Say goodbye to point-and-click applications and chatbots. “Digital humans” (a.k.a. AI-powered avatars) could be the future of human-computer interaction, with transformative effects on telcos.
Why? Apparently, digital humans promise more natural interactions, better customer satisfaction and lower operating costs, according to panelists at the recent Nvidia GTC conference.
“Anywhere you have a chatbot today, you can replace it with a digital human,” said Alan Dennis, professor at the Kelley School of Business at Indiana University.
These digital humans are poised to replace traditional chatbots and point-and-click interfaces in telecom customer service thanks to their use of conversational AI which delivers more human-like, engaging interactions, potentially improving satisfaction and driving sales, he said.
AI avatars will also be used as “digital concierges,” the online equivalent of the helpful salesperson who greets you when you enter a high-end brick-and-mortar store.
These digital concierges can replace application interfaces that rely on point-and-click, searching and filtering. The concierge can also explain how a product meets your needs as a consumer, Dennis said.
(Mitch Wagner)
Additionally, digital humans are useful in generating training videos for employees, partners and customers. These videos, generated for teaching people how to use products and procedures and for regulatory requirements, are now labor-intensive to produce, Dennis said.
“Traditionally, what we do is we hire an actor, we film them, we edit the video and that takes a while. If anything changes, you get the actor back and re-do the whole thing,” Dennis said. “With a digital human, you write the script and press ‘generate,’ and you get the video. If anything changes, change the script and regenerate the video.”
He added, “The cost savings and time savings are immense.”
AI avatars make users happy
Researchers compared the effectiveness of AI customer service agents against text-based chatbots, human agents, and point-and-click interfaces. Customers are more satisfied interacting with a digital human than a text chatbot, by a difference of 10-20 percentage points, Dennis said.
Consumers rate product quality higher when digital humans pitch products, compared with when those products are discovered on a point-and-click interface or via a chatbot, Dennis said. Consumers are also more likely to buy from digital humans.
There are qualitative differences in users’ interactions with digital humans compared with chatbots. With chatbots, users typically use short, clipped language, using the smallest number of words to get the job done. “That’s because, for a lot of us, typing is harder than just speaking,” Dennis said.
He added, “But whenever you speak to an AI agent, the interaction is much more like a human because we just naturally talk,” Dennis said. The engagement tends to be longer, and the language is more like a human interaction.
Users are more forgiving of errors by digital humans than chatbots because users transfer the same tolerance of error to digital humans that they give to real people, Dennis said.
Hallucinations? @%$# no!
But while users will tolerate minor errors by AI avatars, they won’t tolerate hallucinations, Dennis said. Again, this is a case of applying the same standard for digital humans that we do for real people. “Normally, if you hallucinate, it’s a sign you’re high or mentally ill. So that’s why people don’t like hallucinating digital humans,” Dennis said.
Also, users are less likely to swear at digital humans when they mess up compared with chatbots and less likely to complain about service too.
But digital humans have problems as computer interfaces. One is the “uncanny valley” — if the visual representation is too realistic but not perfectly real, many users will be repelled.
Digital humans can have difficulty understanding conversations when not speaking in the user’s native language.
Dennis views both problems — the uncanny valley and the language barrier — as short-term, to be solved in a year or two.
Appearances matter (maybe)
Another problem is deciding how the digital human should present. “Digital humans can look and sound like anyone. So the question is, who should they look and sound like?” Dennis said. Whenever I work with companies, the first question is, should it be male or female?” The university studied Americans and found that both men and women preferred female digital humans.
“The face of AI will be female,” Dennis said. “Sorry, guys.”
How old should the digital human appear to be? The sweet spot for age is somewhere in the mid-30s, Dennis said.
People want to deal with digital humans reflecting their own race.
What type of accent should the AI customer service agent speak with? “All of us have accents, so whenever we program a digital human, we have to decide what accent we want to give it,” Dennis said. “Accent is a cue to where you were born, where you lived and it’s also key to socioeconomic status. So we have to choose carefully.”
