CX-ready AI: Advice from real-world rollouts

Shane Schick tells stories that help people innovate, and to…

You’ve tried using generative artificial intelligence (AI) to summarize contact center transcripts, or maybe to write social media posts. Over in sales, your colleagues are experimenting with agentic AI to manage reps’ calendars and proactively reach out to customers whose contracts are expiring.

Both forms of AI could also do wonders in customer-facing scenarios. Vendors say so all the time. But how do you really know if the tools you’ve bought or are evaluating will work as expected? Yes, you know you need “AI-ready data.” But what truly defines CX-ready AI?

For Bob DePonte, director of customer experience at HR, pay, and workforce management solution provider UKG, CX-ready AI is like a three-legged stool. The first leg represents the idea of customers being at the center of however AI is used. The second leg is a reminder that UKG is focused on nurturing personal and proactive relationships. And the final leg is that AI should leave customers feeling confident in the solution, answer or expertise they’ve been offered.

“It took a lot of time, energy and work – and a lot of trial and error – to get that AI mission in concert with our service philosophy, and reaching that intersection of putting the human element and AI together” DePonte told 360 Magazine. “Our customers should feel like they’re being taken care of an superior way to the rest of the market.”

Test 1: Will customers see AI as better than traditional IVR?

UKG, which supports millions of frontline and salaried workers, has deployed an agentic AI-powered Intelligent Virtual Assistant (IVA) across voice and digital channels that now handles more than 1,500 customer interactions per day, including complex HR and payroll inquiries.

At scale, 38 percent of voice calls remain fully within the IVA, either resolved directly or with a case created automatically. The system has also driven three percent explicit deflection and 16 per cent implicit deflection, reduced average handle time through full-context handoffs, and lowered cost to serve without linear headcount growth.

UKG tested its IVA with a subset of customers, and DePonte said it was clear that many initially treated it like an interactive voice response (IVR) tool or wanted to speak to a human first.

“One of the things that we learned quickly is the AI assistant needs to be smart. It needs to be personal, almost at the moment in which you actually talk to it,” he said, such as recognizing any tickets the customer has open and asking whether there’s a new issue. “That’s when the person recognizes, okay, this is not, just a phone loop. This is actually personal to me.”

Initially, the UKG leadership team read every transcript, and ran another AI agent over all the transcripts to understand their sentiment. This helped identify where the IVA was effective in triaging issues and when customers needed to be connected to a human employee more quickly, DelPonte said.

Test 2: Can your team come together to determine AI’s CX blind spots?

AI offered a similar sense of promise and opportunity at General Assembly, which provides online education across a wide range of subjects to learners all over the world. A key part of the customer journey is when prospective students call in to find out if they’re eligible to take a particular course and if they can afford it.

Jourdan Hathaway, GA’s chief business officer, said the company had set up what she called a “digital worker” to determine eligibility, and then pass qualified individuals to a human admissions professional to enrol them.

While she’s pleased with the results, it has taken considerable effort. Hathaway said she is currently coordinating 11 people across eight business functions in a daily standup-meeting to look at the AI system through different lenses.

Historically, she said, humans in tech, ops and functional teams have figured out how best to navigate the messy reality of business logic, exceptions, dependencies, compliance and customer expectations. In some respects, CX-ready AI is a matter of helping the technology fill in gaps about nuances it couldn’t possibly know.

The team noticed, for example, that a certain number of potentially eligible students were dropping off calls before getting passed onto the admissions offer. Only by carefully going over every step of the interaction did it become clear that when the digital worker said, “I’m going to transfer you to an admissions officer. Thanks, and have a great day!” the person on the other end assumed the call was over.

“It wasn’t a compliance problem. It wasn’t an integration or a data problem. It was human psychology 101,” she said, adding that AI can also be confused by prank calls involving knock-knock jokes (and yes, middle school kids are still making them).

“Every organization has tribal knowledge of what to do when a situation looks like XYZ,” she said. “There is this level of expertise that employees have to know what to do. You can’t possibly forecast every fringe case, but when they come up you have to wire them back into the systems that run the business.”

Test 3: Can your approach to AI overcome customer hubris?

It’s not just that some customers don’t want to use AI or talk to an AI agent. Some simply feel they don’t need it, which means they won’t make the most of the AI-powered experiences offered to them.

Yoodli is a good example of how this problem manifests itself in a business-to-business environment. While it began as a tool for public speakers to get helpful feedback before standing in front an audience, Yoodli has since expanded to help enterprise employees handling high-stakes customer conversations. This could include customer success teams, for instance, who are helping onboard or troubleshoot issues.

Esha Joshi, Yoodli’s co-founder, president and chief customer officer, admitted that human behavior remains the biggest issue in designing CX-ready AI. In her firm’s case, for instance, customers can find it uncomfortable or “cringey” to watch themselves and see how AI assesses their performance. In some cases, the attitude is even more skeptical.

“There can be a hubris component, where some folks are like, ‘I’m great. I’ve been doing this for 10 years, 15 years, I know exactly what to do. I don’t need an AI assistant to help me improve,’” she said.

Joshi said Yoodli addresses that by “pressure testing for realism,” running beta tests internally with frontline reps to better understand scenarios involving emotional customers, ambiguous sales objections and other grey areas.

“It just kind of shows people that we care about their feedback,” she said. “The deployment is deliberate and not rushed.”

From there, Yoodli can provide a mix of both objective feedback (like noting how often someone cuts off a customer or uses “um” too often) and qualitative suggestions that build trust.

Joshi said CX-ready AI is ultimately less about choosing a particular tool than a set of people and process changes, as well as an ongoing operational discipline.

“Sometimes people will treat AI as a feature rollout, but you have to treat it like a behavior change to see the wins,” she said. “And I think if that happens from the top down – and then you also hire for folks that are excited by this kind of behavior change – I think it will go a lot smoother.”