Vapi raises $50M Series B
Read More →
Vapi raises $50M Series B to power the next generation of enterprise voice AI
Vapi raises $50M Series B
Read More →
When used with the right technology, voice agents help brands create higher-quality interactions with their customers, multiplying their revenue while reducing costs and creating efficiencies. What is often the hardest part of realizing these results is not the technology itself, but choosing the use case that enables it.
Our teams have reviewed thousands of voice deployments, many of which have succeeded and some of which have not. Over time, the commonalities and themes of those who have succeeded the most have emerged. The pattern is clear: five characteristics predict whether a voice agent will work. Hit all five, and you have a use case ripe for voice; miss one, and you can compensate, but if you miss three, you are better off exploring alternative use cases.
Voice agents have a fixed cost, not just in monetary terms but also in your time and resources. You're building conversation flows, connecting tools, testing edge cases, and monitoring performance. That investment only pays off if you spread it across enough calls. Fifty calls a week will not justify the effort, while fifty thousand calls a month might. Even if those calls drive your intended outcomes, they will not move the needle to justify the resources. For a use case to scale in voice, it must have the volume required to solve a business problem related to cost, efficiency, or revenue.
Good Fit Example: During Medicare's Annual Enrollment Period, carriers and agencies are processing enormous call volumes, often thousands per day, against leads that have opted into communication. The volume isn't just high; it's concentrated at a certain period of time. October through December is the peak window, with a January restart period immediately after. Companies that can't process leads fast enough during that window lose them to competitors who can. A high volume of calls that human front-liners cannot field is a use case ripe for voice. Seasonal spikes are some of the strongest volume cases because the alternative (hiring, training, and then letting go of temporary staff) is operationally painful and expensive.
Bad Fit Example: A regional insurance brokerage handles about 40 inbound calls a week from existing policyholders asking about renewals and coverage changes. The calls are straightforward, and the agents who handle them could be replaced by a voice agent from a conversation-design standpoint, but at that volume, the math doesn't work. The time spent building the conversation flow, integrating with their policy management system, testing edge cases, and monitoring performance would take weeks of effort spread across a few hundred calls. A senior account manager handling those 40 calls is a rounding error on their salary. The brokerage doesn't have a volume problem; it has a small, manageable workload that happens to involve phones.
Voice agents don't need identical scripts. They need variations on known themes. The conversation can branch, but the branches should be recognizable. If you can map the conversation to a decision tree before you build, the agent can follow it. If every call requires the agent to figure out what kind of conversation it's in, you're asking it to do the part that's hardest to automate.
Good Fit Example: Health insurance qualification is a structured variation. The core workflow is consistent: confirm the caller is enrolled in Medicare or plans to enroll; collect their name, date of birth, and zip code; validate the zip code against serviceable areas; and determine whether they qualify to speak with a licensed agent. The questions shift slightly depending on whether the call is inbound or outbound, and routing logic varies by geography, but the structure is the same every time. The decision tree is short, the branches are well-defined, and the edge cases are all known in advance.
Bad Fit Example: A mid-size health plan's member services line handles calls from enrollees about claims, benefits, grievances, prior authorizations, and pharmacy issues, all through the same queue. If a caller asks why their MRI was denied, the agent must pull the specific claim, interpret the denial code, verify that the provider submitted the correct prior authorization, and determine whether the member's plan even covers that procedure at that facility. The next caller wants to know why their copay changed, which requires navigating a different set of plan documents and a different system. Both calls are "member inquiries," but they share almost nothing structurally, and each call is its own investigation.
The best use cases have unambiguous outcomes. Is the candidate qualified? Did they book an appointment? Has the payment been collected? When you can define success as a discrete event that either happened or didn't, you can measure whether the voice agent is working and improve it, and adapt it when it is less effective. When success requires judgment, the questions get harder. Did this particular customer feel heard? Was the right level of empathy applied? You're in a territory where voice agents may struggle. Not because they can't say appropriate words, but because measuring whether they succeeded requires an evaluation that's hard to automate, and it becomes hard to adjust, adapt, and improve when you cannot measure the outcome.
Good Fit Example: In health insurance qualification, success is binary: was the lead qualified appropriately and transferred to a licensed agent? You can measure it precisely: transfer rate, time-to-transfer, transfer-to-close rate, and data capture accuracy. The clarity of the success criteria also makes iteration straightforward. If your transfer rate drops, you know where to look. If data capture accuracy falls, you can trace it to specific questions in the flow. Every metric points back to a specific part of the conversation, which means every problem has a diagnosable cause.
