I had a truly terrible customer experience with AI today. I called a company with what should have been a relatively straightforward request, but the AI couldn’t consistently understand what I was saying. It asked me to enter information using the keypad, then failed to recognize the DTMF tones it had specifically asked for. When I finally made it through that part of the process, I was given a single routing option that didn’t match what I needed, with no reasonable alternative path and no way to get the system to route me to a human.

This is exactly the kind of experience that makes people hate AI, and unfortunately, it’s still pretty common. What frustrates me is that none of what went wrong represented the limitations of today’s CX AI capabilities. The technology is capable of delivering a significantly better experience. What I experienced was a poorly designed implementation that either wasn’t adequately tested before deployment, wasn’t being monitored and improved afterward, or both.

Bad Implementation Is Creating an AI Perception Problem

Most customers don’t know, or care, whether the problem is the underlying AI model, speech recognition, telephony integration, workflow design, data, routing logic or a decision someone made during implementation. They simply know they called a company, encountered AI and couldn’t accomplish what they needed to accomplish. Their takeaway is that AI doesn’t work.

That matters because every poor experience makes customers more skeptical of the next one. Instead of approaching an AI interaction expecting it to make things faster or easier, many of us have already learned to start looking for the escape route. We press zero, ask for an agent or repeatedly say “representative” because previous experiences have taught us that getting through the automation may be harder than getting around it.

In other words, poorly implemented AI is training customers not to trust AI. That’s a significant problem for organizations making large investments in the technology, particularly because the customer’s frustration may have very little to do with the actual capabilities of the solution they selected.

The Technology Is Capable of Much More

Modern CX AI can understand natural language, maintain context throughout a conversation, authenticate customers, integrate with business systems, determine intent, complete transactions and intelligently escalate interactions when necessary. Importantly, it can also be designed to recognize when it isn’t succeeding and change course rather than forcing the customer to remain trapped in a workflow that isn’t working.

A good customer experience shouldn’t assume automation will successfully resolve every interaction. If the AI cannot understand someone after multiple attempts, the customer’s request doesn’t match an available workflow, authentication continues to fail or the system’s confidence in its understanding drops below an acceptable level, there should be another path. In many cases, that path should be a person.

Transferring an interaction to a human isn’t a failure of AI. Sometimes it is exactly what a well-designed AI system should do. The objective should be successful resolution and a better customer experience, not automation for the sake of achieving a higher containment percentage.

Don’t Automate a Bad Process

I think one of the mistakes organizations make is starting with the existing IVR or contact center workflow and asking how much of it can be automated. If the existing routing structure is confusing, the underlying data is incomplete or the business process requires customers to understand internal organizational complexity, putting conversational AI in front of that process doesn’t necessarily make it better. It can actually make the frustration worse because the customer is interacting with something that sounds intelligent but still can’t help them.

AI gives us an opportunity to rethink the experience instead of simply automating the old one. Start with what customers are actually trying to accomplish, the reasons they contact the company, how they describe those needs in their own words and where those interactions tend to break down today. From there, determine which requests can realistically be resolved through automation and design the experience around customer outcomes rather than internal departments, existing phone trees or legacy workflows.

That distinction matters. A customer generally doesn’t know which department owns their problem, and they shouldn’t need to. If AI is truly able to understand intent, we have an opportunity to stop making customers navigate our organizational structure just to get help.

Deployment Is the Beginning, Not the End

An AI customer experience should never be treated as a project that is designed, tested, deployed and considered complete. Real customers will phrase requests differently than the project team expected. Accents, background noise, mobile connections and different speech patterns will expose weaknesses. Integrations will occasionally fail, new intents will emerge and customers will take paths through the experience that nobody anticipated during implementation.

One of the biggest advantages of AI is that those interactions generate an enormous amount of information organizations can use to improve the experience. Where are customers abandoning conversations? Where does the AI repeatedly ask for clarification? Which intents consistently result in transfers? Where does authentication fail? What are customers saying immediately before asking for a human? Most importantly, which interactions never reach a successful resolution?

Those insights should create an ongoing improvement cycle. The organization can identify where the experience is failing, adjust the workflow or model, test the change, measure the result and continue iterating. Deploying AI without establishing that process leaves one of the most valuable characteristics of the technology unused.

Test the Customer Experience, Not Just the Technology

Technical testing is necessary, but a workflow can function exactly as designed and still create a terrible customer experience. Someone needs to experience the interaction the way an actual customer will. Call from a cell phone, use speakerphone, try different speech patterns, enter information using the keypad when prompted, give the system an answer it doesn’t expect, ask for something outside the defined intents, change direction halfway through the interaction and try to reach a human.

The ultimate test is whether the customer can accomplish the reason they contacted you. That also means organizations should be thoughtful about the metrics they use to define success. Containment may matter economically, but keeping a customer inside an automated interaction isn’t a win if the issue remains unresolved. A high automation rate paired with poor resolution is simply a more efficient way to frustrate customers.

This is where I think the AI conversation needs to mature. We should absolutely measure efficiency, cost per interaction, containment and automation rates, but those metrics need to sit alongside resolution, customer effort, repeat contacts, abandonment and customer satisfaction. Otherwise, it is possible to build an AI experience that looks successful on a dashboard while customers are actively trying to escape it.

We Should Expect Better From CX AI

I remain extremely optimistic about AI in customer experience, which is probably why experiences like the one I had today frustrate me so much. The capability exists to create experiences that are dramatically better than the traditional phone trees customers have tolerated for decades. AI can reduce repetitive questions, shorten wait times, provide support around the clock, personalize interactions and allow human agents to focus on situations where judgment, creativity and empathy matter most.

Organizations won’t get those outcomes simply by purchasing and deploying an AI solution. Technology selection matters, but so do implementation, integration, experience design, testing, measurement and continuous improvement. The companies that get this right will treat CX AI as an evolving customer experience that happens to be powered by technology, not as a technology project that ends when the system goes live.

At Advoda, that distinction is an important part of how we approach AI and CX strategy with clients. We’re not just looking at whether a platform has AI capabilities. We’re looking at how those capabilities fit into the broader customer journey, the contact center environment, underlying systems and data, escalation paths and the operational processes required to continuously improve the experience.

AI is capable of creating a significantly better customer experience. But if customers have to fight the AI just to do business with you, the problem may not be AI at all. It may be how you designed, deployed and managed it.