is based on what kind of architecture you are following today. Right, you could follow you know, an agent which can be a hierarchical agent. It can be have an orchestration agent. You can have level agents and so on and so forth. But you know, they they are they are masters of of basically managing the context, the memory and creating an outcome for an agent work orchestration workflow. That's what they they actually do. Now choosing there are various dominant architecture patterns for agents. You can have a react pattern which is a reasoning and act pattern which is basically you observe something, you think and the tool does it. Like you can search the web. You can analyze the results, summarize it. It's fast, cheap, simple. The other types of agents are like a planner agent which is slightly more sophisticated where you can execute. You have a goal.

[00:16:30]

You can plan a task. You can execute the task and you can you know, get a result. That's it just reduces hallucinations. That's one of the you know, kind of mature agents. Then you have a generator to critique agent where basically you generate a content and you critique it and it the revision is also critiqued by the agent itself. You have a coordinator agent which is a in a hub and spoke model which most of your viewers might be familiar with. Which basically you know, you have research agents, you have coding agents and finance agents which coordinates these you know, which are coordinated by these coordinator agents. And then you have a hierarchical model which it's like a company agent where you can think of as a agent which gives work to a director and a manager and a worker. So, there are various forms of agents and how they are architecturally arranged and what they do.

[00:17:29]

But their goal is simple to basically get some redundant work done and also get in a in a context. >> And what differentiates agents from other engineering and architecture patterns in the past like microservices etc. is they rely heavily on on context, right? How important are are those context of it for agents and how organization should should thinking of surfacing that context for agents? >> So, you know, I think there's there are three patterns. One is like the data in an organization context and what the memory is. So, the data is what is known in an organization. A context is what the agent needs to know and the memory is what the system learns over time. So, if you think like a human consultant, the data is everything in a filing cabinet which exists. All the data manuals, all you know, your your company tax records etc.

[00:18:38]

The context is the document which are spread across your table for today's meeting, right? And the memory is the experience you have accumulated over the years of prior engagements. Now, the consultant needs all three of them. Agents are no different if you want to get achieve a particular goal. So, just like humans, if you have the right context, we can search the data for in the filing cabinet and get our experiences you know, add our experiences to it and gather more experience doing it. So, this this was an interesting use case I saw as part of storing data in memory relationships that can be stored in graph systems also where you have relationships stored. For example, you know, you have um if you're trying to get tax records, those tax records are related are in my tax files, so that relationship is there. You can go to this particular database to get it. So, these kinds of things, um you know, kind of interplay with each other for and and the organization needs to manage that very well.

[00:19:44]

>> Well, the question I have on this is just to elaborate a little bit further. Um look, you must have you and I uh we know typically know what information resides in which locations you need across the enterprise, right? And typically, what is the source of truth for what type of information, right? We know the quality the policies are the source of truth for certain information and CRMs are the source of truth for another information, etc. Now, agents, when you actually start rolling out agentic systems which are completely autonomous, they may may not understand or may not know that what is the source of truth for what kind of information, right? And and if you're surfacing and exposing that kind of agentic system to customer interaction, you want to make sure that they are delivering the right advice and they're using the right sources of truth to to deliver that advice.

[00:20:47]

So, how that is particular problem should be addressed in your view? >> Right. There There are various, you know, precision at K etc. These are some of the metrics which we use in the industry to make sure that, you know, whatever information is getting pulled, is is is is is correct. Now, we also have various guardrails to make sure that the the execution of the agent is is governed and it's it's safe. Like scope guardrails are there. Like every agent should have a clearly defined operating boundary. Your permission guardrails are there. You know, they should be operating on the privilege the least privilege principle. You know, there are action guardrails as well, which is you know, in case you're thinking it's a low-risk thing. Like for example, you know, suggesting an email, you know, action which is basically high-risk, which is sending out an email. Now, we have to have these policy and human escalation guardrails and so on and so forth to make sure that you know, the data which is coming inside the agents and is basically a correct data which is coming in.

