How AI agents will make smart decisions

The average adult makes up to 35,000 decisions every day. That includes everything from the shoes you wear to the things you say—and everything you do on the job. 

Of course, no one could possibly deal with weighing so many decisions, one by one. We can thank our subconscious mind for handling roughly 95% of those calls on its own. But that still leaves 1,500 decisions for your conscious brain. Every day.

AI agents are like a subconscious brain for your work life, handling dozens of low-level decisions so you can concentrate on the more important stuff.

For example, without being explicitly told what actions to perform, an AI agent can query multiple sources of information, ask other specialized AI bots to perform related tasks, and present the results to its human bosses. Agents can put these capabilities to work in roles for software developers, customer service agents, financial analysts, commodities traders, and many other functions—and, in many cases, do it faster and more accurately than humans. 

“Dynamic workflows powered by agentic AI are shifting the paradigm from process automation to process intelligence," says Murali Swaminathan, chief technology officer at Freshworks. "They will enable businesses to automate not only routine tasks but also decision-making—making workflows smarter, more efficient, and capable of delivering higher value.”

While nearly 9 in 10 organizations are either exploring the use of AI agents or have set up pilots, just 12% have deployed them in the business, according to a January survey by KPMG. Closing that gap over the next few years means understanding how AI agents make decisions, and setting them up for success, Swaminathan adds. 

Companies that can pull this off in key business functions like customer support and IT operations stand to make big gains in productivity—and even employee engagement and happiness.

Here’s a look at how AI agents make decisions, and what they need to live up to their hype. 

From process automation to process intelligence

Agentic AI excels at taking static, rule-based workflows and making them dynamic, says Swaminathan. Instead of following hard-coded algorithms, agents will understand the context of each scenario, analyze new data as it comes in, and infer the next-best actions to take. 

An AI agent capable of autonomous decision-making requires three components, adds Swaminathan. 

Brain

A reasoning core powered by LLMs that allows it to analyze context, solve problems, and devise strategies.

Memory

Agents need both short-term memory to keep track of ongoing tasks, and long-term knowledge for personalization and continuous improvement.

Actions

The ability to execute tasks via APIs and automation, and to adjust dynamically based on real-time feedback. 

For example, say you want to request some time off work. Normally you might have to go into one system of record to see how many vacation days you have left, jump to another to see if there are restrictions on personal time off based on your position or your location, and engage a third system to submit the request. 

AI agents can perform all those tasks and automatically schedule your time off for you. They can even determine if one of the days you've requested is already a company holiday, and make sure not to count that day against your yearly PTO balance.

Read also: Josh Bersin on why we’ll learn to love AI agents

In a customer support scenario, an AI agent can detect when a customer sounds frustrated, then bypass standard troubleshooting steps and escalate the issue directly to a senior agent. In a sales workflow, agents can analyze customer engagement data and identify high-value leads. A recruiting agent can look at hiring success rates and adjust candidate screening criteria accordingly. And so on.

While only a small percentage of organizations have agentic frameworks already in place, that number will climb rapidly—starting with relatively simple deployments, such as time-off approvals, questions around refund policies, and tracking order status, says Swaminathan. Eventually, enterprises will adopt a multi-agent framework integrated into mission-critical business systems.

When AI agents collaborate 

In a multi-agent framework, specialized AI agents work together to enable more sophisticated decision-making. To illustrate how this would work, financial firm Moody's created a Meta Quest virtual reality demo of an executive advisory meeting. 

In this virtual meeting, AI agents play the role of executives trying to decide whether to extend credit to a new supplier. The investment and ratings consultant notes that Supplier A almost defaulted on a key loan the previous year. The corporate researcher brings up news reports of political unrest in the supplier's home country. According to the ESG specialist, the supplier is falling behind on its goals for zero-carbon emissions. And the economist and the risk analyst disagree about whether alternate suppliers might be better choices.

