Agentic AI in CX: How We Got Here and What You Need to Know About It

Here’s a refresher to take you step by step.

Evolution of AI Models: From Predictive ML to Agentic

Multivariate Testing

“Let’s test and learn”

Multivariate testing is a method for optimizing customer experiences by testing many variables at the same time. This goes beyond A/B testing, where you might be testing the open rates on a promotional email with different subject lines. If your customer engagement solution support s MVT, you can set up a test with an array of variables to see which combination produces the best outcomes—subject line, imagery, timing of the send, journey step inclusion/exclusion, number of communications over time, and so on. The idea is that you then uncover the “winning” combination and apply it to a broader set of customers to opti mi ze the email engagement.

With MVT, you’re setting up trials and hitting the “Go” button, then allowing MVT to tell you which strategy performs best. Some MVT engines, such as CSG Xponent, can adaptively learn and dynamically flow traffic to winning experiences. But what if you could accurately predict the winning combination without having to set up trials? That’s what machine learning (ML) does for CX optimization.

Machine Learning Models

“Analyze this for me”

Systems w ith ML can take a vast amount of data—even unstructured , like CRM data or traffic data or message data — analyze patterns in it, and make predictions. In the case of CX, these predictions are often about customers’ future behaviors, or the likelihood they’ll take a certain action under a certain set of conditions. ( When people talk about predictive AI, they’re often referring to ML. )

S uppose you’re trying to optimize an email campaign using ML. The system can analyze customers’ history , their preferences, and contextual signals to predict which subject line, image, and send time are most likely to get each individual customer to open or click through the message . It doesn’t settle on one “ winning ” combination purely based on historical trials. Rather, ML models can make inferences about the past and predict future steps and future success.

Now what if you didn’t have to create the output—that is, write the emails with the optimized content? That’s what GenAI can do.

GenAI & Large Language Models

“Create this for me”

GenAI is any AI that can create content—text, images, code, etc. It often uses LLMs , which draw from massive amounts of data to create that content in responses to inputs. While ML recogniz es patterns in customer data and mak es predictions based on them, LLMs go further by actually generating the content .

• ML can: P redict when a customer is likely to churn and inform a downstream decision on retention strategies .

• LLMs can: Compose a personalized email in real time based on the customer’s preferences and context.

In a lot of CX use cases, these LLMs interact with customers on behalf of the brand. Your business would obviously want to set parameters on how they should respond. LLMs require “grounding,” and that usually comes in the form of retrieval augmented generation (RAG) .

RAG is basically giving an LLM a knowledge base—your brand style, your product information, your policies, etc. Suppose an LLM is generating an email to a customer who suffered a major service outage. RAG has a pivotal influence on what content the LLM could produce:

• Without RAG: “We’re sorry you had a bad experience! Here’s a $40 credit to your account.” (Off-brand tone, offer not authorized under your service policies)

• With RAG: “We under stand you were affected by the recent service outage, and we apologize for the inconvenience . Based on your loyalty and subscription tier, we will apply a $ 2 0 credit to your next invoice as outlined in our service guarantee .” ( On-brand tone, policy-aligned .)

As you can tell, grounding will become even more important when it comes to agentic AI—where AI models would act with far more autonomy than GenAI.

Agentic AI

“Get absolutely everything done”

Agentic AI is a network of LLM-centric applications that work together to achieve a goal. That goal will often require the LLMs to complete a series of tasks and make decisions on how to complete them. Agentic AI does this by processing and responding to not just individual prompts (like GenAI), but a whole set of structured inputs that are configured to the tasks.

In CX, agentic AI uses a robust set of tools to not only personalize customer interactions but also handle backend operations — updating systems, logging tickets, leaving notes in CRM systems and even creating executive summaries. Agentic AI can do the job end-to-end, moving beyond human-assisted generation to true autonomous execution.

Related Post: From AI to ROI: What It Means to Navigate the ‘Prove-It Era’ of Customer Engagement

Key Components of Agentic AI

The biggest functional difference between agentic AI and GenAI is the set of tools sitting between the input and output—some used by GenAI, some exclusive to agentic AI. These tools include: reasoning, planning, tool calling, memory, mixture of experts and RAG.

Suppose someone wants agentic AI to design their backyard , and they have three small children.

Reasoning: This component of agentic AI analyzes situations and makes logic-based decisions. It runs like its own internal dialogue:

• What would someone want in their backyard if they have three children?

• A swing set? A treehouse?

• That depends on the ages of the children, etc.

Planning: The AI considers the multitude of factors involved in executing the task. It even maps out the steps to take to address those factors.

• What does the existing backyard look like? How much grass does this person want? What’s the climate? Etc.

Tool Calling: The agents access and use external tools at their disposal to get information or execute their tasks. In this way, they interact with the “real” world and not just internally.

