Transcription
This is Jane Lo and [music] I'm at the UiPath office here in Singapore and with me today I'm very pleased and very privileged to have Matthew who is going to be sharing with us on the latest developments in AI agents or agentic AI. So thank you so much uh Matthew for your time today.
Well, thanks for having me Jane. Happy to be here.
Yeah, great. So a big topic AI agents. I think for many of us, right? When we think about the word agents, right? We we can't help but think about, you know, the kind of agents that we imagine in movies like 007, right? With a wide range of skills, you know, high levels of access and you know, high degrees of autonomy. So when we talk about agents in AI space, it's kind of similar in many ways, isn't it?
Um well, AI agents in the software space is highly autonomous as you mentioned. Uh a software that's able to reason, that's able to break down complex task into smaller task um and achieve the goal that you want, right? So it leverages tools like access to data and access to systems to eventually deliver the outcome that you've asked for it.
Yes, so I think one question that many audience would have is, you know, how different is the current iteration or generation of AI different from the previous generation where there's so much focus on automation, right? Um so, you know, can you share with us some exciting examples that you see in, you know, where the current iteration of AI is displaying some of the this new capabilities?
>> Mhm. So in the traditional sense of automation, you're typically developing automation that's very rigid and in terms of the goal that it needs to achieve, you have you have to prescribe each step, right? As opposed to building agents where again, agents are able to deal with complex environment, ambiguous data, unstructured data, but still uh deliver a certain outcome. So what we see actually is the need to combine both approaches because they're still part of our day-to-day, for example, that's you know, data entry uh which is rule-based. You're going to be entering data into certain fields in an application, you could rely on the traditional automation approach to perform that work, but there's also part of our work where requires our cognitive skills where we're able to then leverage agents to, let's say look at unstructured data like contracts, um medical reports, and then perform the necessary assessment. Uh I'll give you an example like um in in the medical insurance space, right? You have claims assessor that look at medical reports, they look at what are the claims that this particular customer is eligible to. So a lot of these um work where you're it requires you to read documents, uh look at look up certain information from let's say the policy, these are things that you can actually get an agent to perform, which was very difficult to automate with sort of the traditional automation approach.
Right, okay. So I think um the next natural question is, you know, how do users proceed from automation to this current iteration of AI? Or is there a need to actually start from automation first?
Well, it it it automation is a really good starting point. In fact, one of the practical advice that we give to our customers is look at the existing processes that you've already automated and look for opportunities to add AI agents into those automation. Like for example, if I'm in finance and I've automated, let's say, an accounts payable process part of um accounts payable is to ensure that if um an invoice comes in that the details of the invoice matches a purchase order that was released by the company, right? So are we paying for something that we actually purchased? Right? Performing that reconciliation. Um we've been able to automate uh accounts payable for many, many years with our robotic process automation capability, but one of the things that we noticed was there was a lot of false positives. Right? Like the robot saying that the invoice doesn't match the PO. But in fact, in most situations, there is a match just that maybe certain terms were not the same. Like for example, in the purchase order it said one carton of milk. But in the invoice, one CTN of milk. If you're relying on traditional approaches of just, you know, making sure is A equal to B? Is string A equal to string B? Then chances are the robot will say no, it's not a match. But AI agents have the ability to assess semantic similarity. It understands that CTN is a short form for carton and therefore it it will deem the PO matching the invoice, which will eliminate a lot of the typical false positives that we see in processes like
>> put into the agent's configuration to ensure we minimize those risk that we mentioned earlier. Um yeah, I want to explore further about the mitigation measures with you, but before we do that, right? Before I forget the question, you earlier you talked about unstructured data. So for example, feeding the AI with invoices, right? And also you talked about prompt injection. Now, one of the emerging risk is indirect prompt injection where you read a piece of unstructured data and then wherein it contains malicious prompts in it. Have you encountered situations like that?
Not necessarily in our customer implementations of our platform. Um we we know about such possibilities. So again, it's really about how to make sure that the data that the agent's being given access to are sort of tamper-proof and are legitimate, right? So there's of course many ways to introduce data to the agent or documents to the agent. One is to make sure that the agent's able to access that, say, a really secure SharePoint site as an example or secure OneDrive or secure Google Drive where there's data custodians that ensure that the data that the documents and the data that's stored in those repositories are trusted, that they're up-to-date, right? And that you don't just give anyone access to that information. Of course, the risk is if you're giving agent the ability to access external data. Yes. Because then your company you doesn't have control over those external data. So I think that's really the type of you know, assessment that every company has to do is what type of data you're giving to the agent and how secure, how reliable, how up how up-to-date the data is, right?
