Personal AI Agents Are Now Calling the Contact Center, and 64% Reach a Human

For two years the contact center plan has assumed that AI would sit on the company’s side of the line, deflecting calls and shrinking headcount. The newest callers are not people at all. They are personal AI agents phoning on behalf of customers, and early data shows they navigate phone menus, wait on hold and then ask for a human. The automation that was supposed to reduce contact volume is now arriving from the other direction, and it is adding calls.

A New Kind of Caller Arrived in September

Meta’s Muse app moved from novelty to mass distribution almost overnight. According to NiCE, it reached No. 1 among free apps in Apple’s U.S. App Store one week after launch and was downloaded 2.8 million times across the United States and Canada within 12 days. NiCE also points to Accenture research finding that 74% of consumers would delegate routine tasks, such as negotiation or complaint resolution, to an AI agent acting within their preferences and permissions.

Voice security firm Pindrop has been watching those calls land. Its analysis of Muse traffic, published October 9 and covering calls to its customers from September 9 to October 6, found the first matched Muse call a week before Meta announced its outbound calling beta on September 16. By week four, weekly Muse calls had reached eight times the week-one level. Financial institutions received the most Muse calls, followed by healthcare and retail.

Muse is only the most visible example. Across its network Pindrop identified 727,000 confirmed AI agent interactions over two months and 19 distinct bot types and use cases. It labeled 94% of them “good AI agents” and 6% “bad AI agents,” and it notes that some of its customers flag Muse calls as fraudulent while others treat them as legitimate. In Pindrop’s words, “adoption is happening before most enterprises have a policy for it.”

The Bot Waits on Hold, Then Asks for a Person

The most striking number in the data is how often these agents end up with a human. Pindrop found 64% of Muse calls reach a human agent, and 47% press keypad digits to work through the phone menu. A typical call that reaches a person spends 2 minutes 1 second in the IVR and then 2 minutes 18 seconds with the agent. In other words, the IVR is not stopping the bot, and the bot is not replacing the conversation.

That changes the economics of the queue. As Five9 chief marketing and growth officer Jay Lee put it in comments reported by CX Today, “If an AI agent will sit on hold for you, a lot of calls that never got made start getting made.” Friction has always rationed demand. A customer who would have abandoned a long wait can now hand the wait to software, so latent contacts that never reached the queue start reaching it.

Lee’s own analysis for Five9 argues the forecast that AI would replace agents assumed fixed demand. He cites Gartner data showing 38% of contact centers grew agent seats after bringing in AI, versus 20% that cut staff, and Metrigy research finding 42% of companies hired more people because of AI. The conversations left for humans also tend to be the harder ones, because customers accept AI for routine tasks but want a person when the stakes are high.

Two Customers on Every Call

The operating model has to change along with the volume. Pindrop describes the new arrangement as two customers: the bot calling on behalf of your customer, and the real human behind it. NiCE’s Neeraj Verma argues that “neither blocking nor welcoming is a strategy on its own,” and recommends routing every inbound conversation through one controlled entry point that identifies who or what is calling, authenticates the customer the agent represents and enforces what the agent is allowed to do, with step-up confirmation for higher-stakes actions.

Handoffs are the weak point. Five9’s 2026 Business Leaders CX Report, cited in Lee’s piece, found 96% of CX decision-makers believe their organizations preserve context during AI-to-agent handoffs, yet 83% of consumers say they sometimes have to repeat themselves. Add a third party to the call, a bot relaying a request it may only partly understand, and that gap gets wider unless agents can see the full record on one screen the moment the call connects.

A Playbook for Machine-Initiated Demand

• Forecast for more contacts, not fewer: Pindrop measured an eightfold rise in weekly Muse calls in a month. Build staffing models that assume agents will surface demand that friction used to suppress, rather than holding volume flat.
• Detect and label agent callers: Know when a bot is on the line and which one it is. Routing, scripting and fraud controls all depend on that identification happening during the live call.
• Authenticate the human behind the bot: Verify the customer the agent represents, not the agent itself, and require step-up confirmation before refunds, account changes or payment actions.
• Write a policy before the volume arrives: Decide which tasks an AI agent may complete on a customer’s behalf, which require the customer directly, and how agents should respond when a bot asks for a person.
• Give agents one screen: With 64% of Muse calls reaching a human, the agent desktop has to show the customer record and the reason for the call immediately, so the conversation does not restart from zero.
• Measure resolution, not handle time: Lee recommends judging success by resolution, retention and lifetime value. Shorter calls mean little if the bot calls back tomorrow.

The first wave of personal AI agents has not hollowed out the contact center. It has handed customers a tireless assistant that will wait on hold, push through the menu and then ask for a person. Operations leaders who treat these calls as a nuisance to block will lose control of a growing channel. Those who identify them, authenticate the person behind them and staff for the work they create will turn a surprise into a manageable queue.