#333 Adi Kuruganti: Why Your AI Pilot Is Failing and What It Takes to Reach Production

0:00The one customer example is a financial services customer, a bank essentially, who is able to reduce their loan application process for automotive loans from around 12 hours to a little under one hour. In all these use cases, we definitely have human in the loop. Once the human improves, the agent is then taking action to process it as well. I don't think the agent technology is at a point where you can get 90 % plus accuracy all the time. And so there is a trust factor for these mission critical processes that drive operational productivity or regulatory compliance or revenue impact. We believe it's a combination of deterministic and cognitive.

0:36AI is moving fast within the enterprise. Employees are experimenting with personal AI accounts. Teams are building custom AI apps and autonomous agents are connecting to sensitive systems. Innovation is exploding, but governance isn't keeping up. Gartner predicts that through 2026, 80 % of unauthorized AI transactions will come from internal policy violations, like oversharing sensitive data, not external hacks. That's a huge risk. And that's why Island, the creator of the enterprise browser, launched its new AI services. Island surrounds genitive AI, AI browsers, and autonomous agents with the enterprise controls that they were never built for.

1:27Identity enforcement, data protection, auditability, and centralized policy. Now organizations can safely use any AI app, consumer or enterprise, without corporate data loss. Teams can deploy secure permissioned agents with full audit trails. Island's publishing capability makes it easy to safely share internal AI apps across the company. And the Island AI browser brings governed multimodal AI directly into the enterprise browser employees already use. The result? A single secure foundation for AI at work. So CIOs and CISOs don't have to choose between productivity and protection. To learn how to scale AI without losing control, visit island.io.

2:24Adi, usually I start by having guests introduce themselves to listeners and give your background so far as it's relevant, but more importantly, how you came to Automation Anywhere and what exactly Automation Anywhere is doing. Yeah. All right. So nice to meet you, Craig. I'm the Chief AI and Development Officer at Automation Anywhere. I run our product, technology, R &D teams. And I've been here about four and a half years. Before that, I spent about 15 years at Salesforce.com. So I've been in enterprise software now, I guess, going on 20 odd years. Very much in all things product, for CRM, now in all things automation.

3:17a big focus obviously on agentic ai and agentic and i'll talk a little bit more about automation anywhere so uh so that's my background i think just to maybe talk a little bit about automation anywhere yeah we are in uh you know we've defined a category called agentic process automation which really combines what we believe is the best of deterministic automation traditional automation with agentic AI, but with the goal for customers to automate their mission critical processes. Things like auto management, prior authorization, healthcare, anti-money laundering and financial services. So those are the kinds of enterprise grade, mission critical that typically span different applications, different systems, legacy, modern, as well as obviously human decisioning.

4:06So we often touch financial systems, healthcare systems. So it's really important that you get the enterprise-grade governance and controls around it. We started, we created this category before this, you know, we created this category called RPA or Robotic Process Automation, I would say about, I want to say 12, 13 years ago. That was a start and is really more focused on automating repetitive tasks, you know, in a deterministic way. and obviously that category took off and you know you have a couple of our couple of companies still remaining uipath being one automation anywhere another one but with the last four and a half years has really been expanding to what what we're now calling agentic process automation so again we're in that space so a lot of obviously focus on how do we use agentic ai to enable customers to automate mission critical processes.

5:04That's our lens. There's a lot of agent AI. We talked about agent AI for everything. We want to use it at the right time because we still believe there's a lot of value to customers to combine deterministic and agent AI. It's not one or the other. It's the combination of the two. It's a fascinating area right now. I just got open claw working on my computer. of hopefully on a different terminal though different computer because uh you could get some issues if it's your main computer but yeah yeah yeah well i'll take that uh under advisement because i i haven't been very careful with it uh but but the there there is certainly i mean that to me uh has been kind of the gpt moment for agents i mean i i know enterprises have been working on them for a long time you talked about salesforce and and they're with agent force i think the the leading one of the leading uh agent building platforms uh but you guys are focused not on building agents but uh but making sure they work in an enterprise so we are we we obviously you know as part of making sure they work in enterprise we want to we also enable customers to build or what we call process agents because it has to have the process context.

6:31But we all, you know, one of our key values is having an open platform because, again, as a process automation company, we need to make it easier for customers to connect into different systems and different applications. So it can't be a wall garden, right? So we work with Salesforce, ServiceNow, SAP, old school, terminal servers, healthcare systems. So it has to work across and it has to be very seamless. And so for that, we always do our customers build processes on our platform. They use our agents. They're using MCP. They use agents from Microsoft, from Salesforce, other vendors, crew.ai, open source platform.