• There is a tool that can estimate water usage for yards based on a set of factors. We’ll use this tool, input the data we know about the factors, and apply the tool’s output in our design.

Memory: The agents retain an enormous volume of data across interactions, which they continually call on for contextual information.

• Previous interactions suggested that the family might someday get a dog. Let’s avoid recommending vegetation that might be toxic to pets or grasses that might be fragile.

Mixture of Experts: Agentic AI models will often deploy a variety of agents designed to have specific areas of expertise. Instead of having one agent do everything, tasks are routed to specialized agents for quicker, more accurate output.

• Let's allocate most of these tasks to the agent that specializes in soil and climate analysis.

RAG: Agentic AI models can use RAG to ground their answers just like GenAI can.

• Let’s consult government watering restrictions or building codes that might limit how we structure the backyard.

You can see how agentic AI goes beyond generating responses to actually solve problems and execute tasks in the real world. These systems are also working with even more robust sources of information, as well as with each other, to increase the chances they’ll execute these tasks accurately and as intended.

T his example is a simple, consumer “personal assistant” use-case. You can only imagine the applicability to the enterprise environment. When the aver age contact center representative uses more than a half-dozen tools to get one moderate - difficulty service interaction done, imagine the efficiency that could come with agentic AI navigating it at the speed of GPU execution.

Examples of Agentic AI in CX

Most of what you hear (so far) about agentic AI is how it will streamline back-office operations—that is, automate a whole workflow for employees. But here are examples of how agentic AI could be deployed for customer-facing functions.

Let’s say you’re an insurance provider, and you want a more efficient system for offering customers payment plans to prevent them from losing coverage or going into collections. Agentic AI could initiate an interaction with certain customers who are behind on their payments, and the AI could create a payment plan tailored to their situation (subject to your business’s policies). When the customer agrees to the plan, the AI could then go into the billing system and take the same actions that a human agent would to implement the payment plan. It could then send out the necessary communications to the customer to confirm the plan and its terms.

In a loyalty-building example, agentic AI could identify customers at risk for churn, select the right retention offer for each customer, and offer it on the optimal channel for that customer. Where agentic AI goes further than GenAI, again, is that the automation doesn’t end with the interaction: the AI could actually make the “clicks” in the customer’s account to activate the offer, update the billing and send out the communications.

Related Post: AI in the Telecom Customer Experience: Five Insights From Consumers

Challenges and Considerations

A lot of businesses are ready to pull the trigger on agentic AI, with 65% of organizations reporting they’re piloting AI agents in some form. At the same time, the technology is at an early stage, and they have reasons to approach it with caution (especially in highly regulated industries like healthcare and financial services). Here are a few considerations businesses have to make in implementing agentic AI:

Risk Management & Trust: By definition, agentic AI is built to eliminate human intervention and oversight. The fact is, there are a lot of decisions that businesses don’t let their human teams make on their own today. Would businesses let technology make those decisions autonomously? Much of the hesitation that businesses (and their customers) have about AI agents is the breadth of functions they’d be entrusting to the technology. Wal-Mart surveyed consumers and found that 46% somewhat or very unlikely to use an AI agent to handle an entire shopping trip for them.

Brand Impact: With GenAI, businesses need to make sure the AI is giving customers accurate, on-brand information. With agentic AI, businesses would also need to ensure the AI is making the right decisions and taking the right actions in serving the customer. Just like with GenAI, any issues customers experience from agentic AI will affect the brand directly, not the technology provider. According to our research, 76% of telecom consumers said that a frustrating experience with a brand’s AI-powered customer service would negatively impact their loyalty to that brand.

Market Hype: It’s a tricky market out there for organizations looking for agentic AI solutions. Because the technology is still emerging, it’s harder to tell when vendors are overpromising what agentic AI can do today with reasonable maturity or risk management. Then there’s the murkiness about what agentic AI is; businesses need to ask vendors whether their offerings are truly agentic AI, or if they’re existing automation or ML tools the vendor has relabeled as such.

AI Isn’t a Trend. Agentic AI Won’t Be, Either.

AI is the new operating model for CX. It’s raised customers’ expectations for easy experiences. It’s also increased the demand within businesses to deliver these experiences more efficiently, even as journeys become increasingly complex.

At CSG, we've integrated AI deeply into our platform to deliver hyper-personalized predictive experiences—the kinds of experiences where every message feels handwritten, every journey anticipates what’s next, and every customer lands exactly where they need to go. We also help businesses automate mundane tasks so they can focus on the critical ones. Their teams write less copy, fix more journeys and trust AI agents to handle end-to-end tasks without constant oversight.

Agentic AI is about getting things done more efficiently and effectively. Sticking to that principle, we find practical yet innovative ways to apply the technology that don’t just check a box or ride the hype.