Beyond Automation: How AI Agents Are Revolutionizing the Enterprise

Emily Faracca: When most people think of AI in the workplace, they think of technology that reacts to their inputs, usually with information and analysis. What if the most profound change coming to the workplace is a new kind of AI that does more than just inform and analyze? Picture AI that doesn't just respond, but anticipates, simplifies, and acts to optimize your workflow. That's the promise of agentic AI. AI is not just about automating manual tasks anymore. It's about creating a truly personalized and efficient work experience, where AI agents can anticipate your needs and acts on your behalf.

Welcome back to the Future of Work, the podcast series where we discuss the trends and insights shaping the modern workplace. Today, we're diving deep into the world of agentic AI with Kathy Pham, VP of Artificial Intelligence at Workday.

Kathy discusses the power and potential of AI agents, exploring their history, capabilities, and real-world applications. You’ll learn how these intelligent agents are revolutionizing user experiences, automating complex tasks, and driving personalized solutions across industries. Let's hear Kathy's expert insights on unlocking the potential of agentic AI.

Kathy Pham: Hi, I'm Kathy Pham. I'm the Vice President of Artificial Intelligence here at Workday. I am excited to talk about agentic AI, commonly referred to by its core component, AI agents, and how it's transforming enterprise, especially when it comes to experiences, because experience is everything. This technology allows us to move beyond simply analyzing data and making predictions to executing tasks autonomously, combining agents and people to fulfill complex processes. It enables enterprise technology that can anticipate our needs and proactively completes tasks. For the users of our systems, this means delivering a simple, personalized experience while, behind the scenes, breaking down complex processes, combining that with individual contexts, and then coordinating tasks together to solve complicated business problems.

It also means bringing a level of personalization we've never seen before, taking into account levels of expertise, and current individual business needs. AI agents are systems that perceive details of the current surrounding environment, process and reason the next steps, and here's the key, execute on those steps, and sometimes learn from the experience, storing that in memory. My study of AI machine learning first began 20 years ago when we researched early AI agents. Back then, we studied Braitenberg vehicles, originally introduced in the '80s, which are simple robots with basic sensors and motors that demonstrate how complex behaviors can emerge from simple connections. The sensors are used to perceive or see the environment, reason where to go next, and the motors are used to execute the movement. By connecting these sensors and motors, we can program behaviors that mimicked animal instincts, like aggression, love, or fear. It was a fascinating early introduction to the world of AI agents. AI and machine learning capabilities have come a long way in 20 years, but I often come back to how we can find powerful solutions by leaning on basic principles of understanding our environment, processing the environment with machine learning capabilities, and making complex decisions based on those learnings, and then delivering a simple experience on the other side. With these new capabilities, there's a potential for automated processes, better decision-making, and personalized experiences that bring in the context of an individual or team's priorities and needs.

Before we dive into how to use AI agents today, let's take a look back at how it all began. AI has a rich history that dates back to the '50s when the term, "Artificial intelligence," was first coined, and some argue that it began even before that if we take in the prior centuries of understanding statistical analysis and predictions. Since historical context can help us better understand how we build our future, I'll focus here on the recent history that is most relevant to understanding how we can make use of agentic AI today. I'll take us through a progression of early automation with simple rules to prediction capabilities, and now advancements that can connect systems together, helping us solve complex problems. Let's start with the 1950s, when the early foundation for AI began. This is when the concept of the Turing test was formed, testing if machines can simulate human intelligence. Those AI systems were rule-based, so they would take predetermined, pre-written rules and give results based on the request. Then in the '60s and '70s, the early research for neural networks opened the doors to systems that can learn from our behavior and details that we put into the systems. By design, neural networks were inspired by the human brain with nodes that are like neurons that connect with other neurons and firing off signals to learn over time. The '80s brought expert systems, which are large sets of rules that could mimic decision-making from people. They were useful and can seem like human behavior at times, but still rely on roles and could not adapt to complexity, as well as a system that can learn on its own, accounting for real-time information and taking action. The '90s saw the growth of machine learning beyond rule-based systems, building on neural network research before that, and the idea of intelligent agents, which we now refer to as AI agents, was explored in industry. And the concept of reinforcement learning allowed our systems to be optimized based on trial and error over time.

