Frictionless AI CX Depends On Clear Costs And Invisible Interfaces

Generative AI was supposed to make customer service frictionless. Instead, teams get access to powerful new models and often feel pressured to adopt the technology immediately, starting with the tool rather than a business problem. That approach leaves companies with expensive pilots that fail to scale and automations that do not meaningfully reduce effort for customers or employees. All the while, they aren't delivering on their core value proposition: simple, useful interactions for their customers.

This focus on useful interactions is exactly what draws Miguel Guillen to study LLMs. As the Senior Director of Technology and Data at prescription cost optimizer Visory Health—and former architect at Microsoft, Amazon Web Services (AWS), and PepsiCo—Guillen has learned to vet technology strictly on utility. More specifically, he looks at large language models to see if they solve a specific problem, fit within existing architecture, and can be operated affordably.

Guillen applies the same discipline to his view of AI's current capabilities and, in doing so, finds that many enterprises are getting much too far ahead of their own problems and solutions. "We've started the house from the roof and forgot about the foundation," he says. "AI only works when you start with a clear problem and a real business case." Guillen points out that much of what passes for enterprise AI today amounts to better search and text summarization, which is useful, but hardly transformational. “We are still in the 2D era of AI. We haven't made it to 3D yet. The AI isn't smart enough yet to combine the text and visual worlds.”

• Insert coin to continue: Chasing this kind of frictionless customer experience often means that leaders overlook basic economics. Pricing for token usage can be opaque, making it difficult to forecast the total cost of ownership as they move AI from a pilot phase to full-scale production. Because token pricing is highly variable, IT leaders must demand fixed estimates and establish hard spending limits. “This is like a quarter machine," says Guillen. "You put quarters in, but you're not telling me how much a full game costs. You're not understanding that budgets work the other way around."

Because of these uncertainties, Guillen relies on pragmatic budget management to decide what leaves the sandbox. Without that filter, teams default to shipping whatever the technology makes possible rather than what the customer actually needs — adding features that create complexity instead of removing friction from interactions.

• Game over for pilots: Guillen tells his team that pilots are fine for learning, but putting anything into the core customer journey now requires a measurable ROI and a clear business case. “Play with it as much as you want, but if you're truly interested in putting this in production, you'd better justify it and make a business case.”

• Return of the paperclip: He warns that AI tools deployed without attention to the user experience risk repeating one of tech's most infamous usability failures. “Unless AI gives me a better experience—unless it's something seamless, less intrusive, working in the background—it becomes another Clippy.”

To avoid building another intrusive interface, Guillen looks to minimalist design approaches from Google and Apple, with their long-standing emphasis on minimizing buttons and creating a clean, smooth experience, and argues that AI should work the same way. In contact centers, that principle translates into tools like AI-assisted coaching that augments human agents or a unified orchestration layer working behind the scenes—not another screen customers have to fight through.

But even teams that get the design philosophy right still face a structural barrier in data governance. Many leaders are increasingly cautious about how third-party providers handle sensitive inputs. High-profile incidents, such as companies inadvertently exposing conversations from support bots, have illustrated the risks of feeding personally identifiable information into these systems.

• No undo button: The underlying limitation of AI chatbots is that it is currently impossible to cleanly remove that information once a model has been trained. Guillen says that “Vendors haven't figured out yet a way to untrain a model. If you mistakenly enter your Social Security number into an LLM model, you cannot tell the model to delete it.” For Guillen, this leads to a massive breach of trust that can't easily be reversed.

Guillen tells his team to think about AI the way consumers think about a streaming subscription. He uses Netflix as the example: the service started at $3.99 a month and now costs significantly more, but customers stay because the value is obvious and personal. AI in the customer experience has to pass the same test. "If we use AI and tomorrow they raise the prices, can we go back and say, we're getting value from this service, we don't mind paying more?" he says. "Unless you do that exercise, you need to focus on what actually saves money, saves time, makes money, or makes time."