Retell AI: When OpenAI Models Start Driving 100 Million Calls per Month

Opening Dialogue: The AI Voice Agent Behind 100 Million Calls

Shiwen: Recently, OpenAI’s official project mentioned Retell AI, stating that their models handle 100 million calls per month, with an average annual revenue of $1.6 million per person. This number sounds impressive. What do you think?

Yongliang: The number of calls itself isn’t the most crucial thing. What’s important is how much work the AI agents handle. If they only automatically answer and transfer calls to humans, that figure doesn’t mean much. But if AI agents can independently handle customer inquiries, appointments, after-sales support, and even sales conversion, it represents a new form of labor. I’ve led teams of 70 people and know the costs involved: recruitment, training, management, and turnover. If AI voice agents can replace some of these repetitive tasks, the company’s cost structure will be重新defined.

Shiwen: But calls are different from text chat. Voice interactions require higher standards for latency, tone, and handling interruptions. How does Retell achieve this?

Yongliang: From public information, Retell uses models like GPT-6 Astra, which are specifically optimized for voice scenarios. These models typically include a large amount of real call recordings in their training data, allowing them to better handle colloquial expressions, pauses, interruptions, and emotional changes. However, technical details are not fully disclosed, so we can only infer from the results: if they can indeed handle 100 million calls, it means their latency and accuracy have at least reached commercial standards.

Shiwen: Is $1.6 million per person per year really that impressive?

Yongliang: We need to break down this figure. First, it may only include Retell’s own employees, not those outsourced or partners. Second, the profit margin in the AI voice agent industry is already high, and software companies generally have higher revenue efficiency than traditional industries. Third, if each call costs a few cents to a few dollars, the revenue is indeed substantial. But the metric “average annual revenue per person” can be misleading—it doesn’t reflect profits or customer retention rates. What I’m more concerned about is: how much of these calls are repetitive inquiries, and how much truly creates new value.

Q1: What kind of company is Retell AI?

Shiwen: Let’s clarify the background first. What does Retell AI do?

Yongliang: Retell AI is a company that provides AI voice agents, mainly offering automated answering, outbound calling, customer service, and appointment management solutions for enterprise customers. Their clients might be medical institutions, retail brands, financial firms, logistics companies, and other industries that require extensive telephone interactions. From OpenAI’s description, Retell integrates OpenAI’s models into its voice agent platform to handle natural language understanding, dialogue management, and response generation.

Shiwen: So what’s the difference between them and traditional call centers?

Yongliang: The core of a traditional call center is people. The core of an AI voice agent is the model. There are several differences: First, the cost structure is different. In traditional call centers, labor costs are the main expenditure, while for AI agents, the main costs are computing power and API calls. Second, scalability is different. Expanding the workforce requires recruitment and training, which takes a long time; expanding AI agents only requires increasing computing power, and theoretically, it can be expanded instantly. Third, consistency is different. People get tired and their emotions may fluctuate, while AI agents can maintain standard scripts. However, conversely, AI agents lack human empathy and adaptability, and complex scenarios still require human intervention.

Q2: What does 100 million calls mean?

Shiwen: What is this figure of 100 million calls in the industry?

Yongliang: If these 100 million calls are evenly distributed, that would be about 3.3 million per day, 137,000 per hour, and 2,300 per minute. This scale is already close to the daily processing volume of a large call center. But the key is the complexity of these calls. If most of them are simple inquiries, appointment confirmations, or status notifications, AI agents can fully handle them. However, for complaints handling, sales negotiations, or crisis public relations, AI agents may only be able to provide initial diversion, and ultimately, human intervention is still needed.

Shiwen: So, what industries are Retell’s customers mainly from? Yongliang: From publicly available information, typical customers of AI voice agents include: medical institutions (appointment reminders, follow-ups), retail e-commerce (order inquiries, after-sales service), financial services (billing notifications, identity verification), and logistics (delivery notifications, appointment changes). The commonality among these scenarios is: large volume of calls, standardized processes, and sensitivity to delays. If Retell can establish itself in this market, it means their product has found a balance between accuracy and cost.

Q3: The truth behind $1.6 million in revenue per person

Shiwen: How should we understand the figure of $1.6 million in annual revenue per person?

Yongliang: This figure needs to be viewed within the context of the industry. The average revenue per person in traditional call centers is usually between $30,000 and $80,000, due to high labor costs and complex management. AI voice agent companies can have much higher average revenues per person because the core costs are computing power and model APIs, not people. But “average revenue per person” does not equal “average profit per person”. Retell needs to pay OpenAI for API fees, its own R&D costs, sales, and marketing expenses. Of the $1.6 million in revenue, only 20% to 30% may actually turn into profit.

Shiwen: What insights does this figure provide for the industry? Yongliang: It shows that the AI voice agent sector has commercial value, but competition will also be fierce. If Retell can achieve $1.6 million per person in revenue, other companies will also try to replicate it. But where is the moat? Is it model capabilities, customer resources, industry know-how, or cost control? In the long term, companies that simply package APIs will find it difficult to maintain a sustained advantage. Those that can truly survive are those that can penetrate the industry and provide end-to-end solutions.

Q4: Where are the boundaries of AI voice agents?

Shiwen: What can and cannot AI voice agents do now?

Yongliang: What can be done: automatically answering standardized processes, confirming information, arranging appointments, handling simple after-sales issues, and notifying customers by phone. What cannot be done: complex emotional counseling, multi-round negotiations, crisis management, or consultations that require deep industry knowledge. The boundary lies in “predictability”. If the dialogue process can be predefined, AI agents can handle it; however, if on-the-spot judgment and flexibility are required, humans remain more reliable.

Shiwen: So how should companies use this?

Yongliang: My suggestion is to divide it into three levels. At the first level, high-frequency, low-complexity calls should be fully handed over to AI agents, such as confirming appointments and checking orders. At the second level, medium-frequency, moderately complex calls should be initially processed by AI agents, with manual review and escalation. At the third level, low-frequency, highly complex calls should be directly handled by humans. This can reduce costs while ensuring service quality. The key is not to treat AI agents as a panacea, but as a new form of labor supplement.

Conclusion: 100 million calls are just the beginning

Shiwen: Let’s summarize at the end.

Yongliang: The case of Retell AI shows that AI voice agents have moved from proof of concept to large-scale commercialization. 100 million calls, with an average revenue of $1.6 million per person, are the result of the combined efforts of model capabilities, engineering optimization, and business models. But for ordinary companies, it’s more important not to replicate these numbers, but to figure out: which calls can be handed over to AI, and which must be left to humans. No matter how advanced technology is, it ultimately has to return to business scenarios and user experience.

Shiwen: A sentence for today’s issue.

Yongliang: AI voice agents will not replace humans, but they will redefine the role of “humans” in calls. From “responders” to “supervisors” and “decision-makers”, this may be the biggest change in call centers over the next few years.

👇 Has your company used AI voice agents? How was the experience? Share in the comments section, and we’ll discuss some more in the next issue. Follow me for honest insights on the workplace in the AI era, updated weekly.