7:12So it doesn't matter. it's ultimately up to the customer to use the right agent for the right use case and right task uh but that's and but really orchestrate use our platform to orchestrate the end-to-end process and with the governance around it yeah uh and there's been a lot of talk uh but before we talk about uh automation and where's role in this enterprise has been struggling with agentic ai i I mean, there's a lot of promise, a lot of mistrust. And so far, I haven't seen anyone, although I hear about it, but I haven't talked to anyone who's using it in critical business processes, as you mentioned.

7:56So what kinds of critical business processes do you see agentic AI being applied to in the enterprise? So today we have over 5 ,000 customers broadly. But of those customers, we had about 1 ,500 live deployments in production. Now, many more in POC world, and that's one of the challenges, how do you go from POC to production? I will talk more about that. But really typically, it's, again, the three types of use cases that we see. One, kind of the bread and butter, which is operational productivity. So things like cash flow, how do I reduce accounts payable and tie into a process so I get cash, sorry, accounts receivable or get cash upfront or supplier management.

8:47There's certain things that you want to get cash faster. You want to have and really focus on your outpatient productivity. That's one use case. The second is around regulatory and compliance, big use case around that. So things like how do you avoid overpayments, but also underpayments on taxes based on different state or country rules, because it varies for especially these large global organizations. Third, seeing less of it, but it's coming, starting to come is kind of think of it like new product line. So one customer example is a financial services customer bank, essentially, was able to reduce their loan application process for consumers like you and me for automotive loans from around 12 hours to a little under one hour.

9:38And because of that, they were not only able to process more loans, but able to win, you know, basically a deal, a business from one of one of the large automotive manufacturers out there. they were competing against other banks. It's kind of operational effectiveness, but also using that to win new deals. I'd say the use cases are more in the first two. So typically departments like finance, operations, that's where the benefit is. And again, it's not really about using agent AI purely, like just use agents and figure out and let them go crazy. It's about combining deterministic with these agents.

10:15Well, for example, contracts, it's a great use case because agents, you're thinking about unstructured content, like contracts, using generative AI, and then agents to kind of figure out what is the based on certain customer profile, what's the right loan APR you want to give this customer agent, it's a great use case for an agent to look at that customer profile, that task, and figure it out based on the structured content, giving options on loan APIs. So those are the kinds of use cases. But in all these use cases, we definitely have human in the loop, namely human decisioning. None of the use cases so far are just like agents go do your thing and we'll hope for the best.

11:00It's definitely human decisioning at various checkpoints yeah and and so i mean i have an agent you know following my emails uh and it it writes drafts and then i go through and and approve and send them is it that sort of thing when you say human human and loop that the agent in for example a loan application would do all of the necessary due diligence and tee it up for a loan officer to look at. So all the information's there and he just makes sense. The one difference between what I'd say in a consumer, personal productivity, where you can have an agent look at an email, summarize it. And, you know, you can review it before you send out, let's say, email to a prospect, a customer, to a friend, whoever it might be.

11:51In the world of enterprise, looking at this, for example, a loan approval, you have to look at the consumer, you have to look at unstructured content like their bank statements, other information, looking at the financial profile, and then look at your company database, essentially a knowledge database to say, based on the profile, the risk profile, what is the APR that you might, loan rate you might provide. so you'll have to look at different systems systems of record as well as knowledge and then may kind of provide some uh suggestions and that's where the human comes into play where you earlier a human would have to do all that due diligence figuring out all the details but the one caveat is here then you're taking action once the human approves the agent is then taking action to to process it as well so it's not only about the knowledge angle of it to get information but it's also the action piece of it and that's why there's a higher level of governance and guardrails because ultimately then you're processing it through these financial systems or healthcare systems and from a transaction perspective and that's where you know agent tooling for example comes into play where you can you know run a process in sap or run a process in whatever other system and agents have that capability yeah do you do you see a big difference because I'm hearing this from people between companies that have legacy deterministic software running their automations and AI native or agentic native startups that are building everything from scratch.