[00:22:07]

There is there is I'm sorry, it's the data which comes in to these systems is actually checked through the human to through the precision at K, recall at K and other you know, methodologies that we have developed over time. >> One thing I quickly wanted to unpack with you was is a specific question I have in mind because that is surfacing in in in the late conversation in the conversations lately in the AI industry. When do you think of mentioning the context or for agents, right? Or having the reference guardrails for agents? The de facto format that has emerged for in the agent AI industry is the the master format, right? And you see that there is a clear you use Clar whether you use any vendor technology. There is the standards standard context files which are which which are there like agent.md or memory.md and few other things like off-late there was a paper where there was a paper post I don't remember exactly from one one of the employees from Entropic who argued that we probably had one two four and we should explore alternate options like HTML which is much more human friendly.

[00:23:42]

Right? Uh because this is a hot topic right now. I want you to give your your thoughts on this as to uh where which pattern is is one pattern really better than another or or and which pattern potentially would would be better tomorrow more in the future. >> Yeah, I I wouldn't say that um HTML versus Markdown uh we we you know, it it's not a competition there but we have to use where things are apt. Like for example, I'll give you some something which we recently implemented. Uh broker wants to implement recommendation for a $50 million manufacturing client, right? So the orchestration engine needs to coordinate between your client intelligence agent, your benchmarking agent, carrier agents and and so on so and so forth. Here, you know, if you're using a graph uh way of a graph database to actually uh send context, so there your enterprise memory uh comes into the picture. For example, the ABC manufacturing which we were talking about, um it owns property policy placed with Travelers Insurance, right? So that relationship when you're transferring uh context to an agent, that is most important. Your YAML file storage also we did.

[00:25:06]

There the orchestration workflow, you know, we we we we were saying that okay, there's there's some things which you can do sequentially. Some things like for example, step one is client intelligence, step two is benchmarks, etc. Run Run certain elements in benchmarks like carrier scans, etc. You can do in parallel branches. So, that is specified very well in YAML files. The XML HTML the agent to agent task handoff, it is actually pretty good at that. So, because it has got a task tag, it has got an agent tag under it, and you know, you have a client tag, so you do not the agent does not lose context there. It It knows which industry you're talking about, which deliverable deliverable tag you're talking about. So, it doesn't get confused as you're moving from you know, as you're traversing through the XML.

[00:26:03]

I We also use JSON for runtime data exchanges like for example, if you have the benchmark which comes in, these benchmarks were returning premium percentiles and loss ratio, etc. And the renewal risks, etc. So, that was coming in JSON. The markdown was specially important when you have to have human readable knowledge like an executive summary, etc. What I've seen as in case you're debugging, you know, it it actually makes sense to put stuff in markdown files. Otherwise, or it's summarizing in the final output of the agent. Otherwise, you know, the most effective way of agent to agent task handoff is HTML or XML. For YAML orchestration workflows, for graph, if you have enterprise memory you can store there. JSON is for any runtime data exchange and you know your markdown definitely is for human readable inputs.

[00:27:02]

>> Well well now this is a question I have is this is again one of the topics which is emerging in the AI industry as I see is about harness engineering. And and the the the overall emerging acknowledgement now is that AI models are great but they are slowly becoming commoditized in a sense that yes most of the vendors at least open AI at topic and others they have similar kind of models and which you can choose from but the real differentiation value is now how you create the harness around around those those models right? And harness engineering is becoming more important than just choosing a particular AI model. What is your view on what what are the important considerations that leaders should should think of in building that proper harness around AI >> So Soumitra what I would say is like imagine this you know that this as a Formula 1 you know car where you have the model as the engine and the harness is the car.

[00:28:22]

So a Ferrari engine inside a poorly designed car loses races all the time. You know a slightly weaker engine inside a well engineered car often wins. There's a good chance that it'll win. Now take coding agents as an example. People think that agent is you have a user there's an LLM there's a code which it generates right? Um It it's but it's more like uh you know, you have a user, it has a task router, it has a planner, it has a code agent. Uh I think uh then there's a review agent, there's test agent, security agents, and so on so forth. But, so there's a lot of um and a memory layer, and you know, uh observ- observability layer, etc. These are all harnesses. Now, a domain harness in our case, which you know, if you go to insurance industry, is basically, you know, your renewal workflows, your D&O placement processes, your claims triaging, etc. Right? Now, if you if you have these things defined already, like uh you know, what your domain harnesses are, then you can choose any model. The model becomes uh interchangeable. Right now, you can use ChatGPT, Claude, Gemini, Deep Seek without changing the business uh outcome. Because your harnesses are so strong at the moment. Your data with your procedural knowledge to come to a um you know, come to an outcome is determined by the harness and what context you feed into it.

