Art Wittman | Content Director | September 19, 2024
If you’re someone who finds AI fascinating but nebulous, you’ll be intrigued by AI agents. Those large language models (LLMs) that companies have invested billions in? They’re getting real jobs as the brains behind AI agents: What if chatbots could understand your HR policies and have nuanced discussions with employees about them? What if a fraud detection system could act autonomously to shut down bad transactions as they’re happening? What if you could give an AI system a goal and it would autonomously do what it took to achieve it?
All these use cases are possible with AI agents.
You can even equip agents with tools—algorithms, sensory inputs, data sources, and even access to other agents—so they can perform complex tasks under their own steam. Think of a warehouse robot that navigates aisles to check inventory by combining information from a range of sensors, cameras, and scanners with its control software and an ERP inventory management system.
What’s being called “agentic AI” is coalescing as an exciting opportunity for all sorts of organizations by making AI easy to use and vastly more helpful.
AI, or artificial intelligence, refers to computing systems that are trained to simulate human intelligence. Most AI systems are programmed to learn, and some can improve their performance based on experiences and new data, solve problems using a wide range of inputs, and pursue goals and objectives in a methodical manner. In the most recent advancement, generative AI systems can make decisions and initiate actions independently to reach their goals. GenAI is used in applications as varied as self-driving cars, media recommendation engines, and tools such as DALL-E and Midjourney that create images based on textual prompts.
Enterprise AI refers to ongoing work to apply GenAI and related technologies to business workloads, with systems augmented with the organization’s data. Think customer service, personalized marketing, and HR and finance assistants.
AI agents are software entities that can be assigned tasks, examine their environments, take actions as prescribed by their roles, and adjust based on their experiences.
People give AI agents objectives based on the agent’s role and the organization’s needs. With its objective in hand, the agent may make plans, perform task, and pursue the goal based on its training, the application in which it’s embedded, and the wider environment in which it operates. Agents learn and iterate and may take on specific roles, connect with data sources, and make decisions on their own. Advanced agents have specialized jobs that may involve executing multistep processes that require judgment, communicate in a way that mimics human interactions, and often cooperate with other agents. The modular nature of agents enables complex workflows. The autonomy given to agents is determined by the humans who invoke them. Just as in hiring a new assistant, more autonomy may be given as proficiency is proven.
Agents work by combining natural language processing, machine learning capabilities, an ability to gather data by querying other tools and systems, and continuous learning to respond to questions and perform tasks. A good example is a customer service AI agent. When a customer inquiring about an order asks, “Where’s my stuff?” the agent forms its response by checking with the order processing system, querying the shipping carrier’s tracking system via an API, and gathering information on potential weather or other external factors that could delay delivery.
The term agentic AI refers to systems that actively pursue goals and objectives versus performing a simple task or responding to a query. Agentic systems can often initiate actions, such as a customer service AI proactively sending a query to a carrier to ask about shipping delays.
One way to make agents more useful is by incorporating retrieval-augmented generation, or RAG, a technique that lets large language models use external data sources specific to the organization or agent role. RAG allows agents to find and incorporate up-to-date and relevant information from external databases, enterprise systems such as an ERP, or documents into their responses, making them more informative, accurate, and relevant for the audience. For example, an IT support agent could consider past interactions with customers before it decides how best to address the issue at hand. It might include in its response links to helpful documentation or decide to open a ticket on the customer’s behalf if the issue needs to be escalated.
Key Takeaways
An AI agent is a software entity that can perceive its environment, take actions, and learn from its experiences. Think of it as a digital assistant or a robot that can perform tasks autonomously based on human direction. AI agents have distinguishing characteristics, notably the ability to set goals, gather information, and use logic to plan out steps to achieve their objectives. Because they’re underpinned by LLMs that provide the intelligence to understand the intent behind queries, AI agents aren’t dependent on keywords, scripts, or preconfigured semantics. Rather, they can draw on data retained from previous tasks, combined with chat-based prompts, to dynamically come up with solutions.
AI agents also learn by trial and error. Reinforcement learning is where an AI model refines its decision-making process based on positive, neutral, and negative responses. They mimic human ingenuity and can use tools, including cloud-based and enterprise applications and data sources, APIs, and other agents, to achieve their goals. They may also use additional AI- and machine learning-based systems to analyze complex data, natural language processing tools to process inputs, RAG to provide up-to-date and contextually appropriate content, and cloud services for the computational resources required to do their work.
AI agents work by combining techniques and technologies, such as those we just noted, to achieve their assigned goals. For example, a recommendation agent might use machine learning, tapping massive data sets to identify patterns; natural language processing to understand requests and communicate with users; and interfaces to enterprise tools, such as an ERP system, database, or Internet of Things sensors, or external data sources, including the internet, to gather information.