13:34I'm sure you're aware of Sam always. on the first single person with an army of agents to build a billion dollar company. I mean, that's, yeah. Yeah. I mean, so what's exciting about the category we're in is, you know, you have the big enterprise vendors who are getting to this broader agenting automation category. You have the new startups, AI native startups. And it's not either or, I think depending on the use case. We believe that automation anywhere, it has to be a combination. Because when the AI-native startups, I'll give you one example. When a, let's say, AI agent, AI-native startup, like any Dan, whoever, will talk about concepts or crew.ai, let's take crew.ai.

14:24We'll talk about concepts like agentic orchestration. But what they mean by that is that the agent is deciding based on intent, how to orchestrate. And so it's all cognitive. It's probabilistic, right? And whether it's agent calling another agent. So remember, one probabilistic step calling another probabilistic step is going to be a lower, potentially, instead of 100 % accuracy, it might be even lower. But you're literally leaving the agent to go figure it out. And there are potentially use cases for that. And in some ways, Opel Claw is an example of doing that. And that might work in certain scenarios.

15:00In our world, I think one is the level of trust, but also the other is, frankly, customers still have obviously legacy systems behind the firewall systems, that where the deterministic processes play really well. But also, I don't think the agent technology is at a point where you can get 99 or 90 % plus accuracy all the time. And so there's a trust factor. And, you know, it was still early days. This technology is quickly evolving. You know, 18 months ago, there was no consider agents. Now, there's agents and the technology is fast evolving. So today, we think for these mission critical processes, and I want to be clear, not for every use case, but for these mission critical processes that drive these business outcomes around operational productivity or regulatory compliance or new revenue impact, i i do think we believe it's a it's that combination of deterministic and cognitive today we feel it's the 80 20 rule that's what we see with our customers will it become more 50 50 60 40 and eventually 20 80 time will tell i think it will evolve it will definitely evolve the speed of the revolution though it's not only whether one is the technology second is the governance and compliance around the technology but that's also evolving things like you know agent evals still evolving i mean in terms of figuring out how that agent actually perform and then in production does it perform what he thought it would perform because those things are really important in you know especially in a you know business enterprise environment and the third is the just human trust like do i trust this uh system this cognitive system just to do everything without any human intervention that's where i think it'll take a little bit time in in you know industries like healthcare and financial services and other and kind of those kinds of industries but i think it's starting yeah yeah i wanted to ask you about orchestration process intelligence predictive resource execution how those enable uh machine critical automation can and how automation anywhere applies those things.

17:20Yeah. So the core kind of the central nervous system is the orchestration, because that is what orchestrates across these legacy systems, the decisioning within that, does it need to call an agent? Does it call an API, a bot, process some documents, whatever it might be. And I think the benefit of having the orchestration engine, our engine is called the Mozart orchestrator, essentially. It can automate across any application, any system, as connectors, pre-built connectors, customers can create their own. But also you have tools like MCP, both inbound and outbound, to call a process to be orchestrated or to call another tool, which may be another system, through the MCP outbound.

18:07So it's kind of the core system where you can have a mix of deterministic steps, regions, and human intervention. All that, every time a process runs, so that's our core engine, right? So every time a process runs, though, there's a lot of data and metadata captured. So we have over 400 million processes at any given time running through our system. So while we at AutomationMAA obviously don't look at the business data flowing, because that's a little bit irrelevant for us, we can see, and it's also by the regulatory compliance thing that we can't even look at the business data, we can see the metadata flowing.

18:44We can tell how an audit-to-cache process works. We can tell how prior authorization process works. We can tell how AML process works. So it's using that we've created a system called the Process Reasoning Engine with the goal that how do we improve the outcomes for a process? So when our agents are performing, not only do they have the tools and the goals and the memory management, they need the process reasoning engine so that past behavior of those agents can be used to improve the outcome the next time the agent performs. And that's kind of our focus, which is not about general purpose agents.

19:24It's really about how do we enable these process agents as part of these long running processes powered by a Mozart orchestrator to have higher efficacy outcomes and lower the hallucinations. So that's where talking about how do we use the metadata or the data, that's really the process reasoning engine which is driving that. And then obviously we have the full governance and compliance and things like agent evals, masking of PII data. So the full governance piece is also really important because the first question we get from customers when they're looking at any kind of agentic processes. What's the governance and compliance?

20:02Because we've got to make sure not only the CISO, but the legal team are fully aware. Because often customer data is flowing through these processes. Yeah, and on the Mozart orchestrator, is that deterministic or is that based on LLMs and generative reasoning models? So the agents are based on, you know, they have gold, we have gold based agents and reasoning, all that stuff. But the Mozart orchestrate itself is a deterministic process that includes the ability to embed agents as an action. So it's basically think of agents as a first class citizen within the Mozart orchestrator. Most of our customers will build when they do an order to cache process, a prior authorization process.