>> Right. Yeah, okay. So, earlier you talked about the different sort of mitigation measures, right? And I think a lot of the audience have heard about like safeguards and then also ensuring that the access predefined beforehand. Certain actions are allowed or certain actions are not allowed. How is implementing these mitigation measures in you with agentic AI different from, you know, the old AI automation
Mhm. process?
Yeah. Well, it it is a very different world, right? Um because the typical sort of cybersecurity assessment that you do on traditional systems are, you know, you kind of have sort of predefined templates already that allow you to assess um let's say a system whether it's a traditional automation or traditional AI. You kind of have the ability to assess um sort of the risk whether it's from cybersecurity perspective or maybe from a data privacy perspective. But in the world of agentic AI where um as we said, the agents are have the ability to autonomously, you know, derive an outcome. Um you might give it access to different systems whether internal or external. It's kind of hard to in a way predict what the agent will give you and where it's getting the data from and and so on and so forth. So, well, it is indeed important again to make sure that um one is you're really evaluating the output of an agent. So, you have, let's say, a hundred known scenarios with the known outcomes, right? So, you feed this the hundred scenarios to the agent. Does it give you the same outcome that you expect, right? So, that's one way it's to to do an evaluation. And I think the second is really what we see is there are there are teams being formed that kind of you know, tries to break the agent, right? Like it red teaming, correct? Where you try and inject some prompts or kind of let's say have a conversation with an agent and then until you you know, make it share information that you're it's not supposed to be sharing, right? So, I think that that that approach is actually quite relevant today where red teaming. In fact, we do that as well in UiPath to ensure of course that our products are are you know, red production ready, right? So, those those definitely those type of approach that's again very different from the traditional AI or traditional automation. The red teaming approach is very very relevant today.
Yeah, I see that it's becoming increasingly important and a point of focus in terms of making sure that it's production ready. Another question I think many people have is they hear the term human in the loop, right? But it's very difficult I think for many audience to really grasp how and when, you know, you you deploy this human in the loop, right? How how have you seen your clients doing that?
>> Oh, yeah. So, we have, for example, a customer in the telco space whose developed an AI agent for churn prevention. Right? So, basically they this agent assesses the chances that a customer might not renew their subscription. Right? So, we're talking about enterprise customers though. So, the B2B business of the telco, right? So, once a churn risk is identified um we have agents also that will build what we call a strategy playbook. It it will um recommend this is the solutions and products and mitigation that you should propose to the customer and this will be your talk track, Mr. RM, Mr. relationship manager. So, the human in the loop there is we're not asking the agent to go and contact the customer and offer this mitigation plan, this set of new products and services. We're just preparing call script for the relationship manager to then engage the client because at the end of the day businesses buy from people. They don't want to be buying from a robot or an agent, right?
>> [laughter]
>> So, the agent's responsibility is just to create that strategy playbook, that talk track, the call script and then give it to the relationship manager so the relationship manager can have a conversation with the client, right? So, that's what we call human in the loop. In other situations, you have people that are actually making the final decision. Like, for example, when I brought up claims assessment earlier, right? So, you do have sort of the analysis of the agent in terms of uh whether we should be paying a this claim or not, but at the end of the day it's still a person that makes the final decision because the agent has presented all the all the data. This is the analysis that was performed by the agent and eventually the conclusion with um sort of citations and the reasons why it came to that conclusion of let's say paying the claim, right? At the end of the day it's about having that human in the loop which is the claims assessor making the final decision.
Do you see your clients are putting some sort of a uh thresholds in terms of, you know, for example, the claim assessment, right? If it the claim is above a certain, I don't know, dollar amount, the human will step in, otherwise it goes through or not.
Correct, correct. I think I've seen this for many years already, right? Like, for example, if you bought something through an e-commerce platform and that only cost $2 and then you, you know, you decide to return what you purchased, it's cheaper for the e-commerce platform to just give you back your $2 without question. Rather than involving a person to actually look at your claim and then make an assessment of whether to pay it or not, right? Or to refund or not, right? So, in similar situations, you do have those smaller risks type of activities where yeah, if it's a claim below 200, the agent can make the final decision, but eventually for larger claims, you would want that human in the loop, right? That's right. So, yeah, so we've seen our customers sort of apply those thresholds. Um because again, there there's certain um low risk activities where you can make the agent make the final decision like paying out something less than 200 versus a higher risk type of situation where
Right. Okay. Yeah, that brings to my my personal experience on e-commerce sites why some certain purchases where the refund process is a lot quicker than than others.
