Transcription
[Music] Good morning, good afternoon, good evening everyone. Uh, very good to meet you. Uh, my name is Dan Bagley. I run the pre-sales and solutions engineering teams here at uh, Lena AI. Uh, coming to you live this afternoon from the east coast of the US, uh, Northern Virginia, just outside of Washington DC. Uh, surprisingly nice day for August here in Northern Virginia. Normally this is uh prime swamp weather for us, but it's actually a uh very pleasant day. It's been the coolest August on record in Washington. So I will certainly take it. Hope it is uh nice where you all are are as well.
Uh, goal of today's session uh a little bit more of an educational session. Uh, if you've joined this recurring series, you've seen a couple different topics involving uh demos and and interactive discussions. So today we're we're going to focus a little bit more on education and and focus around some of the architectural concepts that one needs to keep in mind as you're building out an agentic AI architecture for the enterprise. Uh, with that being said, I I do want this to be uh relatively interactive. Uh, two main parts of today's webinar. We'll we'll start off with a with a few slides and then the uh center section of the webinar will be focused around sort of the architectural solutions that we put together to to try and solve some of these challenges. So, uh, please feel free to use both the chat and the Q&A feature to to ask, uh, questions and and, uh, make comments throughout. Uh, I may pause briefly before the, uh, the architectural piece, but, I will definitely pause following the architectural piece and we'll be able to go through uh, questions. So, feel free to populate that as you go and and we'll certainly uh, jump into them today.
Uh, so uh, in terms of uh, agenda, we're already talking from from an introductory perspective. Uh, we'll talk a little bit about what agentic AI means both to me and in terms of a a practical enterprise assistant environment. We'll talk a little bit about some of the challenges, what that means for your employees, what that means from you from from an implementation perspective. And then we'll of course talk about the architectural considerations that that we've seen and we've found from a uh a solution team perspective. And as mentioned, I'll we'll follow up with some Q&A following that. All right.
Uh, so to kick us off uh a little bit of a conversation or starter uh I'll I'll ask uh Pashant who's helping me here to go ahead and push a the this poll to live. Wanted to get an idea for for how familiar people are with agentic AI architectures. I think uh AI is something that you'd have to be living under a rock not to be familiar with AI in general uh at this point in today. And I think agentic AI being the the buzzword of the moment is something that uh people have uh increasing exposure to. But if you think about agentic AI architectures particularly as they pertain uh to enterprise use cases uh there's a bit of varying degrees of uh familiarity with that and so uh want to start that off so we can we can tailor some of our conversations. Uh, and and and it looks like uh just seeing the real-time results from the poll coming in, a pretty even distribution ranging from complete novices to to people fairly further along the development curve as we look at that.
All right. Uh, so uh, briefly uh, introduction to Lena AI. Seven, eight years in business. Uh, we actually uh predate the LLM era. So uh, we got our start in the the enterprise digital assistant uh uh chatbot era. So deterministic chatbots that would uh take employees through specific experiences to to assist them. Uh, so the great news is we've been in been helping employers with these use cases for for quite some time. Uh, and as the the LLM era has come upon us, that has provided an extremely large degree of uh uh of information and data sets uh that we've been able to use in order to tailor our experiences. And and that learnings from our seven, eight years in business have been invaluable towards helping some of our 500 customers worldwide uh really build experiences that can really assist their employees across the board. And and that uh has been a journey for us, but it's been a large amount of learning that has been really, really helpful in terms of putting together the architecture that I'll I'll walk you through here today and as you can you can see in some of the complexity uh and uh uh challenges that are associated with putting together an architecture like that.
All right. So so let's talk about uh what's agentic AI in a uh employer-employee relation uh context. And every time I think about AI agents, I I like to to draw back to the analogy of okay, what what's an agent from a from a person perspective? And uh I don't imagine there's too many music or sports stars among us, but uh for those of us who who have done a real estate transaction, you may have used a real estate agent. And and a real estate agent is someone who you have asked to go out into the marketplace and you've given them broad instructions. You've perhaps empowered them to make a certain category of decisions on your behalf and you've asked them to come back to you for a different set of decisions. And I think that analogy is really apt when we think about AI agents because the the the same set of instructions that you might give a real estate agent acting on your behalf uh in a uh person-to-person scenario is the same sort of instructions that you would give to an AI agent acting on your behalf. My AI agent may be empowered to do certain things autonomously, may come back to to me to confirm other things uh and may may not be equipped to handle other circumstances. So that's is is a relatively good analogy as we think about AI agents.
