You want to know the transfer limit set on your account?
– Our customer service is available 24/7, 365 days a year. I was an AI assistant for OO Bank. Thank you.
– Click.
After directly inquiring with one telecommunications company and four banks that have implemented artificial intelligence (AI) call centers, none were capable of engaging in a conversation.
They failed to resolve complex queries like “What’s the lowest interest rate credit loan I can get?” as well as simpler requests such as “Please check my daily transfer limit” or “What’s my current account balance?”
While these companies designed their systems to route customers through AI assistants before connecting to human agents, this approach only seems to delay complaint resolution.
As a result, it’s easy to find online criticism directed at AI call centers.
Comments on YouTube videos related to AI call centers reveal frustrations: “You end up talking to a human agent feeling like you’ve wasted time and with frayed nerves,” “What could be resolved with one statement to an agent turns into a wild goose chase,” and “I called because I was angry, but this makes me even angrier.”
In fact, a recent survey by the Korea Confederation of Trade Unions’ Central Research Institute of 706 citizens with AI call center experience found that 54.2% of respondents were ‘dissatisfied’ with the service, and 87.5% said they ‘prefer human agents.’
AI Consultation Development Hindered Without Wholesale Changes
AI experts explain that current AI assistants are limited to helping users navigate app or website menus. They lack the capability to handle inquiries complex enough to warrant a call to customer service in the first place.
Professor Joo Jae-geol of the Korea Advanced Institute of Science & Technology’s (KAIST) Kim Jaechul Graduate School of AI explained tha just as it often see categorized menus when it clicks on all menus in an app, the structure for AI assistants is predetermined during development. They simply convert user speech to text and classify it within a pre-established menu structure.
This type of recognition-based AI is incapable of organic communication due to its fixed turn-taking nature.
Professor Joo added that it’s fundamentally impossible for current AI assistants to interject with I have a question during a consultation, or for users to correct themselves saying No, what I meant was this. Communication is limited to a walkie-talkie-like exchange where one party speaks and then the other responds.
He suggested that significant performance improvements would require a wholesale replacement with systems based on Large Language Model (LLM) generative AI or agentive AI, integrated with various resources and databases.
Smarter AI Comes At a Cost… But Users Aren’t Beta Testers
Upgrading AI assistants to high-performance models presents challenges in terms of cost and unpredictability.
Professor Joo noted that recognition-based AI assistants don’t require extensive computing power or high-end hardware, making them very cost-effective to operate. However, LLM-based agentive AI assistants would incur significant operational costs.
He added that if it overcomes the cost issue and create smarter AI assistants using agentive AI technology, their capabilities and freedom of action would greatly expand. However, this also introduces the problem of unpredictability.
He warned that for instance, an AI assistant might access and expose sensitive data to users or suddenly cancel or change service plans. Even with high-performance AI assistants, final decision-making should still involve human oversight.
Currently, a compromise solution involves AI assistants quickly recognizing their limitations.
Professor Joo emphasized that users are not beta testers for companies developing AI assistant services. If an AI assistant realizes it can’t understand or properly respond to a user, it’s crucial to have technology that quickly abandons the consultation and transfers to a human agent.