[00:29:58]

Uh almost all um you know, there had been an initial uh you know, kind of a difference in how ChatGPT used to respond versus Claude used to respond versus Gemini, but they have learned from each other, uh you know, in a in a way where they're inference testing each other, and uh uh gathering the knowledge from each other. So, now I think the the the difference between what they uh what individual models can do is is kind of very small right now. Unless you're using a frontier model. Now, a frontier model is is a completely different story, but um with all these local models available, I think that the uh it's level playing the plane has been leveled. >> Uh yeah. So, I want you to um move our discussion towards another important aspect which is another hot topic that is emerging across the enterprises and this is about demonstrating the value of the investment, the ROI and the economics about AI.

[00:31:02]

So, one thing that is becoming very important for CFOs and senior leaders in general is how to manage and predict and control, I would say, the the overall spend around AI and demonstrate what are the what are the different ways they should we should they should be measuring the value. Right? So, let's split into two parts. Let's first talk about how what kind of different ROI models emerging for to measure the value. Then I want to get your attention on the on the bill shocks. That's something some enterprises are seeing and we have seen so many news in recent times on that. I want you to get your thoughts. How how because you must be talking to your enterprise customers, how you clearly advise positions and our devices customers on how they should be measuring value in their investments on AI.

[00:32:00]

>> Correct. Correct. I think one thing which we have to we cannot ignore is the fact that teams are measuring, you know, prompts, tokens, uh conversations and tasks. Uh whereas the executives are measuring revenue retention, margin, you know, risks, growth. Right? So, there apparently seems to be no relationship between these two. So, what say for example if you have an agent function which generates a renewal strategy, etc. Uh we always see the impact of it. Like what is the impact of uh having this renewal renewal strategy done, which is basically renewing your policies uh at the end of the year. And for bigger companies, there are bigger you know, kind of risks, etc. which they have to cover, better rates that they can get from the carriers, etc. So, earlier earlier you know, used to go with this manual process. It used to take like 120 days, 90 days, etc. 60 days and so and so forth. They used to start these process. Now, they can actually do it pretty quickly.

[00:33:04]

So, the earlier renewal preparation, more market options are there, better broker recommendations are being provided. There's the new business wins because of all this all of this and there is a revenue growth because of all of this. Now, see how it grows in form of a tree. So, you have you have basic prompts and token and conversation and tasks which are getting automated are actually leading up to you know, revenue growth, for example, if that is the metric you're chasing, right? Or it could be a expansion in your customer base, etc. So, we help you know, our our customers actually create these KPI trees like this, which actually go from the lowest form of how the agents are using these tokens to a fact that where you know, your it is impacting your KPIs and metrics, etc.

[00:34:02]

I think this was the first part of your question. The second part was that you know, everybody gets AI bill shocks these days. Now, traditionally, I think last year they had started doing this that you know, they they started associated tokens with with the usage that was there. In a way, it was okay to start with because we were trying to get adoption uh moving. But, soon they realized that there's no outcome out of this. There is no uh business outcome which is happening. Now, the traditional software scale, which we learned uh the hard way, was basically you have X number of users, you have X number of transactions, you multiply it, that's the uh you know, scale that you get. But, AI cost scale is completely different. You have users, you have requests, you have context, then you have agents, and you have I I iterations as well. So, one requi- user can spawn hundreds of agents, which can, you know, kind of go in a tizzy, and uh go in some infinite loops, etc., till the time they get a response which is acceptable by uh the Now, you know, the the sources of waste in this are immense. Like, you can have a context float. This is what we have seen in our implementations in implementations of our clients, uh uh where we have to, uh you know, kind of advise them, is that there are context floats, 90% of the tokens are context uh you know, context floats which nobody is using also, and they're transferring from one agent to the other. They are recursive agent loops uh which which they go through, um you know, and some of the agents are actually resolving problems with uh which are known. Uh they've already resolved that problem, and yet they're resolving it again, uh because they were part of a a new context that was never handed off. Right? Uh like premium models, uh we we did this also, where we were um all this context was going to a premium model like a ChatGPT, which is costing them tokens. But, nobody was telling them that, "Hey, why are you using a premium model? Why can't you use a um you know, you can use a local model for it, like a Mistral, which is uh Olama model, which is sitting on your uh desktop, or a container running somewhere. It'll be less expensive for you to run that local model, rather than sending it to chat GPT or cloud or etc.