Uh, I think as we look at uh AI agents and the the enterprise concept, a number of uh both sort of trends and principles uh around it. So, so first of all, I think uh you you've seen and anybody who has uh played with a GPT model has been able to see the immediate power and promise of being able to answer knowledge queries and say synchronize the set of documents and answer knowledge queries based on that. Uh, today it it's it's still amazing technology. Uh, but I I I'd like to put to to view it as the easy part. So, so answering questions is easy, but completing tasks on your behalf is is is hard. And we tend to to view those as two sides of the same coin. The knowledge that your employees have access to will ground the actions that they can take advantage of. So, finding out what my PTO balance is in the time and attendance platform, applying for PTO, that's just as important as understanding what my employer's PTO policy is. And those two together are certainly more valuable than than any one of them offered alone.
Uh, in order to enable those actions, agentic AI has to be connected to multiple different systems and the important systems of record. I I think as you look at employees, there's an ever-increasing number of systems that employees have access to uh and can take action uh on themselves in those external systems. But for a a proper employee assistance, agentic AI, it needs to be connected to those different systems of record. And those are systems of record across HR domains, IT domains, finance, sales, procurement. If if I'm an employee, depending on my job role, I could have access to literally hundreds of systems. Uh, and an agentic AI needs to be connected to a critical mass of those systems such that I can get uh benefit for me in in interacting and the AI agent interacting with those systems on my behalf.
So a a hallucination is a major challenge. So if you if you played with publicly available AI models and and and heck, I even see this with a from a Google search perspective. I was uh searching for a a particular feature on my iPhone actually inside of Google and I I asked Google, okay, well, how can you do this particular feature? And it came back to me with a step-by-step instruction which I tried to follow on my iPhone until I discovered it actually didn't exist. That so so Google's chat search results had hallucinated a very detailed process and procedure that physically didn't exist on my iPhone. And that was really, really uh enlightening even to me. And I and I live in this on a day-to-day basis. So those publicly available AI models tend to be text prediction engines. They predict text that sounds like it's true. It's compelling when you read it, but isn't necessarily. So uh building AI models that one don't hallucinate in the first place and then two have a pipeline associated with them that one will generate an answer via via one model and then perform a fact check on that answer via a second model and only if it if it follows that muster display it back to the user is really important to avoid hallucination, especially when you're talking about something that's important as an employer's policies or or documents that an employee might have access to.
Now the the other side of the coin beyond uh hallucinations is is permission. This think as soon as you you connect an agentic AI system to multiple different endpoints uh the the permission model matters. And me going out to an HRIS system, I have one set of permissions in an HRIS system and the CHRO might have another set of permissions in that that HRIS system. So, uh, I if you're connecting an agentic AI solution to external systems, it needs to follow the permission set inside of that external system and and represent you to that external system to understand what your user has access to. So, permissioning is is really key uh in addition to to avoiding hallucination.
Uh, when I do a lot of my demos and we'll talk about this from an architectural perspective. I I think it's really important for agentic AI to to meet employees where they are. So I I do a lot of my demos inside of enterprise chat applications. So so things like Microsoft Teams, inside of Slack, Google Chat, uh, and for a category of desk-based workers, that makes a ton of sense because as a desk-based worker, we we generally live inside of those tools. But not every worker is a desk-based worker. So uh our recent release and if you caught a webinar about two weeks ago, we had a release for for for voice interaction which has been hugely positively received by some of our prospects and customers because it meets those mobile employees where they are. Clearly mobile devices are part of that as well. So, so mobile applications, uh, text, SMS, uh, uh, the all those sorts of tools in order for interacting with mobile employees, really, really important. And, and that's not a single choice for a given customer. That is a, uh, a multiple choices in order to meet multiple customers wherever they might be.
