A July 2024 Capgemini report reveals 82% of companies plan to integrate AI agents in 1-3 years, expecting automation and enhanced efficiency. AI agents, capable of autonomous actions, are driving workflows across industries, from customer support to financial services. Leading enterprises are quickly adopting this technology for its transformative potential.
In a report titled “ Unlocking the Value of Generative AI, ” published in July 2024 by Capgemini, the majority (82%) of companies surveyed said they plan to integrate “AI agents” within the next 1-3 years, while only 7% said they had no plans to integrate agents. Seventy-one percent of respondents expect AI agents to drive automation, and 64% of companies expect AI agents to free up human workers from repetitive tasks so they can focus on value-added functions such as customer experience. The survey was conducted among 1,100 companies with over $1 billion in revenue.
Andrew Ng said “I think AI agent workflows are going to drive tremendous progress in AI this year, probably more so than the next generation of foundational models. This is a significant trend, and I urge everyone working in AI to pay attention to it.” So what exactly is it about AI agents that is so quickly being adopted by leading enterprise companies around the world?
OpenAI's Greg Brockman posted on X (formerly Twitter) in May of this year : "Users will increasingly interact with systems that are composed of a variety of multimodal models and tools that can act on their behalf, rather than a single model with only text input and output." This description seems to capture the concept of an AI agent well.
An agent is an AI system that can act autonomously, pursuing open and loosely defined goals. The agent's process typically follows these four steps:
Here’s an example you’ve probably heard before: a system that automatically books flights for an upcoming trip. The agent has to check emails and calendars to figure out when and where to travel. Then, it has to remember preferences (aisle/window, night/day flights, special meals, etc.) when booking flights, research and select the most suitable flights, and use the airline reservation system (web browser/API) with payment information to purchase the ticket. If it lacks preferences, payment information, etc., it has to ask the user again.
Now, let’s take a closer look at the technical concepts and future directions of AI agents.
It is difficult to say that the concept of AI agents originated from a single basic paper or specific research. It is a general and broad concept. It seems that the concept of agents has developed further as practitioners have become interested in AI systems that are becoming more sophisticated thanks to LLM.
One of the studies that created the concept of agents was a paper published in 2022 by the Brain team at Google Research that introduced the concept of "chain-of-thought prompting." This paper showed that LLM has the ability to solve the overall problem by breaking down a complex problem into smaller intermediate steps and performing each step in sequence. Although this paper was not researched for AI agents, the chain-of-thought technique greatly improved LLM's multi-step reasoning and planning ability. It is the core of agent behavior.
One of the essential elements of a competent agent is the ability to leverage external applications. This can be done by searching the Internet, sending emails, making online purchases, calling an Uber, building websites, updating databases, and many other tasks. In the realm of AI agents, this ability is called “tool use.”
A landmark study on the use of agent tools is Toolformer , published in 2023 by Meta researchers. The Toolformer team fine-tuned an LLM to learn how and when to make API calls to leverage external applications such as calculators, calendars, and language translators.
A more recent component of agent systems is the multi-agent architecture. The idea is that multiple AI agents working together can be much more powerful. Each agent is specialized and modular. When individual agents focus on a specific subtask, they perform that subtask better than a single monolithic agent could complete the entire project. From a human developer’s perspective, the multi-agent framework is also conceptually useful in that it decomposes a complex system into individual modules that can be improved and evaluated independently.
Financial institutions create credit notes to assess the risk of extending a borrower's loan or making additional loans. This process involves collecting, analyzing, and reviewing various forms of information related to the borrower, the type of loan, and other factors. Given the variety of credit risk scenarios and analyses required, it is time-consuming and requires managers to work with borrowers, stakeholders, and credit analysts to perform specialized analyses, and then submit them for review and additional expertise.
Potential Agent-Based Solution: Multiple agents share roles: one agent handles communication between the borrower and the financial institution; another agent edits and passes the necessary documents to the financial analysis agent; the financial analysis agent reviews the liabilities in the cash flow statement and calculates the relevant financial ratios; and the final review agent identifies discrepancies and errors and provides feedback. This detailed analysis, improvement, and review process is repeated until the final credit memo is completed.
Legacy software applications and systems in large enterprises often pose security risks and can slow down business innovation. But modernizing them is a resource-intensive process. Engineers must review and understand millions of lines of legacy code bases and manual documentation of business logic, then convert the logic to an updated code base and integrate with other systems.
Potential Agent-Based Solutions: Agents can be deployed as legacy software experts to analyze old code and document and translate various code segments. At the same time, quality assurance agents can critique this documentation and generate test cases to help the AI system iteratively improve its output and ensure accuracy. Components of the agent framework can be reused for other software migrations across the organization, significantly improving productivity and reducing the overall cost of software development.
LLM-based customer support cases are on the rise, and it seems like a viable market for AI agents.
Most customer requests (e.g., help with a forgotten password) are repetitive, and a standardized response manual is more than an agent can handle.
Fintech unicorn Klarna says that OpenAI’s GPT-based AI assistant was able to handle two-thirds of all customer service requests (2.3 million conversations in the first month alone), automating the work of 700 full-time human agents and generating an estimated $40 million in additional revenue for the company this year.
Potential agent-based solutions: AI agents can respond to customer inquiries in real time, integrate with internal systems, call the appropriate APIs to retrieve all the necessary customer information, and take the necessary actions to fulfill the customer request (e.g., updating a customer address or canceling an international data plan). The differentiating factor for customer support AI agents is the infrastructure, the orchestration built around LLM. Essentially, they are building a knowledge graph, where each node in the graph is an API call or an LLM call, etc.
Andrew Ng suggests that it is more useful to use the word agent in AI as an adjective (agentic) rather than a noun (agent). Rather than debating whether a particular system is an agent or not, let’s look at how agentic the system is. Once you have a simple agentic workflow, you can iteratively encourage people to make the system more sophisticated.
AI agents are certainly an early technology. It is true that introducing AI agents into an enterprise will require significant testing, training, and coaching before they can operate independently. However, multi-agent systems are moving towards collaborating with other agents and humans to iteratively improve the quality of their work, and given the pace of generative AI development, they could soon become as common as chatbots.
In the past year alone, Google, Microsoft, OpenAI, and others have invested in software libraries and frameworks that support agent functionality. LLM-based applications such as Microsoft Copilot, Amazon Q, and Google’s upcoming Project Astra are moving from knowledge-based to more action-based. Companies and research labs like Adept, crewAI, and Imbue are also developing agent-based models and multi-agent systems. That’s why business leaders should increasingly be interested in AI agents.
Explore how AI agents are revolutionizing industries with automation and efficiency. Don’t miss out—discover how your business can thrive with AI agents. Get ahead of the curve—start integrating AI today, contact us.