Smiling female financial advisor with profit growth report on laptop at startup

How AI agents will take GenAI technology from answers to actions

Authors

7 minute read 15 Nov 2024

AI agents know when to start a series of actions and complete them. Here are three ways to prepare today as the technology matures.

In brief
  • Still a nascent approach, this form of GenAI does not await prompts or falter at executing a series of processes in the way robotic process automation has.
  • AI agents can also be developed with their own traits depending on the types of tasks and processes they need to execute.
  • Early tasks for organizations include defining authorization frameworks for AI agents and the methods in which they can retrieve data and communicate.

While many companies remain in the early stages of adding generative AI (GenAI) capabilities, they have likely been using AI in the form of robotic process automation for decades. These bots are skilled at handling individual rote administrative tasks but not at orchestrating a full process with multiple different activities. Now, GenAI flips the equation, putting these full-service capabilities in reach: when carefully engineered, nascent “AI agents” can complete all the tasks from top to bottom — and know exactly when to do them.

Today’s GenAI use cases are often divided into two groups:

  • Reactive chatbots, which help employees reduce their everyday toil when prompted.
  • Proactive copilots that work collaboratively with a human decision-maker.

How EY can help

But there is a unique third option beginning to emerge: agents (also known as “agentic AI”) that activate themselves to pursue a goal based on changes in their environment, work toward that goal, and recognize when the goal has been achieved and their efforts can cease. Agents do not wait for a prompt, as a chatbot would, or automatically offer assistance to humans as they go about their daily work, as copilots do. Instead, AI agents are entirely different in that they understand when a task is needed and fulfill it from one step to the next, perhaps in concert with other AI agents.

 

It may sound like this world of AI agents remains outside the realm of what’s possible for companies, many of which are already stretching their technology budgets with an uncertain ROI. It’s true that agents are in their toddler phase, targeted with hundreds of millions of dollars from investors searching for the next big development in GenAI. But in just five years, they may reach adulthood. Even today, EY groups are beginning to use agents (in well-defined contexts) in areas such as third-party risk management. Here’s what to know about agentic AI — and how to prepare.

1

Chapter 1

The next evolution of GenAI: what sets agents apart

Business decisions often boil down to how best to prioritize your goals, with one often coming at the expense of another. AI agents will bring about a world of “and” rather than “or.”

For many people, GenAI is a querying tool: You ask it a question or give it a prompt, and it produces text in response. But to develop a fit-for-purpose GenAI solution for an organization, it’s worthwhile to understand the tiers of broader capabilities it can provide and understand how they differ from agents (a term often defined broadly).


The first four are widely available today:

  • Advisory: Provides domain-specific responses and guidance to user prompts
  • Assistive: Helps users complete tasks to achieve their outcomes
  • Cooperative: Involves a collaboratively "back and forth" protocol, in which the AI receives prompts by monitoring user behavior
  • Augmentative: Extends a user’s cognitive capabilities and skills into a different domain

The remaining two levels are set apart by their autonomous nature, which is the essence of AI agents. Consider this scenario to understand the difference: you need to travel for work, and after fulfilling all your meetings, you’d like to return home as quickly as possible within the bounds of your company’s expense policy. You could query a chatbot before a trip about what flights are available at a given time (assistive), or a copilot could fetch options for you based on your online calendar (cooperative).

But an AI agent can monitor your calendar, changes in flight seat availability and dozens of other parameters to dynamically orchestrate your travel arrangements in the background based on your predetermined preferences and changes to your schedule in real time — setting it apart from mere automation.

“Agents will take us away from ‘do this or do that’ and instead move us into ‘do this and do that,’ introducing us to the era of and,” said Sinclair Schuller, EY Americas AI and Digital Platforms Leader. “An agent that does diligence on a deal could process documents in 20 minutes. It can take on deal after deal. AI doesn’t need to lead to doomsdays scenarios about humans doing less — it’s about and, and, and. As a function of human effort, agents can expand our abilities by 10x, while reducing the cost, improving the quality and making human work more meaningful.”

And just like human workers, AI agents can assume a wide spectrum of different characteristics:

Autonomy: Agents can make decisions and act independently with minimal human intervention or regularly seek human guidance or intervention.

