AI Agents: Promise versus Reality in Creating Business Process Efficiency
In today’s tech world, it seems every article is talking about AI agents. They’re being promoted as a breakthrough technology that can observe, plan, and act autonomously to achieve outcomes. One of their main benefits is making business processes more efficient. But before rushing to adopt AI agents, it’s important to understand what they are and if they can actually help your business run better.
Understanding Agents: From Traditional to AI-Powered
Let’s start with the basics. An agent, in its traditional sense, is someone who acts on behalf of another to achieve a specific goal. Think of sales agents who sell products for a company, or customer service agents who help solve customer problems.
This concept has evolved in the software world. A software agent is a program that automatically performs tasks for users or other programs to achieve a specific goal. Taking this a step further, an AI agent is a software agent with enhanced capabilities: it can analyze information, make decisions, and use available tools on its own. What makes AI agents special is their ability to think and remember past interactions to adjust their approach based on the goal.
Anthropic, in its paper “Building Effective Agents,” suggests AI agents “can be used for open-ended problems where it’s difficult or impossible to predict the required number of steps, and where you can’t hardcode a fixed path.” They emphasize the importance of trusting the agent’s decision-making process - a concept I’ll explore further below.
How AI Agents Work
Several frameworks support AI agents, with LangGraph being a notable example. LangGraph, a stateful, orchestration framework based on the ReAct pattern, separates thinking from doing to create more reliable decision-making.
A Large Language Model (LLM) serves as the brain of the AI agent in this framework. It understands and generates text, makes decisions, and guides the workflow by choosing what actions to take. The LLM can be enhanced with tools that allow it to interact with the world - like accessing proprietary company data, performing calculations, or completing specific tasks.
Key Observations from Pratical Experience
I’ve recently spent time studying AI agents and building my own using LangGraph. For those interested in learning more, I recommend taking the DeepLearning.AI course “AI Agents in LangGraph” to understand how to build these agents. Through my work, I’ve made several observations about AI agents:
-
Hallucinations: LLMs still hallucinate. Baldur Bjarnason’s post “The LLMentalist Effect: how chat-based Large Language Models replicate the mechanisms of a psychic’s con” emphasizes that hallucinations are a pervasive flaw in how LLMs work. An LLM’s output is constructed from its mathematical model of language and is only factual as a side effect of the distribution of those facts in the training data. This means human oversight remains essential for validating AI agent results.
-
Opaque Decision-Making Process: lack a full understanding of LLM decision-making. LangGraph states the LLM determines based on the context if a tool is necessary to complete the task at hand. Given this point, how do we know if the AI agent’s plan involved using the right tools to achieve the goal? What if the LLM decides it can perform a given action itself? This means human oversight of the actions an AI agent takes is also important for validating AI agent results.
-
LLM Capabilities: must understand what your LLM can reliably do. When automating a process, you need tools that enable the LLM to execute its plans. Some tools are straightforward, like accessing proprietary company data. Others, like performing calculations or creating new content, need consideration. Can the LLM perform the task itself? Understanding the steps of your business process and the right way to support them is crucial.
-
Error Handling: obust error handling is essential. I found that LLMs sometimes respond with questions or express uncertainty, even for simple tasks. This again highlights the importance of human oversight to make decisions in these situations.
-
Reasoning: I question whether LLMs can truly reason. Recent research suggests that their “reasoning” combines memorization with probability-based processing. If an LLM’s reasoning depends on its training data, can we trust it to create the right plans autonomously? This may point to the value of training LLMs on company-specific data and/or using Retrieval Augmented Generation (RAG) to improve results.
Making the Right Choice for Your Business
These observations raise an important question: Why build an AI agent if we can’t fully trust the LLM’s decisions or outputs? Moreover, AI agents don’t magically automate processes. Investments in tools need to be made outside the context of AI agents to achieve this objective.
Balancing Vision with Current Realities
While there are key challenges to creating fully autonomous AI agents, I am excited about the vision. Given the rapid pace of AI advancements, I expect capabilities will improve sooner rather than later to achieve it.
We also don’t need to view an AI agent as fully autonomous or bust. Andrew Ng suggests we should view AI agents as having varying levels of “agency” or autonomy which improves over time. This perspective allows for a more pragmatic appraoch to AI agent adoption.
Recommended Approach:
1. Future-Focused Learning and Investment
Take the time now to learn by making strategic investments in AI agents to determine what works for your business processes.
Start small: create an AI agent for a simple business process with a couple of clear decision points where you can prove the underlying LLM is able to define and execute accurate plans. As you develop these proof of concepts, assess the amount of human oversight needed and the true value the AI agent creates through its decision making.
In addition, identify the set of opportunities you have to create process effiencies using automation. These investments will need to be planned and delivered to enable future AI agents. These investments will bring value to your business now and influence the development of your AI agent roadmap.
2. Disciplined Problem-Sovling in the Present
As business leaders, we must remain disciplined in our approach to solving problems by defining the what before the how. A clear “what” helps ensure the solution (or “how”) aligns with the desired result.
When thinking about making processes more efficient, consider these key questions:
- Do I fully understand the business problem I am trying to solve?
- Is my process documented, and do I understand the root cause of the problem?
- Can I fix the root cause without technology? Simplifying steps could be enough.
- If my process is working well, would automation create more efficiency?
- What existing tools and proven techniques can I use to deliver the automation?
- As the business leader, can my team implement these tools themselves or do we need to plan the work with IT?
Conclusion
Remember that automating an inefficient process will only magnify existing inefficiencies. The goal is not immediate full autonomy, but strategic, incremental improvement that delivers tangible business value.
By maintaining a disciplined approach and viewing AI agents as evolving tools, you can navigate the current limitations while positioning your organization for future technological advancements.