20:54They build a process as part of Mozart orchestrator. Now, it doesn't prevent a customer from using an agent directly as well, but we see less of that because they want that control of a kind of a deterministic, like I said, a central nervous system where they can control how the agents are being used. And they use the agents at the right time. That's the main thing. It's where the agency uses the right time, you're not using agents for everything. So for example, if you want to, you know, get some information from SAP, and you know, you have to get that information, you don't need to use an agent for that, you can use an API for that.

21:32Yeah, what is predictive resource execution? So predictive, no, so it's not, it's called a process reasoning engine. That's what it is called. But essentially, it is using, conceptually, we have a lot of metadata. Using the metadata, we basically at the core created fine-tuned models for process use cases. So example, we have fine-tuned models for how do you extract and process a contract. We have things for order management. We have different legacy systems that our customers interact with. So how do you automate those legacy UI systems through traditional RPA, as well as how do you build processes?

22:19So there are lots of different models that we have created on top using that metadata. And then there's an entire context graph around it, which is saying, okay, now we've created these generic models built on LLMs, but it has to be contextual within a certain within the customer's specific business. So for example, a though, you know, a manufacturing customer is going to be really very different from a financial services customer. So that's where the context graph comes into play, which is a combination of their business data, could be the SORs, systems or records, could be their knowledge graphs, but also all the processes that are running and the agents that are running, that is also context for us.

23:00And so it's a combination of that. And then you add, you know memory management and um and self-reflection basically learn constant learning that is what we call the process reasoning engine and again the goal there is how do you improve the outcomes of these process ai agents because uh you can build it but it has continuously evolve and because these are highly probabilistic in nature yeah okay and yeah i'm sorry i had pre i had the wrong uh that's why i enter more is yeah yeah uh so um how is ai transforming traditional uh robotic process automation into more adaptive and autonomous systems i mean you've talked about that already but uh where does automation anywhere see the biggest impact i think the at a macro level AI, agentic AI, and we believe this is why we've gone all in on this.

24:04It's completely transforming the industry and our customers in a couple of different ways. One, at a core, we believe that customers can now automate more. The reality is with traditional automation, whether it's RPA or just kind of process automation, there were limitations because you could only automate what you knew. It was very well defined. It had to be very well defined. But now when you combine what you know with what you may not know with the exception scenarios, now you can use agentic AI to deal with those exception scenarios or deal with scenarios where, let's say, you have unstructured content.

24:47And so that combination at the core allows our customers to automate more of their operations, which is, I think, a big deal for them because ultimately it leads to real business outcomes. And I think from a technology standpoint, I'd say they are big. They've been big benefits of our customers. One example I'd give you one of the challenges of traditional, let's say, RPA robotic process automation, because essentially it was mimicking user behavior. And so it was going into, let's say, logging into an SAP and automating that screen and blah, blah, blah. But if that screen changed, then the bot would typically fail.

25:25Because ultimately the screen is changed, the experience is changed. well, the field is no longer there or new fields are there. We rolled out this product called Genitive Recorder, which really uses vision models as well as Genitive AI, but also vision models. And we have a lot of understanding of the DOM structure because we have this history of RPA. So we created a model around that, a fine-tuned model. Now we're seeing enabling our customers to drive 60 % higher resiliency of even pure play RPA. That's just one example. There are many other examples like building processes using natural language, but essentially allows, you know, higher resiliency, faster time to build, right?

26:06You can build these processes and you don't need to be an expert to build these processes. You can build it with natural language because you know what you want to build, but you don't need the technology behind it. So the barriers to drive business change have lowered tremendously. And you can see this, whether it's OpenClaw or other cloud co-work, which is in preview beta, all these tools are reducing the barriers as far as you and I can just, and anybody can go build these. And that's the biggest benefit I'd say for our customers, because frankly, you don't want a central IT team to be the only ones to do it.

26:49You want the line of business and their folks to be able to drive efficiencies and outcomes within their own business. That I think is another big benefit. Yeah. Can you talk about how do you work with a customer? I mean, say that you know, a highly regulated company in a highly regulated industry, say a bank or a hospital comes to you and they have some low-level automations in place, but they want to expand that automation with a Gentic workflow. So how do you go in and do an audit initially, or how do you work? So one is the big three industries that we've traditionally worked with are financial services, healthcare manufacturing.