He added, “That’s just the tip of the iceberg. There are dozens of things we have to think about in the design of digital humans that we don’t think about in the design of tech-based interfaces.”
However, surprisingly, none of these questions of voice and appearance are reflected in trust scores. “We were very surprised. What we found is that people have clear preferences, but whether they got their preferred age, race, and gender had no effect on satisfaction, trust, or willingness to use,” Dennis said.
Dennis’s panel also featured speakers from AT&T, Amdocs, ServiceNow and was moderated by Lilach Ilan, Nvidia global head of business development and telco operations.
In its own research, Amdocs found that allowing users to customize digital humans increased trust, creating a connection beyond technology. Generation Z and Generation Alpha might prefer an anime persona to looking at a human, so Amdocs provides those types of avatars, said Anthony Goonetilleke, Amdocs group president, tech and head of strategy.
AI behind the scenes
The digital human is the user interface or presentation layer. Behind the scenes, on the back end, are agentic AIs, LLMs and old-fashioned rules-based automation, Dennis said. And the digital human and back-end intelligence need to be balanced against each other.
“If you have a highly realistic-looking digital human, you need to make sure the intelligence behind the scenes is up to the task, or else people are going to have expectations you cannot meet,” he said.
If you have a highly realistic-looking digital human, you need to make sure the intelligence behind the scenes is up to the task, or else people are going to have expectations you cannot meet.
Alan Dennis, Professor, Kelley School of Business, Indiana University
On the back end, ServiceNow is building agentic AI for network automation. In March, the vendor released an AI agent to automate service, test and repair. “We found that a lot of time in resolving trouble tickets is spent on identifying what kinds of tests to carry out, what sequence to carry them out in, and, based on the results of those tests, what kinds of repair actions to do,” Romit Ghose, ServiceNow head of telecom and media said.
AT&T is aggressively using AI for support at a rate of about 2 billion tokens daily, processing 500,000 API queries, with heavy use of agents, Mark Austin, AT&T VP, data science, AI and automation, said. The company uses AIOps to troubleshoot network problems using agents. AT&T uses AI agents for its entire software coding lifecycle, even including an agent that joins scrum calls and updates Jira support. Other agents create code and test for security vulnerabilities, resulting in 20% more commits and pull requests.
In 2023, AT&T had a rules-based chatbot on its customer-facing website that could answer about 60-70% of incoming questions. By adding generative AI support to the chatbot, AT&T was able to vastly improve the rate of questions it could answer, Austin said.
Beginning in 2023, AT&T began working with Nvidia to experiment with digital humans, using a cartoon character for support. However, the latency was not adequate, Austin said. The telco is now in a trial using realistic digital humans.
AT&T measures effectiveness against a “three-legged stool” of cost, latency and accuracy, Austin said.
AI might make business less fun
ServiceNow sees digital humans as an opportunity to improve call center return on investment, as measured by the key performance indicators of average handling time, whether the problem resolves on the first call and the net promoter score measurement of customer satisfaction. In production trials, all three KPIs are significantly positively impacted by using digital humans, with double-digit improvement, Goonetilleke said.
British Telecom saw mean times to resolution reduced by 33% using AI, and Amdocs expects vastly better performance with agentic AI, Ghose said.
AI will likely become significantly more autonomous within two to five years, Goonetilleke said. He illustrated the anticipated change with an anecdote: The morning of the panel, he learned the Wi-Fi Internet at his home was out. He called his service provider, and they sent someone out to resolve the problem. He expects that soon, an AI will discover the AI outage and call directly to get it resolved.
Austin agreed. In two years, AI agents will communicate with each other without requiring human involvement. Transactions will occur more quickly because AI doesn’t have to talk in English; they just talk in bits.
However, the transition to AI agents might make business less enjoyable. “For some of us, it’s a bit of fun to negotiate,” Austin said. “Agents don’t really negotiate.”