Since the arrival of ChatGPT, the evolution of GenAI has accelerated rapidly. For companies, it is challenging to keep up with the fast developments. Agentic AI heralds the next revolution that organizations cannot afford to miss. There is even talk of a genuine digital workforce operating alongside your human staff. But will it really progress that quickly? And where do we draw the line regarding the autonomy we grant AI agents?
To answer these questions, we must first examine what Enterprise AI agents actually do. Essentially, the foundation consists of the same enterprise data and AI assets we’ve been using for years. However, agents go a step further than chatbot applications developed with Large Language Models (LLMs), which quickly hit their limits. An LLM does not have access to tools and lacks connection with data, analytical reports, or AI models.
For example, if we ask an LLM why a particular customer wants to leave the company, the model will not be able to provide an answer. Enterprise AI agents can use information about the customer in their communication. An AI agent is essentially a system that connects and utilizes various intelligent modules to assist you as effectively as possible.
Key Components
Perception
First and foremost, an agent must understand what the situation entails. Everything begins with gathering information from various sources. This brings us back to the foundation that is crucial for any form of AI: the data. Without quality data, no AI system can deliver good output. An AI agent is no exception.
Interpretation
In the next phase, the AI agent must give meaning to the question. The agent must interpret the question and determine what information is needed to arrive at a correct answer. Here, traditional LLMs hit their limits: they often understand the question but lack domain-specific knowledge or context to formulate an accurate answer. For this, enterprise assets such as company data, knowledge bases, and existing AI models are essential.
Decision
The third component is the decision-making process. Based on the insights from the previous step, the enterprise AI agent makes a decision about the best strategy or action.
Action
What truly distinguishes agents is the fourth component: turning the decision into action. An agent does this more autonomously than we are used to and requires limited human guidance.
Fifth Component: Governance
To ensure the reliability of AI agents, we add a fifth layer: governance. Governance is important if we want to increase the autonomy of agents. Enterprise AI agents are primarily productivity accelerators for employees. In practice, people still provide the glue between what the AI agent produces and what is communicated or decided. By collecting data on performance, you can make an agent better and more accurate. Agentic AI requires a growth process where you gradually allow agents to work more independently. First, clearly define the autonomy of an agent and then build it up gradually.
Fraud Detection
From the perspective of productivity enhancement and as an assistant to humans, AI agents have actually existed for quite some time. What was missing until recently was integration with Large Language Models (LLMs). In the financial world, banks have been using systems for fraud detection for decades—a task that is virtually impossible manually and lends itself well to automation with AI, and thus with AI agents.
An AI model must collect the necessary data, such as information about the location, time, and amount of a transaction. Additionally, the system has access to a customer’s transaction history. With this data, AI searches for patterns and discrepancies so that the system can decide whether or not to block a credit card.
Twenty to thirty years ago, we called this a data mining solution, ten years ago we spoke of machine learning, and five years ago of AI. Now we call it an AI agent that, like a human colleague, screens whether your credit card transactions are trustworthy. For fraud detection, an LLM is not necessary. However, a bank can benefit from integration with such an LLM in certain applications. For the National Bank of Greece, for example, it is useful for a model to translate into Greek, allowing agents to communicate and analyze in the customer’s native language.
How Do We Make an AI Agent Reliable?
In essence, Enterprise AI agents do not differ much from the models we have been using for years. But especially as these systems gain more autonomy and are enriched with LLM applications, we must pay sufficient attention to governance. Therefore, here are four good rules for setting up an agent:
- Determine how autonomous your agent should be. Not all agents need full autonomy. Therefore, it is important to define the capabilities and manage the agent well. Always ask yourself whether it is necessary to connect an agent to an LLM. Ultimately, these tools consume a lot of energy and cost money, so it’s best to use them sparingly. In principle, an LLM is only necessary if there is human communication in natural language involved.
- Build guardrails that guide the behavior of AI agents. Just like human colleagues “, AI also needs certain rules within which the system can operate. Governance is needed to define the boundaries of that environment. What is an AI agent allowed to do? And more importantly, what is it not allowed to do? Also provide fallback mechanisms that detect undesirable behavior. With LLM” s, the possible answers are endless. A model may also seek answers outside the permitted environment. A fallback mechanism will slow down the model and inform it that it should not answer a particular question.
- Ensure that the reasoning of an AI agent is traceable. If you know how an answer came about, you can continue to train and improve the agent. This way, you gradually increase the autonomy of the tool and determine which decisions you can leave to the AI agent and when a human colleague should take over.
- Opt for a low/no-code approach when building agents. If you don’t need coding skills, it’s easier to maintain the AI agent. And if you can visually see how a decision was made, you can immediately improve the tool.
Since the rise of GenAI, many organizations have started to view AI in a more mature way. Experiments around AI and GenAI are growing, and organizations are beginning to see the value of agentic AI as a productivity accelerator for employees. At the same time, it is important in the evolution of AI and AI agents to carefully weigh the benefits and risks against each other. If we do this well, Enterprise AI agents will indeed generate a significant impact on business.
This is a guest contribution by Véronique Van Vlasselaer, Analytics & AI Lead, South West & East Europe at SAS. Click here for more information about the company’s solutions.