>> Correct, correct.
>> So, when it comes to AI, a very popular question, right? Is, you know, how is it transforming the workspace or the jobs, right? How what has your experience been?
Yeah. Well, personally, I would say with the help of AI and AI agents, I have a lot sort of a lot more access to information, right? It allows me to, let's say, prepare for some of my customer engagements because I'm able to do, let's say, deep research on a client that we're we're supposed to see, maybe even some deep research on a specific person that that we're meeting and so on and so forth. I think this that aspect of sort of personal productivity um that we're seeing is like um helping you write a proposal, help helping you write emails. These are examples of where AI agents has provided sort of a lot of gains when it comes to personal productivity. But at the same time, I think there's also um an opportunity for AI agents to be embedded in business processes so that business processes become more efficient. Um we can help um companies achieve certain KPIs that they want to hit. I'll I'll give you an example. Um UiPath actually developed um AI agents for Omega Healthcare. Omega Healthcare is a revenue cycle management and healthcare outsourcing um company based out of India. So, um Omega Healthcare leverages UiPath UiPath's agentic AI platform to help them automate a lot of accounts receivable, medical claims type of processes. So, when it comes to let's say medical claims, as I mentioned earlier, you got to look at the medical report and this is the the claim that the the patient is submitting. You got to make sure that when they say they went for an x-ray that the the diagnosis actually matches the recommended recommended treatment, right? Like for example, if the medical report says you have a fever, why are you getting an x-ray? As an example, right? So, a lot of these things are actually now handled by AI agents, right? Which allowed Omega Healthcare to drastically cut down accounts receivable processes by as much as 20 to 30% in terms of turnaround time. And they were able to also increase what they call the clean clean claim ratio up to 97%. Clean claim ratio is the percentage of claims that are paid uh after the initial submission without denial, without asking for more information. So, 97% is actually above industry average. So, these are the sort of the the impact that we've seen agentic AI bring to our customers.
Right. Okay. Amazing. There's a lot of productivity gains that we have seen in the last, I don't know since ChatGPT came into the scene, right? So my final question is, you know, you have provided a few exciting examples or opportunities where AI agents can add to enterprises and also to our personal lives. Looking ahead, where do you think are some of the untapped opportunities?
Mhm. Yeah, I think it's really around moving from individual tasks automation where you might leverage AI agents for a single activity to more of how do we orchestrate a full end-to-end process, right? Like employee onboarding, customer onboarding, vendor onboarding, or whether it's to orchestrate a whole lead to cash type of process, right? Where we're now envisioning a future where or maybe not the future because we can make it a reality today. But what what we're envisioning for our customer is the real full orchestration of those journeys. Yeah. Where you're leveraging traditional automation, AI agents, and of course keeping people at the center of the process, right? So at UiPath we actually have a solution we call Maestro. It's like the maestro of an orchestra, right? Where you have your wind instruments and your string instruments being orchestrated by a maestro, right? Where we're able to model an end-to-end business process, indicate whether certain steps within that process are manual or whether they're executed through traditional automation or whether they're executed by an AI agent. So I think what where we see this going is really as I mentioned earlier also is really embedding that agentic capability within a business process.
Right, yeah. There's also a lot of conversations now about AI fleets, right? Like you say orchestrating or coordinating AI I guess teams.
>> Correct.
>> Right. Building a chain of AI agent processes.
>> That's right. That's right. Because we see a future where you have certain agents working with each other, right? An agent, for example, that does dispute investigation to another agent that does dispute resolution. And maybe these two sort of working hand in hand. Dispute investigation agent handle identifies these are the disputes and why that dispute came about and then a resolution agent that might use the data that or the outcome that the dispute investigation agent has created to help it perform the resolution. And maybe it needs more information, goes back to the dispute investigator to get more information before it's able to resolve the dispute, right? So yeah, we see this future as well and I think UiPath is one of those companies that's able to orchestrate multiple agents, ensure that agents are sort of working well with each other and eventually again help our customers, you know, move to the future of agentic AI and reap benefits
Yeah. Yeah, there's a lot of talk that, you know, the future organization is no longer just managing humans, right? It's managing a hybrid of humans and AI agents, right? And perhaps in the next podcast I'll be talking to a avatar of you or you talking to avatar of me, right?
>> not not that soon.
>> [laughter]
>> I do love my job, so I hope to keep it. Or you delegate you delegate opportunities are there?
>> there are certain things that I can delegate to sort of my digital twin. Yeah, and me too.
>> [laughter]
>> All right, but for today so thank you so much for your time today.
>> My pleasure. Thank you.