Customer support -- whether internal or external -- is a day-to-day imperative for enterprises.

But even today, it can be difficult to get it right. Simple prescribed chatbots are often on the front lines, but customers can tend to get caught in frustrating loops that eventually result in human hand-off.

The leveling up of this process — what some call the next frontier of artificial intelligence (AI) — is agentic software, or that which can automatically address (and ideally resolve) customer inquiries and interactions without the need for back-end human involvement.

These AI agents supported by large language models (LLMs) are considered “rational” and are designed to be completely autonomous, acting on data and iteratively learning from it — and their capabilities are increasingly expanding far beyond just customer service.

“We’re seeing the discourse in gen AI changing from ChatGPT-style interactions that are very human-driven, to these AI-driven interactions,” Mike Gozzo, chief product and technology officer for customer service AI agent Ada, told SDxCentral.

Understanding nuance in human conversation

Traditional chatbots, while impressive, are scripted, meaning that everything has been predetermined and explicitly programmed by developers with constrained rules. This ultimately limits their scope, as people can’t think through “all the different permutations that conversations can take,” said Gozzo.

No one can possibly think through a full conversation on any topic, he pointed out. “There’s too much chaos, too much entropy in natural conversation.”

Similarly, most businesses have a great deal of individual variation and extenuating circumstances that require human judgment, making it impossible to anticipate different scenarios.

Instead of a script or a step-by-step guide, then, an AI agent is given a role.

For instance, AgentGPT uses GPT-4 to automate repetitive or data-intensive processes; AutoGPT automates prompts by creating subtasks and loops; and ChartGPT transforms raw data into interactive charts and graphs. Other agents are specialized to fintech (retirement planning or investment advice) or ecommerce, while more generalist models are emerging — Hugging Face’s Jack of All Trades, for instance, or BabyAGI, which can handle a range of tasks including performing research and managing to-do-lists.

Autonomous agents are proactive and reactive, can spot patterns and have “freedom” and the capability to “make their own picks,” writes Matt Pogla of Auto-GPT. “Unlike standard AI, which follows strict rules, these agents decide their next move based on their surroundings and goals. This skill makes them stand out in the world of tech.”

Humans set the goals for AI agents, which are trained to independently choose the best course of action to achieve those goals.

“It works like the human brain,” said Gozzo. “Think of the system as being modular.”

That is, a frontier language model serves as the core reasoner (the “what’s the plan?” element), while secondary models linked to it will be tuned for more specialized tasks such as scanning articles for information, he explained.

When an AI agent initially fields a question, it first determines whether it has all the information it needs to answer it, Gozzo noted. If not, it will send follow-up questions and “clarify everything it needs to clarify.” The model knows all the tools in its toolbox, has all the sourcing documents and supplementation it needs, and is also given certain constraints.

For instance, an agent handling flight bookings needs to understand that situations involving deaths in the family need to be handled differently than an everyday cancellation.

“If you give an AI agent constraints, ‘you should be empathetic,’ the system knows that it has that tool available to it, it’s able to decide,” said Gozzo.

Companies also need to consider transparency. Some AI agents are trained to proactively disclose the fact that they are synthetic, while others don’t explicitly reveal themselves unless asked.

“If the agent is asked, it will always reveal it is an AI agent,” said Gozzo. “If interrogated, it’s going to respond correctly.”

Supplementing humans, making lives more convenient

Increasingly, agents are becoming even more specialized, serving essentially as personal assistants — and we’ll only continue to see this increase as the technology gets ever more sophisticated, experts predict.

For instance, said Conor Twomey, head of customer success for KX, you might receive an email from your boss asking you to travel to Washington D.C. for an event. The agent knows you like to fly Delta, what your preferred airport is and that you are a Hilton Honors Member. It also knows everything you need to get ready so that it can suggest an itinerary. There is no need for users to ask a question or provide a prompt because it will be “all automated into the day-to-day,” said Twomey.

These user-specific AI agents are intended to “make everyday lives more convenient,” he said. And, at least for now, they’re not intended for use for mission critical or life-or-death situations. This is because machines make mistakes, and it’s generally accepted that agents aren’t going to be perfect or get everything right.

“We know that just about anything it suggests needs to be independently verified — and that’s OK,” said Twomey.

Eventually, though, applications such as ChatGPT will be as quaint as the rotary phone of old.

“We’ll probably be laughing in a few years from now about how we used to login to ChatGPT and physically ask questions, because all of these capabilities will be naturally embedded into our systems and processes,” said Twomey.

Onboarding is important with AI agents (just as with humans)

There are many benefits to autonomous agents, experts point out — they can streamline operations and speed up time-to-market, improve customer service and free up human talent.

For instance, in the case of Ada’s customer service automation platform, Gozzo pointed out, one customer reaped enough savings to create an entirely new product line. Workers who had been doing lower-level support were trained up to handle more “in-depth relationship-style questions” for specific clients.

To ensure that AI agents are deployed successfully, enterprises need to have all the content pertinent to their business documented so that AI can access it. Architecture should be prepared with bidirectional flows so LLMs can retrieve information and clearly communicate. Importantly, enterprises should also enforce self-imposed guardrails and hire talent with the skills to work alongside AI tools.

Also, when deploying, approach it as if you're hiring a human, said Gozzo: AI needs to be onboarded, taught and provided structured feedback.

“Companies need to spend some time sitting next to their AI agent and providing feedback continuously to keep it fresh,” Gozzo advised.