Blog ·
Customer Support: AI trends to focus on - Agents act across accounts with money and identity
Voice AI, text ordering, and browser agents that complete purchases are now live. Support leaders must govern identity, payments, and escalation when AI acts across accounts and channels.
Jeroen Erne
This week support automation broke out of the chat window. Voice agents became faster and more expressive, text-based ordering went mainstream, and browser-based AI agents gained the ability to complete purchases on a customer's behalf. For support leaders, the question is no longer whether to automate—it's how to govern identity, payments, and escalation when agents act across accounts.
What changed this week
Voice AI took a measurable step forward. Inception launched Mercury Voice, targeting sub-300ms latency for agent conversations—fast enough to interrupt naturally. ElevenLabs released its v4 speech model with expression control across more than 90 languages. Tavus introduced Griffin, a full-duplex video agent capable of real-time face-to-face interaction. These aren't demos; they're production tools entering enterprise pipelines.
Transactional agents arrived in customer-facing channels. DoorDash launched a text-based AI agent for food ordering. Shopify opened its checkout to browser-based AI agents, letting them complete purchases directly. xAI released Team Bots for shared enterprise workflows, while Meta expanded its Muse AI agent to small businesses. The pattern is clear: agents are moving from answering questions to taking action with money and accounts.
Security and trust became urgent topics. A deepfake voice scam targeting a grandparent inspired a new identity-verification startup. Reco raised $55 million as demand for AI-agent security grew. Meta faced claims that Muse accessed private messages without permission—claims the company disputes. Instinct's proactive product recommendations triggered user discomfort, with some customers calling the behavior intrusive.
On the infrastructure side, Cohere released Parse 5 for enterprise document extraction and Embed 5 with a retrieval metric focused on rank consistency. ElevenLabs added instruction-based transcript editing to its speech-to-text API. OpenAI expanded ChatGPT plug-ins with app-like interfaces and automations, and launched Dots as always-on agentic avatars. EliseAI raised $350 million at a $4 billion valuation, signaling investor confidence in verticalized support AI.
What it means for you
Your agents—human and AI—will soon operate across voice, text, and video, often in the same customer journey. A customer might start with a text order, escalate to a voice call, and complete a transaction through a browser agent. If your identity verification and escalation paths don't span those channels, you'll create gaps that frustrate customers and invite fraud.
When AI agents can spend money or modify accounts, permissions become your most important design decision. Shopify and DoorDash are building consent into their flows, but the responsibility sits with you when those agents operate inside your support stack. Define exactly what an agent can do, what requires human approval, and how you'll recover when the agent gets it wrong. If you can't answer those three questions for every automated action, pause deployment.
The voice improvements are real, but they raise the stakes. Low-latency, expressive speech agents feel more human, which means customers will treat them like humans—sharing sensitive information, getting frustrated, expecting empathy. Your quality measurements need to capture resolution accuracy, not just containment rate. A fast wrong answer is worse than a slow transfer to a person.
Multilingual support is becoming table stakes. ElevenLabs' 90-language coverage and the broader push toward multilingual automation mean customers expect service in their preferred language. But language fluency doesn't equal cultural competence. Test your agents on local idioms, complaint patterns, and escalation expectations before you scale.
What to focus on next week
- Audit every automated action your agents can take across accounts. For each one, document the permission boundary, the approval trigger, and the recovery process if the action is wrong. If any of those three is missing, restrict the action to human-only until it's resolved.
- Pick one voice channel where latency matters most—likely your phone support line—and test whether your current stack can handle sub-500ms response times. If not, start evaluating low-latency options like Mercury Voice or ElevenLabs v4 with a specific use case, not a general deployment.
- Review your identity verification flow for voice interactions. Deepfake scams are no longer theoretical. Ensure your system can distinguish a real customer from a synthetic voice before any account changes or payments are authorized.
- Set a per-resolution cost target for your automated support. As agents handle more complex transactions, costs will rise. Know what you're willing to spend to resolve a billing dispute versus a password reset, and build those thresholds into your routing logic.
- Run a trust audit on any proactive or outbound agent behavior. If your AI reaches out to customers with recommendations or follow-ups, verify that customers explicitly opted in and can easily opt out. Instinct's user backlash this week shows how quickly proactive features can erode trust.
These stories and more are collected in the all Customer Support AI news archive, updated daily throughout the week.