Agentic AI in enterprise workflow automation

Agentic AI is rapidly enabling a new paradigm in automating business process workflows for enterprises. AI Agents, the program components that power an agentic AI system, complement the traditional rule-based processing and decision-making with cognitive approaches that rely on thinking, reasoning and planning.

Many organizations across the industry and market are looking at generative AI, and more specifically agentic AI, for a way to improve their automation and reduce their costs.

This article introduces the key features and architectural concepts of agentic AI such that you can explore agentic AI adoption in workflow automation. With a focus on the financial industry, this article uses a bank loan processing workflow as an example and lays out a phased approach to leveraging agentic AI, starting with incorporating AI agents in less risky processing steps and gradually expanding adoption to cover full end-to-end workflows.

Agentic AI core features

Three core features of agentic AI systems include:

• Use of LLMs for reasoning
• Varying degree of control
• Cognitive system architecture

Use of LLMs for reasoning

In an agentic AI system, the AI agents use large language models (LLMs) to reason, plan the steps, collaborate, and take actions. This is done using unstructured natural language instructions where the AI agents work cognitively – without any hard-coded rules or action sequences – to continuously understand the context and determine the path or actions to complete the request. Using an LLM, AI agents can also reflect and re-evaluate their plans to ensure optimal results.

Varying degree of control

The autonomy of an agentic AI system or an AI agent to take actions can vary from just planning and providing suggestions, to asking a human for approvals, or to fully automatic decision-making and actions. Although systems are usually tested in multiple ways as best practice, generative AI that drives the AI agents is prone to issues like hallucinations that can result in errors in automatic decision-making. Depending on the use case, there can be ethical, financial, and even life-threatening risks involved in a fully autonomous system. Complex enterprise workflows often have human-in-the-middle at key steps to mitigate such risks, while many smaller tasks and actions can be fully autonomous.

Cognitive system architecture

There are multiple architectural patterns that enable an agentic AI system.

At the fundamental level each AI agent operates with a specific purpose and within a specific scope. The purpose is its expertise, and the scope is defined by the tools it can use. Typically, the purpose is the role or persona of the AI agent, and the tools are any specific resources, such as specialized data or computation often available via internal or external APIs that help the AI agent to complete the request.

A simple architecture for an AI agent is often referred as ‘single agent with tools’. For enterprises, single AI agents with tools is a great way to start experimenting with agentic AI. Then, you can incrementally augment or replace specific steps of the enterprise workflow.

Once enterprises build individual single agents that are experts in specific domains, the next step is to make them work together. Referred to as a multi-agent architecture, the multiple individual AI agents collaborate to complete complex workflows end-to-end. The architecture has various patterns depending on how the agents collaborate; for example, a supervisor pattern where one master agent delegates and coordinates with other individual expert AI agents, or a networked pattern, where the expert AI agents coordinate directly with each other. Often for complex systems, a hierarchical or custom hybrid pattern is suited.

Loan processing workflow example with a Loan Risk AI agent

Let’s use a generalized bank loan processing workflow as an example to understand the concepts and implementation of a simple agentic AI system with a Loan Risk AI agent. This AI agent was developed using IBM watsonx on IBM Cloud.

There are multiple domains in bank loan processing.

In an agentic AI system, the different domains can be handled by specialized AI agents working together, each using their own tools, exchanging data, and collaborating when needed. For the scope of this article, let’s focus on just the risk evaluation domain. The following figure shows the conceptual architecture of the Loan Risk AI agent that uses AI to determine the risk and associated interest rate for a specific loan requests by clients. When requested, the Loan Risk AI agent interprets the question context, uses the LLM and information about the available tools to decide the logic and sequence for completing the request, and then acts on the information to complete the request. LLMs receive the questions and natural language prompt instructions from the AI agent, but there is no hardcoded traditional programming logic for interpretation or tool sequence to find the response.

To implement our example, the Loan Risk AI agent was developed following the ‘single AI agent with tools’ pattern and was programmed in TypeScript/Node.js using LangGraph. IBM Cloud Code Engine was used to host the user interface and code for the AI agent. We used watsonx.ai for the LLMs and retrieval augmented generation (RAG). We created embeddings for risk and interest rates documents (as sample for enterprise content) and stored it in vector database for RAG querying.

Refer to the GitHub repository for details.

Comparing traditional and agentic AI systems for workflow automation

While AI agents are still programs with input and output, the agentic AI approach contrasts with the traditional rule-based or deterministic programmed systems that work with structured requests, fixed input, and output parameters with all conditions and options hardcoded.

AI agents work cognitively. Using natural language prompt instructions, they can perform much wider and more flexible contextual interpretations, decide on the logic, and then take actions to drive complex end-to-end enterprise workflow. This enables flexibility in the type of requests and input parameters. If the context in the request is within the scope of the instructions and available tools, most likely it can be handled. The table below compares the traditional rule based vs. agentic context-based approach.

Loan Risk AI agent in action

Let’s put the Loan Risk AI to use. The following figures show the Loan Risk AI agent interacting with a human risk analyst who is trying to assess the risk and interest rate for customers.

In the following example, the analyst asks the AI agent about the risk for a customer Matt and how it was determined. The AI agent responds that the risk is “medium,” and it also explains how it was determined based on credit score, account status, and the published risk document. Note that risk was not determined by hardcoded rule, but rather the AI agent using generative AI RAG query that used the published risk tables in the bank’s documents.

The context and reasoning capabilities extend even further as the AI agent is able to interpret the policies in the bank document. When asked about interest rate for low-to-medium risk, a category that is not defined, the AI agent understood that the policy is to be conservative, and also that conservative means assuming higher risk and higher interest rate. Based on that, the AI agent presented the interest rate for the medium risk.

Summary

As shown in this financial industry example, AI agents can reason and take actions to complete requests. What sets them apart is their ability to ‘think’ and determine the logic instead of using hardcoded conditional logic as in traditional programs.

AI agents derive context from natural language by using large language models (LLMs). Given a certain level of autonomy and a set of available tools in their scope, one or more AI agents can drive a complex workflow with minimal to no human supervision, which is where we start to see value from AI agents.

As a first step towards agentic AI adoption, enterprises can review their complex workflows and identify the processing steps that are heavy on unstructured content, often require human intervention, and have multiple possible decision paths. Typically, these processing steps are a challenge to program and maintain and are great candidates for replacement by a cognitive ‘single AI agent with tools’ like the one described in our example.

A single AI agent can use its inherent reasoning, leverage its tools, and drive the step in the workflow based on natural language instructions with minimal programming effort. Over time, multiple processing steps can be enabled with AI agents, such as an AI agent to handle the loan application, another one to conduct document verification, and so on, thereby creating a multi-agent driven end-to-end workflow.