  • Adaptability: A defining characteristic about AI is that it can learn from its experience and rebalance its decision-making and behavior based on context. But it can also follow predefined rules and change little (or not at all).
  • Collaboration: Agents can work with other agents, humans or both, or operate without requiring or desiring interaction with either.
  • Lifespan: It could be indefinite or last until a certain amount of time passes or a goal is finished.
  • Reentrancy: This is a programming concept concerning the ability to resume a task after being interrupted, either by external actors or events. Agents can also be designed to cease work on a task when interrupted once.

Each trait impacts the cost and complexity of an agent, as well as its ability to be controlled, so reflexively assuming high levels of each characteristic may not be unwise. For instance, if you want a cybersecurity AI agent, you’d likely want it to rigidly follow the rules with an indefinite lifespan, depending on the goal. Whereas an AI agent that acts as an M&A legal diligence agent would likely need to be collaborative with a low level of autonomy.

2

Chapter 2

Being prepared for the era of AI agents

While agentic AI technology becomes more mature, organizations should be evolving into “agent-friendly” businesses — to be ready to capitalize on the power of this disruption.

As is typical with a new technology, the build-or-buy debate is top of mind. Rather than building their own AI agents from scratch, most companies will likely rely on the dominant tech players and new AI startups for general-purpose platforms that help create agents for their own specific needs, such as a customer service agent that leverages proprietary account information.

While AI agents crawl, and then sprint, into adulthood, organizations should be assessing how to become “agent-friendly.” The foundation for an AI agent involves leveraging data and codifying business processes because the most powerful agents are not necessarily the most capable — instead, they have access to the best data.

And by “data,” it’s not just what happened yesterday — the rows and rows of historical transactions in a database. It’s “dark data” (which is captured but often unused) or tacit knowledge: the institutional knowledge and expertise of your people about how tasks are completed, what guides decision-making criteria, and what unique factors determine one action or another for better outcomes. Very few organizations have strategies around tacit knowledge, where their AI systems develop differentiation. A fully AI-ready data strategy is the collective consideration of all of these boxes in the chart below to help shape the business and AI agenda.


Additionally, executives should be thinking about the following topics to be “agent-friendly”:

1. Authorization: You may think of authorization as access to a system or a data set. But in the world of agentic AI, the question concerns whether autonomous agents are authorized to perform a task such as booking a flight or canceling a hotel accommodation, as in our example — just as a human worker often needs a keycard to get into the office. Without a robust security model in place, your agents are free to make decisions you don’t want, or they’re not able to deliver value at all in an overly restrictive model.

2. Comprehensible data: Most AI agents will run on a base large language model (LLM), and their corpus gains specificity and contextualization through other sources of data, accessed through retrieval augmented generation (RAG). For AI agents to complete their authorized tasks, a RAG architecture must fetch the necessary data — and that data must be structured in a familiar way or the agent won’t be able to put it to use. Today, organizations rely on data fabric or data mesh as an enabling layer between the RAG architecture and data sets, facilitating how data is connected to the AI agent and how that agent comprehends it.

3. An appropriate messaging protocol: As we’ve seen, AI agents need freedom and data to complete their tasks. A third component of their interactions relates to how they reach out of their computing resources — whether from a laptop, or in the cloud — and into other domains through networking infrastructure. For instance, your organization may want to hire a third-party contractor that brings in agents, but its network may not be set up in a way that effectively connects with yours. (And you can’t send an agent a laptop!) Or your agents may need to team up with other external agents. Messaging protocols may need to be revisited to enable the interplay of agents’ data requests and authorizations.

Summary

The advent of GenAI introduces the potential for AI agents to autonomously initiate and execute tasks based on environmental changes, aiming to achieve specific goals without constant human direction. Currently in their developmental phase, AI agents are expected to mature significantly within five years, with applications already emerging in fields like risk management. Today, companies can prepare by creating frameworks for agent authorizations, implementing data fabrics or meshes, and establishing messaging protocols.

About this article

Authors

Related topics
AI
Consulting
Emerging technology

Related articles

How to implement generative AI for 400,000 employees

EY leaders have been racing to effectively leverage GenAI across the organization. Here’s what we learned in rolling it out to 400,000 employees.

Dan Diasio + 1

    EY refers to the global organization, and may refer to one or more, of the member firms of Ernst & Young Global Limited, each of which is a separate legal entity. Ernst & Young Global Limited, a UK company limited by guarantee, does not provide services to clients.