All right. So, so let's talk about some of the challenges. And let's say you say, "All right, well, this is all great and we we want to we definitely want to move forward with this sort of thing." What are some of the challenges that you might face? Uh, I think anybody who has deployed any enterprise system these days has seen the proliferation of multiple different agents from those first-party vendors. You've seen the challenge of sort of operating across those platforms and uh it's of course impossible to to to read the press these days without seeing the latest and greatest AI advancements. So that presents a number of sort of fundamental challenges. So I think if you look at the enterprise landscape uh and and there was a study that Okta put out that said that the average employee has 231 applications uh and Okta would know since they tend to to authenticate those applications. So it's it's a hugely overwhelming concern if I'm an employee, where do I go to do this? And I think uh as something that we do from a pre-sales perspective because I talk to a lot of customers who who are coming on board to the Lena platform and even us understanding, okay, well, you might do time and attendance here, you might do pay and performance here, you might do core HRIS here, you might do HCM over here. Understanding where data lives and where an employee takes action is difficult for for me as a professional when I'm talking to people who specialize in this. If you're an average employee, understanding where you go in this landscape of uh of uh applications is difficult and that that's already a challenge.
Uh, now when you layer that out and say, okay, well, every one of those applications is now coming out with an AI assistant and saying, okay, well, I'm the the Workday AI assistant. I'm the ServiceNow AI assistant. And I'm the Salesforce AI assistant. Being able to help you inside of those applications, uh, that's helpful in theory. Uh, but now instead of 231 apps, you have 231 AI assistants. And, and that doesn't really solve the core challenge. In fact, it can make it more complex over time. Uh, and then if you take it to the next level and look at one of those AI assistants offering an AI studio product and try to integrate an AI assistant with other platforms and and again, this is theoretically solving a challenge. So an AI assistant being able to to to build an agent to connect across platform cross-platform is good. Uh, but uh every one of these vendors is going to tell you that okay, well their system needs to be at the front and center of the employee experience and rarely does it have from a capability perspective the ability to do that. And so that doesn't solve that fragmented challenge either.
Uh, now look, I'm a technology guy and uh I'll happily tell you about my home theater system or my home automation system or all those things that are kind of technology for technology's sake. But AI assistants can't be technology for technology's sake. They need to solve a real-world problem in a meaningful way. And when those uh both AI assistants offered by the individual enterprise uh service vendors and even the AI studio products of those individual vendors don't talk to each other, it doesn't save the fun solve the fundamental challenge of a uh of a fragmented ecosystem.
Uh, furthermore, if you go down the AI studio route, and we have a number of customers who who have gone down that AI studio route, and and I talk to a customer every day who says they've built something on Copilot Studio, you're essentially doing something that's outside of your core business. You're you're onboarding essentially a development team in order to develop a set of capabilities on an AI studio platform. Uh, they all typically start with the the very easy-to-use drag-and-drop low-code no-code workflow builders, but fundamentally to operate something like that at scale, you're looking at for uh DevOps engineers, data scientists, data engineers, an entire coding and integration pipeline. And it's a it's a heavy lift from a personnel perspective to operate a shop that's iterating on an AI studio product. Uh, the the fact is that and I talk to customers and prospects all the time, nearly every one of them without fail, like I said, has has gone down that road in some way. Some team in that organization has either tried out a Copilot Studio or a Joule Studio and said, "Hey, look, I want to try to build something." And and frankly, they probably saw a little bit of early success and said, "Hey, this is easy. This is cool. We're accomplishing things." I mean, that's the that's the promise of AI, and we're all, even me, I'm I've continue to be impressed when you see something work for the first time. You're like, "Hey, this is progress." But the reality is when the rubber hits the road and and you look at those uh those customers of ours who have attempted to bring those projects to the light of day, less than 5% of them have because you get into complexity challenges of okay, well, how do we grow across the ecosystem? How do we ensure that it's offering in an intelligent manner? How do we eliminate hallucination? All those challenges that I were speaking of become more uh transparent as you think about bringing a homegrown AI studio-driven product like that uh to fruition.
So the the the solution. So we we've spent as I've mentioned at the onset, a a a large amount of time working on this. Uh, and this is the the evolution of something of our collective eight years in business in order to put together an agentic AI architecture that we think is both scalable in nature, is secure in nature, can be cross-platform in nature, and can uh solve for many of those challenges that I that I outlined at the onset. So uh in a second I'm going to go ahead and and and take you through that architecture. I do want to remind everyone of the the chat and Q&A as I'm going through the architecture. At the end of that, we'll have a chance for questions, answers, and and some other overall engagement.
