By Nicholas Khami, Founder, call4.me
By the end of 2027, AI agents will routinely complete work over the telephone, not just inside browser tabs. The important change will not be a voice that sounds more human. It will be software that can call a business, navigate a phone tree, wait on hold, speak to either an automated system or a person, and return with a result the user can verify.
Today, most discussion of agentic AI focuses on software that operates websites, writes code, or moves information between APIs. That leaves out a large part of the real economy. Many businesses still require a phone call to cancel a service, change a reservation, resolve an account problem, or ask a question that is not covered online. In 2027, the phone will become another tool an AI agent can use.
At call4.me, I build systems that let AI agents make real calls, and I record and publish test calls to major support lines. On October 8, 2026, we tested nine companies. Four of the nine put an AI assistant in front of the line: Delta, PayPal, DIRECTV, and Hertz. This is a small sample, not a universal benchmark, but it shows what agents will encounter in practice. A calling agent will often be speaking to another automated agent before it ever reaches a person.
The first prediction is that voice agents will need to operate across mixed environments. In our Delta test, an AI voice assistant answered first and a live agent picked up after about 16 minutes. The audio and transcript are published here: https://call4.me/blog/delta-customer-service. In our PayPal test, an AI assistant also answered first, but a person picked up at 4:30: https://call4.me/blog/paypal-customer-service-number. An agent cannot assume the other side is human, and it cannot assume an automated menu will behave like a clean software interface. It has to detect what it is talking to, preserve the user’s intent, and adapt without inventing progress.
The second prediction is that waiting will become machine work. The immediate value of a calling agent is not replacing an entire call center. It is taking over the dead time and repetitive navigation that people currently absorb. Hold music, repeated prompts, transfers, and requests to restate the same problem are poor uses of human attention. Software can wait, keep notes, and alert the user when a decision or sensitive disclosure is required.
The third prediction is that reliability will matter more than voice realism. Some human-first support lines were faster in our tests. Hulu connected us to a person at 1:28 without an AI assistant: https://call4.me/blog/hulu-customer-service-number. At Pottery Barn, saying “customer service” reached a person at 0:49: https://call4.me/blog/pottery-barn-customer-service. The lesson is not that AI or humans always win. The lesson is that a useful calling agent must learn the actual path that completes the task. A natural voice is cosmetic if the agent cannot recognize that a short phrase is the fastest route.
The fourth prediction is that proof of completion will become part of the product. Users will not be satisfied with an agent saying, “It is done.” They will want a recording, transcript, confirmation number, or precise account of what changed. This is especially important for cancellations, billing questions, reservations, and financial services. The agent should separate what it attempted, what the business confirmed, and what remains uncertain. In my view, that audit trail will become as important as the conversation itself.
The fifth prediction is that a human escape hatch will become a design requirement. Our calls exposed several failure modes. Fifth Third’s account-number prompt looped before the call transferred to a person at 1:44. DIRECTV only connected the caller to a person after the caller said “cancel service” and said there was no account. Hertz’s virtual assistant dropped one call when our caller went quiet after she promised a transfer. These are not dramatic science-fiction failures. They are ordinary interaction failures, and that is exactly why they matter.
Calling agents will therefore need clear rules. Saying “representative” and declining to provide an account number usually still reached a person in our tests. Staying silent could get the call dropped. A capable agent should know when to keep trying, when to ask its user for information, when to request a person, and when to stop rather than claim success.
By the end of 2027, the best AI agents will be judged less by how convincingly they speak and more by whether they finish the task, preserve evidence, and hand control back cleanly when they cannot. The phone will not disappear. It will become part of the agent tool stack, alongside the browser, email, code, and APIs.
##
ABOUT THE AUTHOR
My name is Nicholas Khami. I have 10,000 followers on LinkedIn and 18,500 on X with over 75 million impressions in the past six months. I am a Y Combinator founder, have raised $3.5 million in venture capital personally, and was head of engineering at a Series B a16z startup. I am the founder of call4.me, which lets AI agents make phone calls and publishes recordings and transcripts of real calls to major support lines.
