AI Agents: What They Are and How to Deliver Agentic AI Processes

As enterprises look to turn AI interest into business value, a design pattern is emerging known as AI agents. AI agents work autonomously in business processes to help organizations reach their goals. They help enterprises extract greater value from artificial intelligence. But what exactly are AI agents?

Artificial intelligence agents make decisions and perform tasks. They can handle activities such as analyzing data, routing work, updating systems, and completing process steps. An AI agent functions as an entire process in itself, using AI to understand the context of an input and then orchestrating which actions should occur based on that understanding. AI agents can act autonomously to perform tasks and meet goals without requiring human input or work in processes alongside humans to ensure human oversight, driving enterprise value by adapting to and acting within specific contexts.

The success of agentic AI hinges on two components: strong reference data to give context to AI decisions and structured processes that empower AI to take action.

So how do AI agents work? And what value do they provide to businesses?

An AI agent may be best understood by contrast. Consider popular generative AI tools based on large language models like ChatGPT or Google Gemini. A human has to first submit a prompt to receive an answer. The chatbot lacks agency, requiring a human operator to direct the AI. These chat-based services are good for contextual understanding of a conversation, content creation, or summarization, but lack any ability to take action.

Agentic AI combines powerful generative AI capabilities to not only understand context, but also plan a future outcome, and take action to achieve a goal.  

Let’s take a simple example in a customer service use case.

An email comes in from a customer stating: “Please cancel my order placed last Thursday.” 

The AI can gather a base level of context from this message. The email sender address provides context to verify and discover the customer’s information. The AI can understand the phrase “cancel my order” and derive a potential action. Finally, the AI can note that the order was from “last Thursday,” requiring a contextual calculation from when the email was received and calculating back to the date of last Thursday.

All this context understanding is well suited for modern generative AI, but taking action requires further information and a plan to execute the action. 

The agentic AI process must now take this base information, and dynamically determine how to retrieve the order information from across enterprise data sources. To make this more easily accessible, effective agentic AI requires a data fabric to empower AI services to easily retrieve data for more context that informs further actions.

After retrieving data from the data fabric, we gain additional facts that answer questions such as:

• Was there an order for this customer last Thursday?

• Does the email sender address validate to be an actual customer?

• What was the order number?

• Is the order able to be canceled or was it shipped already?

Given this greater context, the agentic AI process can now plan an appropriate response. By checking corporate policies and prescriptive rules, the AI service can initiate proper action. 

Rules to check may include:

• Customer email address is validated.

• Order is valid.

• Order is not shipped.

• Order amount is less than a dollar threshold for auto-cancelation via email.

• Order amount exceeds auto-cancelation threshold requiring human approval.

• Credit card information is still valid to issue a refund.

Depending on the conditions, the agentic AI process can take several actions, such as:

• Requesting more information from the customer.

• Responding to the customer that the order cannot be canceled.

• Escalating to a human to approve order cancellation.

• Canceling the order and issuing a refund.

• Canceling the order, but informing them that a credit card refund could not be processed.

As you can see, even this simple example can result in many possible outcomes. But this demonstrates that the key to realizing this autonomous vision for agentic AI requires mixing AI with enterprise data and business processes. 

Now that we have a definition and use case of AI agents, let’s describe the types that might be created. Understanding agent functions helps you apply AI agents to the right processes. Here’s a brief definition of each and how they apply in an enterprise environment:

• Goal-based agents achieve specific objectives by following a sequence of steps or rules. They’re ideal for environments with clear goals and straightforward decision-making. In customer service, a goal-based AI agent might route an email to a specific team by analyzing keywords, issue types, or current team workload.

• Utility-based agents make decisions based on “utility functions.” Utility functions quantify the value or desirability of an action or outcome. For businesses, this might mean optimizing time, cutting costs, or managing resources efficiently. A project management tool could use an agent to allocate team members to tasks based on workload balance and skill alignment.

• Model-based agents build an internal model of an enterprise system to make decisions and adapt to changes. This model could include an organization’s common workflows or resource constraints. Their adaptability makes them helpful for complex tasks. For instance, an ERP system might adjust inventory levels by simulating demand trends from sales data, seasonality, and supplier schedules.

• Reflex agents rely on simple, rule-based responses without considering environment models or historical data. This makes them fast but limits their applications in complex scenarios. For example, a simple reflex agent in insurance might flag expensive claims for fraud review. Reflex agents enable rapid responses, but they may miss nuance that model-based agents would catch. 

• Learning agents use machine learning to analyze data to improve performance over time. In financial services, a learning agent might be deployed to optimize trading strategies. They could start with indicators based on historical performance, then refine the approach based on profits and losses.

Each of these agent types brings distinct advantages, allowing businesses to choose the right agent for each task. But what makes AI agents so valuable?

When AI first gained mainstream attention, business leaders envisioned unprecedented productivity gains. And many saw those gains. Employees used AI chatbots for a wide range of tasks like generating content, analyzing information, and writing code. AI played the part of personal assistant, giving employees real-time answers to questions and improving productivity.