27:43So we have a lot of data, a lot of history and forget the technology used, but we have a lot of history with the outcomes that we can drive. So typically we have this process called the BVP or business value proposition, which really looks at outside in looking at maybe the financial statements and other areas where what do we perceive as the potential areas to impact, areas of impact. So what are the outcomes we can drive? But that's an outside interview, right? Because we don't, at that point, we've not necessarily got in-depth with the folks within the company to know more detail of what rack actually they care about.

28:23So we have outside interview, and then we work with the key influencers, whether it is somebody in the CFO's operations department, the CFO, somebody in line of business who really then map that to what are the outcomes that they care about? because all of us have outcomes that we care about and some other outcomes are not as high in priority because you always want to look at that first couple of areas or use cases to say this use case actually there might be benefit because it has a mix of okay there's a process maybe a long-running process but it has unstructured content and it impacts maybe cash flow or there's some outcome the mistake i think folks made at the start of this entire push into generative ai and to ki was it was folks are trying to use it more as a technology tool versus thinking about okay what outcome we want to drive because you know the technology is cool if it doesn't drive an outcome at least for our customers you know it's kind of useless at that point because what am i going to do in an enterprise setting so that's the first step of bvp with trying to identify the top two or three outcomes we always have a you know buyer or a influencer who's kind of owns that outcomes on the customer side.

29:38And then, you know, we obviously work with it, we do a POC, a solutioning, we have hands on sessions for the technical teams where they can actually build using our platform. So it's kind of a multi level effort. And that's typically what drives, you know, the selling motion and then post sales, what we found what now we're, especially last 18 months, what we realized is AI and agentic AI is not the same as deterministic. At deterministic, we could allow a customer to just go do it on their own, frankly, or maybe a GSI helps them. We are very high touch with agentic AI because even defining a goal and define the right tools and kind of optimizing that output, it's not easy for our customers.

30:24So we have this concept of pilot to production where we help our customers roll out the first set of use cases and train them while we're doing it. And so while once the first set is successful, then they can take it forward. So it's the end to end. It's not like a traditional what I was used to at Salesforce where you're selling the software and then yeah, you have services, but when you're not thinking about it end to end. Now I think all of us have to think about it end to end because a customer is not just going to be able to go figure out how to deploy these agent requests. Yeah, and that's something I see a lot of these agentic builder platforms that are being pitched to enterprise, unless you have someone that really understands agentic AI on the customer side, just having the platform without some guidance doesn't really solve the problem.

31:23And as you said, when I mentioned using OpenClaw, you know, there are all kinds of security vulnerabilities or potential security vulnerabilities. And you want a company that understands that to make sure you're not opening yourself up to things. So what are the challenges that enterprises face when deploying agentic AI at scale? And what do you see the main challenges has? I think a couple of things. So one, we just talked about it. I think this technology is evolving so fast and so much hype sometimes. And there is also a lot of vendors who interchangeably use the term agentic AI, assistants, chatbots.

32:13It's all confusing. But everybody is also talking about agentic AI from a slightly different vantage point. There's a different lens, right? How Salesforce talks about agent force is really focused on sales cloud, service cloud, their applications. How we talk about agentic process automation is from a different lens. It's a process lens. How ServiceNow we'll talk about is a little bit different. So everybody is a little bit different how they are approaching agentic AI. So that's one big challenge for a CIO to understand, okay, what do I use? Because you also don't want like 100 tools that are all having silos of application.

32:53That's one. But once you figure that out, and I think that's the forward leading CIOs understand when to use what. The second is, you know, I think things like pilot to production are really important to enable that deployment. Figuring out the right outcomes that you want to go after versus treating as a technology problem is super important. That is like a make or break for companies and those who know what outcomes they want to go after. And there's a line of business and technology at the table. Similar problems from the past is no different from when we were talking about just deterministic automation.

33:30Because even deterministic automation of processes or deals, if you don't have a problem you're trying to solve, you're just using for the technology, it's not going to work. But because even more now, because there's a cost of tokens, there's a cost of using LLMs, and there's a pressure from the top to say, what are the good outcomes that you're seeing? And then I think if you go to the technology, there's definitely a knowledge gap for some customers. For example, what type of data context do you want to use for an agent? That's what we would see is, okay, hey, here's all my catalog, my SharePoint servers, my knowledge.