All right. So, with that being said, this is a relatively busy slide. In fact, I'll go say a very busy slide, but uh this is essentially what we view as the architectural solution uh to some of these challenges. And I got my little laser pointer up here. I'll try to break it down for everybody into individual components. So let's start at the top here from a touch points perspective. And so so you heard me mention early on that it's important to meet employees where they are. So so certainly Microsoft Teams and Slack and other chat applications, very, very important for desk-based workers. Uh, as I mentioned, voice is getting a ton of traction for us uh uh internally to the point that we have uh customers and prospects of ours who are looking at entire call center replacements and say, okay, you're operating a call center today, but if an agentic AI solution can can answer directly via voice, you don't need a call center tomorrow. Uh, certainly browsers, both desktop and mobile. Uh, we partner with a number of the major internet vendors and embed inside of those internet uh tools. Service portals as well. Uh, a, uh, API and sort of machine-based interaction is something that we're seeing more of. Uh, I think as as you look at protocols such as MCP and A2A, which are are coming out in the marketplace and seeing more adoption in the marketplace, we'll have uh, more cases where that's being used. We're technically capable today. I don't think that the rest of the ecosystem is there yet, but we'll see more of that in the future. And then, of course, legacy grounds of communications, email, text, WhatsApp, things along those lines.
So regardless of the touch point that a given employee uses to enter the the Lena ecosystem, they are met by our orchestrator. So the orchestrator is what uh the large language model that builds an intelligent execution plan for solving an employee's query. Now it's important to note we're not a core model company. So we don't build models from scratch. We will take best-of-breed models of commercially available and we will tune them based on our extensive data set. So we bring those to market as the Work LM family of models and we leverage best-of-breed commercial models, bring them using our synthetic data set that we've built over our eight years in business. We have a minority of our customers who will bring their own key to external models and we'll bring those, but the majority of our customers use one of our Work LM family of models to take advantage of that seven to eight years of uh training and learning that we put together. Uh, regardless of what's used from a model perspective. Uh, the first thing uh the the orchestrator will break down the query into individual tokens. So an AI large language model doesn't read from left to right like you or I would do. They look at an entire query holistically, breaking it down into subpieces or tokens. It then looks at those tokens and will build an actual intelligent execution plan in order to solve for the challenges represented by that prompt. Uh, now in Lena world, it will use agents that it has access to in order and skills that belong to those agents in order to build the intelligent execution plan and of course, it'll observe and adapt over time.
So let's talk about agents. So agents are are sort of core to the the the LENA offering and you can think of agents in two ways. You can think of system-level agents and and I think as you look at early efforts into to agentic offering, we were seeing system-level agents much more often. So this could be, I am a Workday agent. I can update your employee directory. I can apply for time off. I can see uh job change requests. I'm a ServiceNow agent. I can update and close tickets. I'm an Oracle agent. I can uh interact with procurement offerings and ERP offerings. Uh, so I think those system-level agents are important and and probably the majority of our customers are still on the the system-level agent uh perspective.
As we go a little bit more forward-looking, it's a useful thought process, both a thought process and a and a technical architecture process to to think of agents as functional agents. And and we're even evolving both our branding, the way that we talk about this, and the technical architecture around functional agents under the the the moniker of an AI colleague. Think about you ask a colleague in HR or a colleague in finance or in in procurement to help you and they might be doing work for you anyway, even if you don't ask them to help you. That's what a functional agent is and where we're seeing more and more of our deployments head towards this functional agent view and and we'll have some more architectural changes coming up to even support that further. So it could be, I am an IT support agent and I have access to reset passwords, install software, open tickets, things along those lines, and similar for these other examples. Think of it as an entry-level job role, person in that sort of uh uh experience.
