|
Welcome back! Artificial intelligence has gotten great at automating customer support emails and text chats. But automating customer support phone calls is far more challenging because AI that can understand and generate human-sounding speech is still too costly, slow and unreliable, said Jesse Zhang, CEO of Decagon, a developer of AI for customer support. Decagon gets around the voice AI limitations by transcribing audio to text, then feeding it into an AI model that processes text. After the model produces a text response, Decagon’s system uses models from ElevenLabs, OpenAI or Deepgram to convert the text back to audio. That process means Decagon’s voice AI product is relatively slow compared to models that only handle audio, but it’s necessary to verify the AI doesn’t give false information—a reputational risk for companies like Hertz that use Decagon software, he said. “I can’t imagine many other folks are actually doing speech-to-speech models” [for customer support] because after you’ve talked a bit it’s very clear that it just will hallucinate,” said Zhang, who co-founded Decagon in 2023 and has less than 100 paying customers. The slowness of voice AI is another big reason it hasn’t been widely adopted compared to text-based customer support, according to executives I’ve spoken with. This reality could be challenging for OpenAI and another startup, Sesame AI, both of which have been looking to build a business around voice AI models. Decagon rivals such as Salesforce, Crescendo and Sierra also said they are launching voice AI for similar purposes. Zhang said AI researchers have continued to find ways to improve voice capabilities—meaning costs could come down significantly in the next few years. Customers have also started to warm up to the technology after some initial skepticism about its performance, he said. “Last year, the buyers weren’t really ready for voice, but now they are,” Zhang said. We’ve written about the blurring of the lines between functions like design, engineering and product management because workers now have access to AI tools that can write code or spin up a webpage in a pinch. That’s prompting Figma, the design software startup headed for one of this year’s most highly anticipated initial public offerings, to put AI at the front and center of its products. At its annual Config conference yesterday, the company announced Figma Make, which turns text prompts into working code for designs or apps. The launch also lets designers easily write code or engineers to create product mock-ups, highlighting how the two areas have been coming closer together. Vibe coding startups like v0, Lovable and Bolt have taken off with customers in recent months thanks to how easy they make it for nonengineers to design webpages and apps. “People are designing via prompting, people are designing via coding, and those are all legitimate ways to design,” Figma Chief Product Officer Yuhki Yamashita told me ahead of the Make launch. “We think that it’s really important to have a platform that allows for all that to happen and kind of commingle all those different inputs or mediums.”
| 