And then the 2000s saw an expansion in computational power and data and movement to cloud computing, which further expanded both data storage and processing capabilities for many organizations. In the last 15 years, we saw the growth in machine learning capabilities, data storage, and the interconnectedness of systems and devices, making it now possible, and perhaps, necessary for agents to connect these systems together to simplify our software and technology processes. And with the capabilities of large language models and natural language processing, agents can process and respond to new requests from people in real-time. This is transformative for the experience of how people use technology. This moment is also different from what we've seen in the past with technology, where AI just did one task. Let's take weather, for example. Machine learning alone can help us train on past weather patterns and predict what the forecast will look like in the next few days. We can type in or ask, "What's the weather?" And it can return the results without knowing the context of who is asking or why. It could be a meteorologist prepping to give the forecast, or a parent trying to decide how to dress their kids for school. Now let's look at the same scenario from the lens of an agent. A weather agent can perceive the current environment, including the people interacting with it, like the meteorologist or the parent, and why they are asking, such as to tell a news story versus how to dress a kid. Agents then reason the next steps based on the environment and execute on those steps to provide tailored responses, like generating rain graphics for the meteorologist or suggesting appropriate clothing for a family. The agent learns from these interactions to improve future responses.

Let's follow the path of traditional software development to agentic AI with another example. Traditional software development generally follows a set of instructions. A programmer gives the system instructions, and it executes. Then early AI systems were rule-based, so they can sometimes mimic human behavior with enough sets of rules, but ultimately, they couldn't handle complex scenarios or new unknowns. With the rules-based system, for our weather example, we would put in inputs like, "Temperature is greater than 95 degrees Fahrenheit or 35 degrees Celsius," and, "Humidity is greater than 80%," then we'd predict a storm. If wind is greater than 40 miles per hour for longer than one hour, then issue a high wind warning based on data from the United States National Weather Service. Given enough complex rules, it can seem like a person is on the other side, analyzing. Ultimately, with more complexities, though, like unpredictable storms and other data points that lead to changing weather patterns, we'd need more advanced tools to predict and gain insights.

Advancements like machine learning, deep learning, neural networks, and especially natural language processing, led to the ability to take those historical weather patterns, combine with needs of different audiences around the world, pull in other current weather data, make predictions, and take focused, unique, personalized actions based on all of those factors. So, as you can see from the weather example, AI agents can learn from the preferences of individuals and continue to execute on their behalf. Agentic AI shifts the field from processing data to now also executing on both predefined tasks and learned tasks. This is meaningful for enterprise, where we have countless business processes and tasks that often require us to remember where to go, where to click, where to upload a receipt, where to find a list of candidates, or how to message a hiring manager about candidates. Agents can automate tasks that previously required manual clicks and expertise or know-how, streamlining workflows.

Let's now talk about the different types of agents in agentic AI. These agents perceive the details of the surrounding environment, process the information, reason the next steps, and take action to achieve the goals. Some agents can learn from the experience and store that in memory. I find that understanding agent capabilities can help us figure out what methods are best suited to solve our problems, so let's talk about a few. The first is reactive agents, which is the simplest type, built on rule-based AI. Reactive agents respond to changes in the environment based on preset rules. These agents don't have additional learning capabilities or store those learnings in memory. This might be a chatbot that pulls answers from a preset list. Next is model-based agents. These agents are built with models that process the surrounding environment and have a level of reasoning to make decisions based on understanding of the environment. This might be an agent in retail that balances the cost of holding onto inventory against the risk of a shortage, predicting order lead time, future demand, seasonal trends, reorder requests, and adjusting stock levels based on real-time sales data, making sure that in-store supply is most effectively stocked to minimize storage costs.

Now let's talk about goal-based agents. Rather than react to an environment, these agents have specific goals. This could be setting a goal to win a chess game, finding any way to get there, or creating an optimized weekly schedule that considers the latest weather events and upcoming deadlines. Then there's a utility-based agent, which looks at a utility or a domain-specific area and maximizes outcomes based on that utility. For example, in healthcare, the agent's utility might be to improve patient health, while minimizing cost and risk. It would then create treatment plans optimized for those factors, taking into account each step of the treatment. Or, in chess, if an agent's utility is to win the game, it also evaluates the position on the board after each move, from the worth of each piece to the safety of the king looking beyond winning. Next up are learning agents. These are agents that are built to change and adapt dynamically based on new data points and ongoing experiences. This might be a subsequent question in a chatbot, reactions to performance plans at work, or new data points from medical clinical trials for research. And finally, the increasing need for collaboration across teams and organizations has driven the development of collaborative or multi-agent systems. These systems connect and coordinate multiple AI agents to solve complex problems, such as optimizing transportation routes by incorporating real-time weather data forecasting.

Now let's explore how AI agents can be applied to solve enterprise problems. First, we apply human-centered design practices and user research to help us identify the right opportunities. This lays a foundation so we can responsibly design the system to solve the right problems for our people and our business. We also need to ask if we should even build these agentic AI systems at all. We assess risk of training our models on the data we have, the risk of connecting systems together, and how much automation is acceptable and responsible for our environment. This assessment needs to happen throughout the development cycle from early ideation to design to delivery of services. Next, we look at the technology fit. We consider scenarios where we have the data and the historical context, are seeking to find predictions based on that data, and want to automate actions based on those predictions. Now we can take a look at a few examples where agents may be used in industries.