So regardless of whether we're thinking about a system-level agent or a functional agent, the agent is essentially a sub-orchestrator. So it has a plain text description, not dissimilar to the descriptions I gave you when I was talking about a system and a functional agent. So that is the way that the orchestrator actually calls an agent. It looks at the plain text descriptions of those agents. Then it can makes a list of agents which may be in scope for a given execution plan. Now there are three parts of each agent which will then uh help build the execution plan. So the first is the agent operating protocol. So the agent operating protocol is essentially a set of instructions to your agent that will give information about your business processes. So uh this can be plain text in nature. Doesn't need to be something that you've built out via code or or or complexity, but can be a a plain text description of, think about a job change. And a job change might involve five, six, 10, 15 different discrete steps or descriptions, but you can write that business process in plain text. And an agent can understand that from an agent operating protocol perspective.
Now, the two pieces of action that an agent has access to are skills. So these are discrete bits of work that an agent can accomplish, as well as memory. So this is data that an agent has access to. So let's talk about data first. So first of all, we have a purpose-built set of agents called knowledge access agents. And these knowledge access agents actually go out and synchronize with sources of knowledge across your enterprise. So uh things like Google Drive, Dropbox, One Drive, Confluence, ServiceNow, Zendesk, other places where knowledge might live. So knowledge access agents synchronize with that and they use policy documents to respond to questions. They're relatively purpose-built agents. Agents can also access structured and unstructured data via memory. So if I need to query an opportunity inside of Salesforce or a ticket inside of ServiceNow or an employee inside of Workday, those are examples of structured data that an employee that an agent has access to via memory queries.
Workflows are actually how we encapsulate skills. So, think of a skill as the smallest discrete bit of work that an agent can accomplish. So, uh thinking of about not a a job change that's too big. There's too many things involved with a job change, uh, but a the components of a job change. So thinking about things like uh uh looking up an employee's cost center, looking up a manager's cost center, uh looking up an employee's title, changing a title, each of those discrete bits of work uh or a skill that's available to an agent, and they're tied together via the operating protocol. Skills are enumerated via the workflow studio. Now, our workflow studio differs in in two ways from other studio products. First of all, we're implementing smaller bits of work. The skills are are are tiny, and second, we come with a entire set of pre-built templates. So the the templates are out of the box. We have out-of-the-box integrations, out-of-the-box connectors to all these systems, and out-of-the-box templates which can be used to enumerate those skills.
Last things uh worth calling out: highest degree of safety, security, and trust, ISO 27001, SOC 2 Type 2, and a robust set of observability and governance toolset. So, like I said, this is a relatively complex slide, but this is something that we've been working on for a really long time and and frankly, we're quite proud of in terms of uh being able to offer to our customers the ability to take out-of-the-box integrations, out-of-the-box templates in order to uh uh resolve these sorts of problems.
All right. Uh, so, I think I've got time for just maybe two questions uh while we're wrapping up here. So, I had a question. Are the functional agents pre-built at Lena or are those separate integrations needed? So, uh, the answer is both. So, we had, uh, functional agents with for all of those systems and, uh, persona levels that I described, which are essentially pre-built out of the box. We have pre-built connectors, pre-built templates for all those skills. They can, of course, be customized. Uh, and that's a big value proposition that we offer. They're essentially out of the box, really, really short time to value while still giving you the ability to build custom agents.
Uh, how much of the AOP or agents are pre-built and how much is needed to be done by every company? So, a little bit of the a similar sort of question, but I I'll answer for the AOP, which is a different sort. AOPs do tend to be company-specific because the AOPs are your business processes. So how you would like your agent to interact with your employees is going to be more likely to be specific. But remember that's plain text. If I give you a Workday agent or an HR ops agent or an ITSM agent, I'll give you the raw connectivity to that source system, the templates of skills of work that that agent can accomplish, and you can tell me what are your business processes. Oh, when you open up this sort of ticket, it needs to be escalated directly to level two. Those sorts of things that are going to be unique to your your organization.
All right. So we're about two minutes up to the or actually one minute up to to the end here. I think that's uh all the time we have for for questions. Uh, do thank you everyone for joining. Uh, if you answered the the the the poll is yes, we'll reach out to you for more information. You can also scan this QR code. Uh, put any other questions in the chat. Uh, we really do appreciate you taking uh 25 minutes, 30 minutes of your time here today to join with us. Uh, this is obviously a really complex topic. I could speak for for two hours on this. Uh, so uh, look forward to uh continuing the conversations and uh uh being in touch with you soon. Thank you.