Ultimately the group recommends not extending further credit to Supplier A, as that would represent too great a risk.

The goal was to simulate how collections of agents can work together to augment human decision-making, says Sergio Gago, managing director of AI and quantum computing at Moody’s. (The firm released its first gen AI-based product, Moody's Research Assistant, last December.)

"We wanted to see what would happen if we created a collection of agents, each with different perspectives and access to different datasets, and have them collaborate toward a specific goal," says Gago. "It's like having an army of assistants working together as a team to provide a solution, rather than just answering a question.” 

Not yet fully autonomous

But just as you can’t climb into your Tesla, curl up for a nap in the backseat, and wake up safe and sound in your driveway, you can’t just set agentic AI loose on your data and expect it to do all your work for you—yet.  But it can save you the hassle of navigating bumper-to-bumper data traffic, and it may present alternative routes home you hadn’t thought of.

“The vision of agentic AI is that you just give it a goal to achieve and it carries out all the actions on your behalf without any human intervention,” says Tom Taulli, author of the upcoming book, “Building Generative AI Agents: Using LangGraph, AutoGen, and CrewAI.” “But we’re not anywhere near that yet.”

Taulli says the closest we come to that today is AI-assisted code generation systems like Cursor and GitHub Copilot. In addition to writing code, these bots can also create all the files the code is dependent on. But even when you have high confidence in the results, you’ll still want to do a code review and a security scan before you put it into production, to make sure the AI hasn’t introduced bugs or vulnerabilities. 

“An LLM is essentially a probability mechanism, and sometimes it gets things wrong,” he adds. “There are certain decisions where you want guardrails in place and to bring humans into the loop.”

Drawing limits

As with many machine learning models, a key stumbling block to agentic AI is transparency. When a rule-based system makes a mistake, you can usually trace it back to a flaw in the code or a poorly designed algorithm. When agents are making decisions autonomously based on real-time data, debugging errors is much harder, says Swaminathan.

Companies will need the ability to hit rewind, figure out where the agent took a wrong turn, and reinitiate the process using new sources of information or more explicit instructions. Right now, the technology is not mature enough for this level of troubleshooting, he adds.

Finding the right mix of agents, data sources, and foundational models requires experimentation, adds Gago. Moody's runs agent simulations using different LLMs, then compares their output against a set of gold-standard documents. 

"We use a collection of extensive golden data sets, like real credit memos for example, that we use to benchmark and evaluate the accuracy and alignment of each of the outputs," he adds. "In the same way you run software testing, you need to run LLM and AI output testing."

Taming the wild AI agent

There are also high-stakes areas where you shouldn’t trust bots to make decisions on your behalf, like payroll, invoice fulfillment, or any other area of the business that’s vulnerable to financial fraud, adds Taulli.

The European Union's Artificial Intelligence Act identifies four levels of risk for AI systems, notes Art Kleiner, co-author of “The AI Dilemma: 7 Principles for Responsible Technology.” 

The act bans use cases that could cause harm to individuals, such as using AI for surveillance, assessing creditworthiness, limiting access to education or public services, and meting out criminal punishments. 

Though U.S. companies are not bound by EU rules, they should exercise extreme caution in how they deploy AI agents, he adds. 

"The old line from Spider-Man applies: 'With great power comes great responsibility,'" Kleiner says. "You need to ask, what should AI systems be allowed to do, and what shouldn't they do? You want to design your systems so that the benefits they bring aren't outweighed by the risk to your users or damage to your company's reputation."

Gago suggests enterprises start out by giving AI agents free rein to see what they're able to produce, then initiate the process of "taming" them by reducing their autonomy.

"You start by giving them absolute freedom, then begin evaluating their outputs and putting limits on what they're able to do," says Gago.

"Sometimes these agents come up with ideas you wouldn't otherwise think of. By analyzing how these conversations work, you can identify potential workflows that can be used by real humans."