[00:36:32]

So, uh we had advised our customers that instead of measuring, you know, kind of these token costs etc. Um and AI costs, you you should you should actually determine what is your cost for uh per outcome. You know, and allocate budgets per agent. So, that way they were able to rationalize it better and also see how it is impacting their KPI much better. >> Uh one thing I wanted to get your uh view was quickly around the governance and risk aspect around around agentic AI deployments. Um and we you you briefly mentioned this, but uh um the in traditional governance when we talk about uh in in the enterprise context is all about um lengthy processes and the forums and where the the risks are discussed and and decisions are made etc. Uh which most of the time is typically reactive in nature.

[00:37:34]

Right? Um but agents in the agentic world, that may be a bit too late, right? Because uh they are working continuously and autonomously etc. So, you want to make a governance as real time as possible. Right? Um so, what's the shift is required to uh how do you see the governance mean embedded in the in the flow of work uh itself. Um Uh so, that the govern agents follow the right guardrails policies. Whether it's in the the coding agents, whether it's the operational agents um in any scenario. >> Right. Right. I think uh what we have seen so far, the you know, in in autonomous agents especially, uh I've seen this evolve in four stages. First, you have your human approval stage where you know, you have your agent, it has given some recommendation, there's some human input, and uh that leads to action, right? So, agent, it gives some recommendation, the human checks it, and then the action get proceeds back, right?

[00:38:43]

Uh that is you're you're completely involving human in the loop uh before any action gets taken. Um well, it is it is a rudimentary form of uh when you're entering into agentic AI. The second stage when uh you're slightly mature, you would go to human oversight. Uh an agent will take an action, it will complete that action because it is within your risk parameters, and then a human is going to review that, right? Um the third is policy constrained autonomy, where uh you have a policy, an agent acts on that policy, and takes it uh any action. At the end of the day, you will see that is the policy being followed or not. You're not worried about the action, you're worried about the policy more. And then comes the next evolution of this stage is fully autonomous domains. That's where I think you're autonomous agents uh it's it's a holy grail for that. So, there uh these these are well uh you know, uh bounded environments. You human humans just govern the systems, and they they do not um take into account uh the agent decisions. They don't even look at that.

[00:39:53]

They were might be doing some audits here and there, but they actually look at the ROI, they look at what the outcome the business outcome is, etc. And they work along with that uh you know, kind of uh uh with that input. And uh basically say that, yeah, it's it's it's working because there's a level of abstraction that I have brought myself uh you know, on, where I'm just going to observe the system rather than individual uh decisions. >> And the other aspect I wanted to get your thoughts on is around the impacts on people, the workforce, both positive and negative, but mostly the negative uh it's kind of perception that is built around AI particularly in enterprises. Um Uh and we are seeing some many cases where uh there have been reductions in the workforce. Um Uh but at the same time many of those organizations have also invested heavily on AI. Uh AI-related capabilities. So, it looks like there is a kind of um a reallocation restructure of how the funds are allocated and spent um across across the business, right?