AI agents are planners. They can identify the tasks and steps needed to achieve the assigned goal. For our customer service agent, understanding where a given shipment is requires a series of actions. It would first access databases with information about the specific order, such as the shipment ID, delivery method, and date placed. Next, it would use that data to query the shipping carrier’s database using a web services interface to provide real-time tracking and an estimated delivery date. The agent could also look at where the shipment currently is and how long it has taken in the past to make the next leg of its journey. If it’s in an air freight terminal in Boston and a hurricane is moving up the East Coast, the agent might infer that a delay is likely and convey that information to the customer.
AI agents, like any AI technology, can deliver benefits commensurate with their training and the data they have to draw on. A feature that separates agents from their more static predecessors is that they can recognize when they don’t have enough data to make a high-quality decision and take action to get more or better data. The formulation of agents within applications is a highly applied version of AI. As such, organizations will find that for success with agents, AI gurus aren’t needed so much as those who understand business processes and, possibly, data quality experts. These specialists can help define agent objectives, set parameters, and assess whether business goals are met, calling in IT or the software vendor only if they believe the AI itself is malfunctioning.
Specific benefits cited by early adopters of AI agents include
AI agents can be challenging to develop and put into production primarily because they rely on complex models, powerful compute infrastructure, and vast amounts of data that must be curated and kept up-to-date. Moreover, IT talent oversight is required to confirm that agents can effectively interact with humans and adapt to unexpected situations, and business and data experts need to help with setup. Make sure you have expertise in natural language processing and machine learning and watch for these problems.
AI agents depend on a range of inputs to do their work, with the specific mix depending on agent type and use case. A customer support agent will converse with customers, consult their purchase and support histories, and access support libraries to answer questions. Some agents will interact only with other agents. A database query agent might create SQL queries to retrieve information requested by other agents. Agents functioning as virtual assistants measure success by how well they accomplish tasks, often based on human feedback. All require a unique mix of components.
Ideal AI agent use cases generally have related data and other systems, such as a CRM or ERP, that AI agents rely on. They’re also task-oriented: Think answering a customer question or driving a passenger from point A to point B. Look for jobs that take advantage of agents’ ability to improve their performance over time and to make decisions based on their understanding of their environments and assigned goals.
Current popular use cases include
As with any technology investment, you want AI agents to cost-effectively deliver the desired functionality, now and in the future. For agents embedded within applications, best practices are similar to those you’d use for a new employee, such as carefully monitoring early outputs and ramping up the complexity of work as the employee progresses with assigned tasks.
For organizations looking to create their own agents suited to their unique needs, the process is more involved. Consider these six requirements and recommendations to address them.
Your AI center of excellence should play a pivotal role in overseeing and managing the rollout of AI agents. Don’t have one? Here’s how to get one up and running now.
The steps to implement an agent are similar to any AI deployment. First, you’ll define the task: What do you want the agent to do, being as specific as possible with goals and objectives. Then, identify the functional process the agent will follow, the data it’ll need to access, the relevant business experts, and the tools and other agents it can access as part of its work.
It is often best to start by assigning a small beta test group, closely monitoring use and outcomes, tuning the agent based on results, and increasing autonomy based on proven success. Where applicable, you might model the process on provisioning a new employee. Let’s consider a demand forecasting agent coming online to help a retailer plan for the back-to-school season.
One note: You should have sufficient computational resources to run the AI agent—laggy performance will kill enthusiasm before the project gets off the ground.
These are just some of the AI agents currently available. Organizations should look at their pain points: What roles are you having trouble filling? What are some opportunities you’ve identified but lack the resources to test out your hypothesis? Is there a persistent employee or customer complaint that might be addressed by AI? Also, talk to your cloud and enterprise application providers to see what agents they’re baking into their products and services. Those roadmaps can spur ideas.
Examples of AI agents include
OCI Generative AI Agents combines the power of LLMs and RAG so that employees, partners, and customers can directly query diverse knowledge bases enriched with your enterprise data. Quickly create and embed custom AI agents in your enterprise business applications and processes.
The service provides up-to-date information through a natural language interface and the ability to act directly on it. Looking to try AI agent technology? OCI Generative AI RAG Agent—the first in a series of Oracle AI agents—is generally available.
Most of us have asked a chatbot a question and received a response that didn’t solve the problem. Ending that frustration is the ultimate goal of smart AI agents. Giving people contextually accurate and relevant information is good for them and for your organization.
What are the types of AI agents?
Types of AI agents include simple reflex, model-based reflex, goal-based, utility-based, and learning agents.
What are real-life examples of agents in AI?
Real-life early examples of AI agents are Alexa, Google Assistant, and Siri, virtual assistants that can perform tasks including setting alarms, sending messages, and searching for information. For businesses, Oracle Digital Assistant is a conversational AI platform that lets businesses create chatbots and virtual assistants for customer service and other applications—essentially an AI agent that helps companies create their own agents.