In higher education, AI agents can personalize academic planning and advising for students. This includes understanding past class offerings, trends in past successful combinations of classes, individual student needs, and combining that with other events, activities, and factors across the university and taking action to help students. In healthcare, AI agents can assist providers with understanding diagnosis, treatment planning, and resource allocation for the proper amount of staff required for complex inpatient admissions. In retail, agents can provide multi-step customer service when a customer wants to return an item and take multiple-step processes to send the customer a shipping label. Agents can also run reports across multiple systems to manage inventory and staffing and help project hiring needs and generate job descriptions for those needs. I will also share a few of our examples here at Workday. We've implemented AI agents to streamline expenses, optimize succession planning, and transform recruiting. For example, our expense agent automatically itemizes receipts and creates expense reports, while our recruiting agent sources candidates, automates outreach, and recommends top talent.

Let's take a look at these in more detail. For expenses, once an individual checks out of a hotel, they can simply snap a photo of the receipt right then and there with their phone. An AI agent automatically extracts the relevant information from the receipt, creates a new expense line item, and adds it to the correct expense report. The agent can even send notifications to the employee and their manager, keeping everyone informed and ensuring timely reimbursement. For succession planning, and today's quickly changing workforce, having a strong pipeline of successors across career stages is crucial. An agent can analyze factors like current business needs, required team skills, predicted attrition rates, and suggest potential successors across the company. For us, built on 20 years of historical context, data, and processes, the agent can proactively identify high-potential employees and even generate personalized growth plans to help those employees prepare for future roles. For recruiting, finding the right talent is more crucial now than ever. A recruiting agent, which incorporates higher scored capabilities, goes beyond traditional methods by proactively sourcing passive candidates who may have expressed interest in the past. By understanding the current needs of the business and analyzing candidate profiles, the agent can automate outreach, recommend top candidates, and even schedule interviews. It seamlessly integrates with other workplace tools, allowing for efficient communication feedback throughout the hiring process. This streamlines recruiting, reducing time to fill, and improves the quality of hires.

To successfully deploy AI agents, we can lean on practices we've developed in software and product development. Here are some key strategies: discover and define the problem. What are we trying to solve? Beyond features and tasks like finding skills or analyzing an image, what is the problem? Perhaps it is finding the right candidate for a frontline role or automating certain parts of the expense reporting system, while ensuring compliance for an expense. It's important to clearly articulate the business challenge we want to solve. Understand user needs and context. Continuously get user feedback for pain points in the system and keep testing new updates. Make note about new moments of joy and challenges so we know where to focus. User experience is everything. Develop a data strategy. Ensure we collect the right information as needed, maintain security and privacy appropriately, and clean and label data sets to prepare for training. Find the right tools and models. Select the appropriate AI technologies for your specific needs. Sometimes this is the latest new large language model capabilities. Other times, it might be relying on rule-based systems and other prediction models paired with new design and experience capabilities to surface our results to our teams. And sometimes, no AI or machine learning is needed at all.

Prioritize privacy and security. Build in mechanism to maintain both privacy and security, including protecting sensitive information of individuals and the business, and set security protocols in place throughout the whole development process. Assess risk. Before we build and throughout the build process, continue to assess the risk of the models, the data that is used, the connected system, and if the problem should be solved with AI at all. Connect different systems. The power of agentic AI comes from connecting different systems to gain insights. For example, connecting data in our anonymous employee surveys with succession planning analysis can be helpful for the future of the company. Or connecting systems with partner technologies or external data sources, like weather patterns, might help with our supply chain planning. Be sure to brainstorm with expertise across your organizations to find those use cases. Test and monitor. Testing our systems is a critical part of AI development. Each step can yield new results that may or may not align with what our users and our customers need. Maybe connecting one dataset was not as helpful as we needed to gain insights. Or maybe the set of documents for our policies did not have enough details if we need to revisit that data set of documents. Testing will surface that early. In addition to testing, we need to be able to continue to monitor and maintain our systems, especially the resources that we use for training. And this brings us back to the first point of discover and define: "Never stop asking questions to understand what our users need and how they interact with our systems, and bring that insight back into our build and deployment process."

I hope this has helped unpack how AI agents are transforming how we work by enabling us to automate tasks, make better decisions, and create more personalized experiences. By understanding the different types of agents and their capabilities, we can unlock their full potential and transform our organizations. I also hope that this has sparked your curiosity and inspired you to explore the possibilities of AI agents. Happy learning and exploring, and thank you for spending time with me today.

Faracca: That's it for today's deep dive into the future of work. A huge thank you to Kathy Pham for those incredible insights. If you’d like to hear more from Kathy about how AI agents are transforming the enterprise, head over to our AI Masterclass series for a comprehensive outlook on the future of work.