[00:41:09]

Um But primarily what's your thoughts What are your thoughts on the the overall perception that is built that AI would most probably take over jobs in the enterprise context? And what kind of advice you provide to to your customers on that? >> Right. I think more than customers, it's uh it's the employees who are uh who are getting impacted. Uh but I always give them a John Henry um you know, example. John Henry was one of the railroad workers, I think when when steam um uh you know, when when steam engines came or etc. steam um drills etc. had come. He had gone uh with um you know, competing with the machine, and he uh unfortunately he lost his life in that. But um the fact is because he he got tired and um exhausted. But the fact is do not compete with the AI, right? It it is there to automate your work. Even after John Henry's era, the the fact that railroad workers were now the managers of the steam press machine.

[00:42:26]

Rather than the steam drilling machine rather than drilling it, you know, the holes themselves. So, your work would considerably change as you go along. You are you may be doing 80% of your work, which is the iceberg model which MIT came up with. Um which may be under the belt, um which basically would be, you know, uh you know, working with with different interfacing with different people, coming to know about different opportunities here and there. AI is not going to do that. They might find, if they have exposure to other models, that's that's very mature model. But, most of the work which they do is like like "Let me get these claims. Let me process these claims and move forward." That takes away 80% of your work. Your employer is employers are after that. Not to that they would like to reduce your job completely. So, you so that you can focus more on that 20% which is now the exploratory work, the new work which you know, which would actually improve ROI of you know, sorry, your your your top and bottom line top line metrics. That's where the push is right now.

[00:43:46]

They're trying to create margins by reducing the bottom line by giving it giving all the redundant work to AI and so that you can focus on the top line of, you know, finding new opportunities, new ways of, you know, interacting with the AI where you can find new patterns, you can find new things to do. So, that is that is something which, you know, we we are constantly telling people that do not think that the AI is taking your job, it is making your job easier. >> Yeah. And listen also for going to the backlog, typically also the guys are on the other work to be done. And uh that backlog keeps piling up because people just don't have capacity to to focus on that. And with agentic AI gives opportunity to for people to to for organizations to then um accelerate the backlog and and get more and more features and more Yeah, more features on the market.

[00:44:50]

>> It's fail fast and you know, that's that's the strategy these days that experiment and fail fast. >> Yeah, absolutely. Uh adjacent to that is quickly a question around the operating model, how it should evolve or potentially would evolve. Um As as agents become more more common in in the in the enterprises and you see that there are agents and agents existing being being built which are more focused on particular area of the business, right? Whether this is a back office, claims you you mentioned, right? Anything else. Um Um how do you see that people agent interaction relationship and what responsibilities will evolve? Do you see different people role will will emerge or are already emerging? Um and will they have the same responsibilities now?

[00:45:49]

>> Right, right. So, what what we've seen is that there are a lot of you know, changes which are happening in the roles which are there. You are initially Uh, people were like there there's a huge um, you know, uh, force of getting into coding, learning coding languages. Now, those developers are becoming architects. Your analysts are becoming supervisors. Your SMEs are uh, are now looking at themselves as knowledge curators. Uh, your product management guys, they were uh, they are now becoming capability managers. So, it's not just a a change in the title, it's the the way that they are functioning because now, uh, we are at a different level of abstraction. Right? Uh, previously, when handlooms were there, there used to be uh, people who were uh, actually doing all the handloom work. Now, nobody looks at that because now there are machine managers who would be just adept at seeing that the input is coming in, the cotton is coming in, and the the fabric is going out, they're checking quality, etc. So, uh, so the level of abstraction has moved one layer up. And, uh, you know, the leaders are now measuring um, like, for example, number of reports, um, you know, from number of reports being produced to now they're measuring uh, to measuring uh, you know, the quality of decisions influenced.

[00:47:14]

Uh, so this is a leadership mindset shift which is happening. Um, the and even people who are who are leaders now, they should stop measuring things in an older way. They should start uh, with the AI world, they have to start um, you know, kind of uh, measuring stuff in the new way. The biggest mistake leaders will make is viewing agents as digital employees. Uh, they should be viewed as force multipliers rather than just digital employees. >> Look, as we start wrapping up now with uh, I wanted to get your thoughts on how the future is evolving. Uh, and in in the agentic world of future, could be another 6 months but we are entirely new future and another 12 months but we are entirely um, new future but uh, um, the do you see how the industrialization of agentic systems happening across the industries verticals which verticals you see are taking or progressing ahead rapidly versus which in a way verticals are you see potentially are more cautious on the journey I would say and could be a reflection of their being more regulated industries as a That's one thing I want to definitely get to get your insights on and the other question is if someone is some in the business is just embarking on the journey right what kind of advice you will offer to the CIOs and senior leaders as to how they should approach transitioning from traditionally and now to agentic AI and typically also transitioning from experimentation production. >> Right right. So So first of all where exactly is is the future of AI?

