AI customer service by phone: what an AI voice agent can actually handle

Which customer requests can an AI voice agent handle on its own?

The ones whose answer is already written down somewhere in your systems. Opening hours, a delivery window, the status of an order, whether a slot is still free, the steps for returning an item: the agent reads it, says it back in plain language, ends the call. Same logic for booking, moving and canceling appointments, and for qualification before a transfer (name, order number, reason for calling), which hands your reps the first thirty seconds of every call back.

Retell AI gives you the whole platform from its pay as you go tier, with 20 concurrent calls included. Vapi, which we score 4.5 out of 5 on features, lets you pick the model, the voice and the transcription yourself, so you can tune the agent request by request. If all you need is answering the line and taking messages, our guide to the AI receptionist already covers that simpler case.

When should the agent transfer the call to a person?

As soon as the answer commits your company. A goodwill gesture, a discount, a refund, a dispute, a cancellation, a complaint that is heating up: those are decisions, not lookups. The same goes for anything touching sensitive data, or an identity you genuinely need to verify.

There is a second, less obvious category: calls where the customer has already tried. A second call about the same problem is a signal to escalate, not a request to run through the script again. In our reading, the transfer should not be treated as a failure to minimize but as a feature in its own right, and your concurrency limit is what decides whether it goes through fast or leaves people on hold. Retell AI includes 20 concurrent calls and charges $8 per month for each extra one. At Vapi, concurrency is tied to the plan: 4 with no subscription, 10 on Core, 30 on Pro, custom on Premier.

What does a minute of AI customer service cost at Retell AI and Vapi?

Both vendors bill by the minute, but they do not put the same things behind the posted rate.

Vapi’s $0.05 looks lower than Retell AI’s $0.07, but the two figures do not cover the same ground: at Vapi, models, transcription and voice are passed through at provider cost with no markup, and they stack on top of the hosting fee. We break the Retell side down line by line in what a minute of calls really costs.

What do you pay on top of the per minute rate?

The rest of the budget sits elsewhere. At Retell AI the side items are legible: $2 per month per phone number, $8 per month for each concurrent call beyond the 20 included. Support stays limited to the community and email outside the Enterprise tier, a custom quote that brings volume pricing, SSO, role-based access control, an optional dedicated server and 24/7 support.

At Vapi, the middle step is Core at $29 per month on top of usage, which brings 30 days of retention, ZDR and email support within two business days. Above that, the jump is steep: Pro is 10% of your Vapi hosting fees with a $999 per month minimum, for a 99% uptime SLA and dedicated Slack support. HIPAA compliance, with standard DPA and BAA, is an add-on billed at $2,000 per month. We score Vapi 2.5 out of 5 on ease of use: every component is yours to assemble, and it is built for developers.

Where does your European customers’ call data end up?

This is the real dividing line between the two. Retell AI’s documentation describes processing and storage in the United States on AWS, and states that the vendor does not yet operate services inside the European Union. The contractual guarantees do exist: SOC 2 Type 1 and Type 2, HIPAA, and a DPA with EU standard contractual clauses you can sign online at no cost. Hence our GDPR score of 2.5 out of 5.

Vapi documents infrastructure mainly on AWS with deployments in the United States and the EU, each organization belonging to a single region. Its Trust Center lists SOC 2 Type II and SOC 3 (July 2025 to July 2026 period), HIPAA, GDPR, CCPA and PCI DSS v4.0.1, for a GDPR score of 4.5 out of 5. One caveat to know: the documentation does not state how you get access to that EU region, whether self-serve or on request. In our reading, that is the question to ask in writing before you sign, ahead of any talk about the per minute rate: a customer service agent records voices, names and complaint reasons, which means personal data on every single call. You can check both vendors side by side on our Retell AI and Vapi pages.

Which one fits your situation?

If you want to find out whether a voice agent is worth it on your actual call flow before committing money, Retell AI is the shortest path: no subscription, $10 in free credits, 20 concurrent calls included and a price list broken out by component. Its ease of use score, 3.5 out of 5, mostly reflects the fact that you have to estimate the cost piece by piece.

If your customers are in Europe and hosting is a decision criterion, Vapi is the one to look at, with its documented EU region, provided you have someone comfortable with APIs in house or at your agency. If you want a guided, end to end deployment instead, Synthflow’s enterprise contracts start at $30,000 per year with no published per minute rate, an order of magnitude that rules it out for most small businesses. Our comparison of the three platforms walks through that choice.

How do you start without putting your support line at risk?

With a deliberately narrow scope: one or two recurring requests, a dedicated number, and an automatic transfer the moment the agent leaves the script. The free credits ($10 at Retell AI, $5 at Vapi) are enough to see how the agent behaves on your own phrasing before you spend anything real.

Three things to lock down before you open the line to real customers: how long recordings are kept (configurable per agent from 1 day to 2 years at Retell AI, 14 to 180 days by plan at Vapi), the processing region, and a signed DPA. The rest, voice, model, wording, can be corrected in production. Those three cannot.