Bad Fit Example: A behavioral health provider uses a nurse triage line to handle after-hours calls from patients in emotional distress. The goal is to assess whether the caller needs immediate intervention, a next-day appointment, or just reassurance that what they're feeling is manageable. On paper, that looks like a triage decision tree, a route to one of three outcomes. In reality, there are nuances that require human involvement; the nurse is listening for tone of voice, sentiment of conversation, and other key indicators. Success isn't whether the call was routed to the right bucket. Success is whether the nurse caught the thing the patient wasn't saying, and there's no metric for that. When the definition of success is "the right judgment was applied to an ambiguous situation," you don't have a measurement problem; you have a use case that requires human touch.
A voice agent is only as capable as the systems it connects to and can take action on. If your CRM has a clean API, the agent can write customer data in real time and have full context on the information you have already gathered about the customer. If your current use case requires a human to manually navigate across three screens and copy and paste between windows, the voice agent can only collect information and will not be able to do much with it. In this case, it's not automation; it's a complicated answering machine. The question isn't just whether your use case is a good fit for voice; it is whether your infrastructure can actually support what the voice agent needs to do once the conversation is underway.
Good Fit Example: Health insurance qualification typically involves a small number of integrations with existing APIs. Teams that have deployed voice agents for this use case report that small engineering teams can stand up the full integration in a matter of weeks. That timeline reflects the backend simplicity: they're not building custom middleware or reverse-engineering legacy systems. They're connecting well-documented APIs to a well-defined workflow.
Bad Fit Example: A workers' compensation insurer wants to automate the first notice of loss, which is the call a policyholder makes to report a workplace injury. The conversation itself is structured enough: what happened, when, where, who was involved, and what treatment was received. But completing the intake requires the agent to enter data into a claims management system built in the early 2000s that has no API. Today, adjusters enter data through a desktop application that requires a VPN, a specific browser version, and a sequence of screen navigations that varies by the state the claim originates in. A voice agent could perfectly collect every piece of information from that call, but it has nowhere to put it. The data would land in a holding queue, waiting for a human to key it into the real system, which means you haven't automated the process; you've just moved the bottleneck from the phone call to the data entry desk.
Voice agents answer immediately without hold times or callbacks. If that urgency matters to your business, voice agents have a structural advantage that human teams can't match at any meaningful scale without incurring enormous costs. At the same time, some conversations are worth just as much whether they happen now or tomorrow morning, while others lose most of their value in minutes. The use cases where voice agents have the strongest structural edge are the ones where delay has a direct cost.
Good Fit Example: When a prospect calls in about Medicare plans, they have intent right now, and the best-performing teams target qualifying and transferring an inbound caller within 90 seconds. That's not a nice-to-have; it's the difference between capturing the lead and losing it. When there is a voice agent available to answer the call, regardless of the time of day or the number of callers, it will help deliver immediate value to the brand.
Bad Fit Example: A specialty pharmacy calls patients to coordinate refills on maintenance medications, cholesterol drugs, blood pressure medications, and thyroid prescriptions. The calls are routine: confirm the patient still wants the refill, verify the shipping address, and check whether anything has changed with their insurance. At the same time, the patient's prescription doesn't expire today. The pharmacy batches these calls and works through them over a two-week window before the refill date. Whether the call happens Monday morning or Thursday afternoon makes no difference to the outcome. The calls that benefit most from voice agents are those where a 30-second response time yields a materially different outcome than a 30-minute response time. When the value is the same either way, speed is a feature nobody asked for.
These five characteristics don't just add up; they compound. High volume means the fixed costs get amortized. Predictable patterns mean the agent handles most calls without escalation. Clear success criteria mean you can measure and improve. Strong backends mean the agent can actually do something, not just collect information. And time sensitivity means the speed advantage translates directly to business value.
Health insurance qualification meets all five criteria, which is why it's one of the strongest use cases for voice AI in production today. But the framework applies to any use case you're evaluating. Before you pick a platform, before you write a prompt, before you design a conversation flow, run through the five characteristics to make sure you have the right use cases in place before proceeding with implementation. If you're hitting four or five, build with confidence. If you're hitting three, proceed carefully and know where you'll need to compensate. If you're below three, the technology isn't the problem; the use case is.
This checklist is Chapter 2 of the Voice Agent Playbook. If you're hitting four or five of these, the next step is scoping your first agent. Chapter 7 covers how to prioritize use cases. Chapter 8 covers how to scope your v1 without overbuilding.