[00:49:02]

You know AI is going to be more dominant in your work and this is because of the fact that if you're doing redundant work over a period of time you have lost the you would lose the capability of thinking outside the box. People who are going to think outside the box are going to be rewarded. People who are doing stuff which nobody has done or thought of before they are going to they are going to be risk-takers and you know again taking risks is something that you you have greater risks versus you know with greater rewards. So I think your incentivizing those people has to change as we go along right and how is your management going to look at it Uh you know it's it's going to be start with the work, not the technology. Uh uh you you you have to create your champions early uh to the senior leadership. That's what my recommendation would be. Uh don't show uh don't uh just show, you know, implement it also. Uh don't do show and tells. Do implementations as well. Uh focus on quick wins just to make sure that you are um you're implementing something which is easy to do.

[00:50:17]

Um you you have to focus on that. And um you you have to redesign your incentives, like reward for business outcomes, invest in uh role evolution. Where you are your your your roles are now going to evolve. Um your your developers are going to be architects, etc. So, their metrics of measurement is going to change. Not lines of code are developed. They should be now rewarded on what kind of stable system, how many issues came in that system. That's what they we should be rewarded on, right? Um you should train for judgment uh and not just for prompting. So, your judgment should be like, "Okay, well, I think the system is going to work." Rather than, you know, that this prompt is uh you know, you know, it it worked or not worked, etc. And you have to make your adoption visible. So, if people can see that yeah, something has worked in AI, they have to incrementally build on top of that. Now, how are the leaders going to stay ahead? Um again, you have to map work, not jobs. Uh you can't think that um you know, I have got um you know, five people doing five pieces of job, so I can just eliminate one or two. Uh you have to you have to map work. If five people are doing uh pieces of work which are redundant, yeah, you can collapse that. That will affect the job. But, um you you have to see what these guys are doing. You just cannot be at the surface level saying that yeah, this guy is just doing something without looking at what he does, you know, what innovation he brings along. Otherwise, you will lose some good employees.

[00:51:59]

You know, you have to identify your institutional knowledge because that's where it actually sits with the user with the with your you know, with your workers. You have to build an AI native operating layer. Operating layer has to be AI native rather than okay, well, I'll add it as a feature. It should that shouldn't be the case. It should be embedded in your DNA. You know, and also one of the other things which I advise my customers is develop an AI finops discipline early in the game. How many tokens are you using? What is the you know, kind of how many agents how many tokens are is consumed by each agent etc. That needs to be clearly called out. Right? You have to again, reiterate that because you have to redesign your leadership metrics and also develop agent supervisors. You know, you just cannot let the agent run amok right now in the initial stages. You have to see the outcome. You have to see how it's working. And you know, experiment aggressively.

[00:53:08]

Scale selectively, but experiment aggressively. Not everything which works in your demo environment and your your sandbox will actually be scalable. And you have to be very cautious about how you scaling stuff. Just focus on your decision advantage and be prepared for a hybrid workspace because when you are preparing for a hybrid workspace, you are you know, kind of contending with the fact that there's going to be some level of digital workers who will be working going going to work with you. Whereas there're are going to be some actual people who would be working with you. So, make sure that you have that clarity and that accommodation that this is a digital workspace and a digital worker would be able to respond much faster or even on Sundays, etc. But, don't have the same expectation from some of the people who are, you know, kind of in the non-digital space.

[00:54:05]

>> Monday night, thanks for the joining interesting talk. I think so the time it was as usual really good discussion. We covered so many topics. >> Thank you, Sumitra. Pleasure being on the call. >> If you found this discussion valuable, please follow and subscribe to Enterprise Tech Talk. And thanks for listening. I look forward to seeing you in the next episode.