AI agents are coming to a business near you, and you might have already chatted with one to ask a question, facilitate a return, or get help with any other number of issues. While AI agents are beginning to strike a more human tone , we know it’s not a human on the other side. So what exactly makes up an AI agent?
In customer service, sales, or any other domain, AI agents are not a single piece of technology but rather a collection of connected systems that are carefully weaved together.
“It’s important to consider each component individually and how they integrate with each other,” said Sinarwati Mohamad Suhaili, a computer networking researcher at the University Malaysia Sarawak.
In fact, dividing the system into constituent parts should be “the first step” in designing any sort of AI system .
It all starts with the AI model, of course. But from the knowledge source to the analytics layer and everything in between, here’s a look at the nuts and bolts of an AI agent and what’s most important to know about how to successfully integrate them into your business.
Knowledge provides the AI agent with the information it needs to respond to customer inquiries. According to Mohamad Suhaili, it's important the knowledge source is up-to-date and relevant to the users' needs to ensure user requests are responded to accurately.
Companies deploying AI agents can choose from a variety of knowledge sources to inform them. The most common is a hosted knowledge base that exists on a platform like Zendesk or Contentful , explained Gordon Gibson, director of applied machine learning at Ada, who agreed knowledge is “an essential component” of AI agents. However, the knowledge base could also be as simple as text documents. Ada has a knowledge base API that allows companies to tap content in various formats including text files, Word documents, and Google documents, he said.
But this doesn’t mean you can simply connect AI agents to piles of information and expect helpful results. The knowledge provided should be written in explicit language, have some repetition to enforce context, and be organized into logical subsections.
“At a high level, the way to think about this is that your AI agent is going to be using this material to deliver answers, so you want the content to be quite clear and not use jargon that would be hard to interpret,” Gibson said. “Each piece of knowledge should be a standalone piece of information that's not depending on deep inherent knowledge of a system, because your AI agent may not always be aware of all aspects of the knowledge base.”
Every component of an AI agent has a unique and important role to play, but it’s hard to overstate how directly knowledge impacts the quality of the AI agent and subsequent user experience. It’s key to both response relevance and problem resolution, which a 2022 systematic review of 40 studies on AI agents cited as “the most influential factors” for achieving positive customer satisfaction, increased probability for usage continuation, and product purchases.
Configurations are another vital component of AI agents, accounting for where and how customers interact. This includes dictating which languages are supported and the channels through which the AI agent is deployed, such as web chat, a custom app, voice, email, social media, or any combination of these options.
“The configurations such as supported channels and languages are important to consider as they determine the accessibility to the user,” said Mohamad Suhaili, adding that it’s important to ensure the AI agent is available on the channels and languages that users prefer.
When it comes to language support, for example, the importance goes beyond preference — it’s vital to whether or not a customer can interact with the AI agent at all. Connecting with more customers in their native languages is still one of the biggest challenges left to solve in customer service, and it can make a significant difference in their experience and customer loyalty.
“Meet the customer where they are” is a long-held saying in business, and from supported languages to channel deployment, configurations dictate where — and how successfully — companies can do so.
AI agents don’t only converse with customers, but also look up the status of orders, process information, and take other actions on customers’ behalf in order to resolve their issues. That’s where business integrations come in, allowing AI agents to directly interact with other functions of the business.
To set up these integrations for success, Mohamad Suhaili said companies need to identify the most common tasks customers need help with and ensure the AI agent can perform them accurately and efficiently. They also need to ensure it has access to the domain-specific knowledge it needs and consider how the AI agent integrates with other systems and services in order to provide a seamless experience for the user. Companies should also monitor these actions to identify areas for improvement and make updates as needed.
Sometimes referred to as rules, policies are directives put in place when designing an AI agent that dictate how it behaves — and ensure it does so in a way that’s appropriate and consistent with how the company wants it to function. This includes guidelines for handling sensitive information to ensure it doesn’t violate any regulations, Mohamad Suhaili said.
Overall, policies allow a company to specify hard logic for an AI agent to determine how it should respond to a customer inquiry, which knowledge or business integrations are applicable in a specific situation, and more, so it can act accordingly. Think of them as a set of if-then statements for the AI agent to follow.
“For example, if you have a business integration that's only relevant to customers residing in a certain region, you might want to have a strict rule that makes sure that integration is only available to the agent when it's speaking with the customers in those areas,” Gibson said.
Representation of the customer refers to information known about them from the company’s databases, the conversation itself, or their history of conversations with the company. It also allows the AI agent to set some of those configurations. This could include language, what region they’re in, what product they’re using, and any other user parameters that could be relevant to the AI agent both in figuring out which policies apply and responding to questions in the most helpful and accurate way.
Along with keeping the knowledge base up-to-date, the representation of the customer is also important to ensure personalized responses to the user's needs.
Personalization is another area where AI can bolster customer service . Imagine a customer who’s had to reach out multiple times about the same issue. It would be frustrating if the AI agent started each new interaction from scratch, whereas having a clear picture of the customer, the issue they’ve been experiencing, and their previous interactions with the company can enable more tailor-made support and a faster resolution.
On the topic of monitoring, an analytics layer is another crucial component of any AI agent, enabling the company to analyze how conversations are going and if the AI agent is working how they want it to.
“Analytics is about analyzing the conversations the AI agent has and determining which ones were able to resolve the inquiry without the need for a human agent, or human support, and other attributes like what the customer sentiment was,” said Gibson.
The analytics layer should provide both core metrics and intelligent reporting, making it possible to aggregate themes and see how the system is performing within different categories. Overall, the analytics layer provides vital insights, but even more importantly, informs what changes need to be made to improve the AI agent’s capabilities and the overall customer experience.
Guidance, also called feedback or instructions, refers to directives a company can give its AI agent to improve its performance and help it better resolve customer issues. It’s not all that different from the policies mentioned earlier — the difference is that policies are rule-based restrictions that can be hard-coded, while guidance is qualitative feedback provided in natural language.
Guidance is “basically, any kind of feedback you want to give the system to help it understand how to interpret the tools it has available,” Gibson said, saying this could be instructions around when to use a certain business integration in a conversation, how to format responses to a customer, or when to hand off to a human agent.
For the latter example around handing off to a human agent, for example, the company can give the AI agent guidance to detect a customer’s tone and pass off the conversation if they seem frustrated. A company might want to tell its AI agent to “try to be friendlier” or to “use simpler words." The insights discovered through analytics are crucial for informing what additional guidance an AI agent may need.
Overall, each of these components serves a specific and crucial role for making an AI agent function. Breaking down the AI agent into parts shows the importance each plays and how it maps directly to achieving the function of a successful AI agent, and more importantly, to the customer experience. AI agents are not a “set it and forget it” type of technology, however. Even after these components are in place, they’ll need to work together in concert and be continuously refined through analytics and guidance. AI agents may be powered by large language models, but all in all, they’re the sum of their parts.
Evolve your team, strategy, and tech stack for an AI-first future.
Get the toolkit