AI agents go beyond basic generative AI tools like chatbots. They offer a few key benefits that first-gen AI tools cannot do alone.

Because AI agents don’t always need direction from human users, organizations can embed them in processes. These AI agents save employee time, prevent errors, and reduce manual work. Plus, they can initiate intelligent actions that solve business problems fast.

For example, an AI agent can automatically generate an email response to a customer query for a customer service rep to review. Without the agent, the rep would have to draft the email manually or prompt an AI to create the draft. This manual work takes time. Sometimes, business is a game of inches, so these savings can really add up.

In the beginning, AI was a bit of a wild west scenario. Two employees could prompt an AI agent on the same task and get vastly different results. The quality of the output depends on the prompt fed to the large language model tool. A process platform lets you create a generative AI prompt or an AI model to run in the background. This ensures consistent outcomes by enabling you to deploy high-quality models and follow prompting best practices. 

Additionally, AI agents can be heavily specialized. This specialization gives you greater control and lets you optimize processes. If any smaller AI task in a multi-agent system goes awry, you can isolate that section of the process and fix it.

For example, an AI agent can take on tasks for a financial advisor. Let's say a client inquiry comes in via email. An AI agent could classify the email, extract information like PII, summarize the key points, and then draft an email response. The advisor could then review the response prior to sending. This saves the advisor significant time, freeing them for higher-value tasks. Plus, since you designed the prompts needed to do this ahead of time, you know they’re likely getting a quality outcome. If not, just find the link in the process causing the problem and fix it.

To maximize AI’s value, it must be placed in the context of a wider business process. A good process platform provides the tools to design, automate, and optimize your processes—and AI is a critical component. Using process intelligence tools, you can review metrics to decide where to deploy AI to improve processes. You can access multiple automation tools like robotic process automation (RPA) to handle procedural, straightforward tasks. Data fabric provides seamless access to all your enterprise data including those in external systems—giving greater context to your AI agents. This lets you build smarter AI agents and make more informed decisions. 

Most importantly, a platform orchestrates AI agents as part of a wider enterprise process. Enterprise processes are made up of complex, moving parts. Orchestrating data and tasks across systems, human employees, and digital workers is critical for efficiency and effectiveness. This ensures your AI agents (and your broader processes) have the most impact. 

AI agents bring a lot to the table, but they also carry some concerns. From runaway decision-making to security and compliance, you must understand the risks when delegating decisions and actions to AI. Here, we’ll break down the risks and offer practical tips to keep your AI agents secure, compliant, and aligned with business goals.

The very benefit of AI agents—their autonomy—also brings risk. One of the biggest fears surrounding AI is its ability to be autonomous and run away with decisions. This ranges from science fiction (hello, Skynet) to the realistic (such as mass insurance claims denials due to misinformation). 

AI agents can improve productivity, but adding a human in the loop for oversight reduces this risk. Employees can check the output of an AI agent for accuracy to correct errors. Failure to do this consistently can scale the problem and cause widespread damage. 

Consider a financial institution using AI agents to review loan applications. Without human intervention, AI agents may reject applications based on data related to protected classes like gender or ethnicity. Having employees review decisions allows the organization to catch and correct biased outcomes like this, ensuring ethical, responsible AI usage.

Like any other technology, AI can introduce security vulnerabilities. AI may have access to critical data and information that could be very damaging if breached. Plus, it adds a new attack vector in the software supply chain, increasing your attack surface. If using an adaptive AI agent that updates in real time, it can be fed misinformation to throw it off (this is called "data poisoning"). For example, an industrial supplier using AI to set prices could have prices manipulated via data poisoning. This could lead to them underpaying or charging too much. 

Make sure to implement guardrails like data access controls for your AI agents. Just as zero-trust policies ensure employees don’t maliciously or unintentionally leak data, good AI policies ensure agents only have access to the data they need for the job. A process platform gives you this level of control and more—the best platforms prioritize offering the tightest security controls.

Lately, there have been a slew of regulations around artificial intelligence. Lawmakers have grown increasingly concerned about the potential impacts of artificial intelligence. On top of that, AI agents can potentially run afoul of even non-AI related compliance regulations, such as HIPAA, if they inadvertently access or leak patient information without the proper guardrails. As a result, organizations must actively monitor and configure their AI agents to comply with both AI-specific and general regulatory standards. A process platform that enforces data privacy, retention, and processing rules ensures AI contributes to the business without creating risks.

When we talk about agentic AI, we must focus on the value of process. For complex, enterprise-level tasks, AI must inform itself, collaborate, and act independently. This requires giving AI access to data, humans, and system actions. A process platform provides the tools and structured environment for scalable AI integration. When you focus on process, you get a better version of agentic AI. As the Process Company, Appian empowers businesses to bring agentic AI into their business processes and unlock transformative value.