AI Voice Agents
$0.10 per minute of talk time, all in. Autonomous voice agents as a managed product, not a do-it-yourself build stacked on rented VoIP.
What you get for $0.10 a minute
An agent that carries the call, on rails we own, billed on the only unit that matters to you: minutes of talk time.
- Autonomous conversation. The agent greets, asks, listens, handles being interrupted, and drives toward the outcome you defined for it.
- Tier 2 rails underneath. The same carrier interconnects, DIDs, and routing our human traffic runs on, not somebody else's resold VoIP with a model bolted to it.
- Outbound and inbound. Agents place calls against a campaign or answer a queue on your own numbers, under your own brand.
- Warm transfer to a human. When the call needs a person, the agent hands off with the conversation context attached instead of a cold ring.
- Simple usage pricing. $0.10 per minute of talk time. No seat license, no platform fee, no per-agent build charge.
- Production ops. Recording, transcripts, per-leg CDRs, and multi-carrier failover are already part of the network, so an agent call is as auditable as any other call.
Why the media path matters for machines
A person repairs bad audio without noticing. A recognizer does not. The network carrying an agent call has to be treated as part of the product.
- A latency budget shared three ways. Speech recognition, model inference, and speech synthesis all spend from the same budget the caller experiences as a pause. We keep the network's share of it small so the agent still gets a natural turn.
- Jitter buffers tuned for speech recognition. Buffers are tuned for recognition accuracy rather than music on hold, because a resequenced packet costs a recognizer far more than it costs a human ear.
- DTMF that arrives. RFC 2833 and SIP INFO are both handled. Digit capture fails silently when one side signals digits the other side is not listening for.
- Symmetric RTP. Predictable, symmetric media paths so turn detection is reading the caller instead of fighting the network.
Concurrency sized for fleets, not demo lines
An agent product is only interesting when hundreds of sessions run in parallel and the hundredth one sounds like the first.
Concurrency is raised as your volume grows, and the multi-carrier failover underneath human traffic sits underneath agent traffic too. A degraded carrier path costs a few seconds of setup time rather than a shift of dead air. Live route health is on the network status page.
Three steps, then it is just minutes.
There is no build phase you have to staff, and no second telephony vendor to integrate before the first call goes out.
- Define the job Tell us the outcome: qualify an inbound caller, confirm an appointment, answer the main line after hours. The boundaries and the escalation path get written down before anything dials.
- Stand up the agent We attach numbers, caller ID, routing, and the warm transfer target, then run it against live calls with recording and transcripts on so you can hear what it actually says.
- Pay per minute $0.10 per minute of talk time on the same invoice as the rest of your usage, with every leg on a CDR that carries the rate that applied.
What an agent call costs
One rate for the agent, and the published rate card for anything else you switch on.
Local outbound rates vary by destination. Rates shown are standard published pricing and may vary by destination, DID type, or account configuration. Volume, platform, and AI Voice Agent pricing is quoted. See the full rate card.
This came out of insurance calling, not a demo
We did not set out to sell a voice agent. AgentTech needed agents that could qualify inbound callers and follow up on leads at volume, on insurance traffic where a dropped turn or a mangled digit capture is a lost sale. Solving that meant owning the media path, the routing, and the records underneath it.
Once it was running on production insurance calls, the interesting part was not the model. It was that the rails around it, the codec handling, the failover, the recording, the attestation, and the CDRs, were already ours. So the agent became a product other platforms can buy at a published rate instead of assembling for themselves.
Solved Telephony owns AgentTech Dialer, and AgentTech runs on these rails as a customer. See how the stack fits together, or read the routing layer the agents plug into.
FAQs
Questions about AI Voice Agents
What does the $0.10 per minute cover?
Talk time on an agent call, all in. That is the conversation itself: speech recognition, model inference, speech synthesis, turn-taking, and the managed operation of the agent. It is billed per minute of talk time with no seat license, no platform fee, and no per-agent build charge. Standard PSTN rates apply when you use our connectivity without the agent product.
Is this a platform I build on, or a product you run?
A product we run. You define the job the agent does and where a human takes over; we stand the agent up on our own rails and operate it. There is no separate telephony vendor to integrate, no media server for you to tune, and no second invoice for the minutes underneath the agent.
Can an agent hand a call to a person?
Yes, and that is usually the point. The agent warm transfers to a live human with the conversation context attached, so the person picking up knows what was already asked and answered instead of restarting the call. The handoff target can be a queue, an extension, an external number, or a routing rule you already run with us.
Why does the media path matter more for an agent than for a person?
People repair bad audio without noticing. A recognizer does not. Latency is shared between speech recognition, inference, and speech synthesis, so anything the network adds comes straight out of the budget for a natural turn. We tune jitter buffers for speech recognition rather than music on hold, handle DTMF over both RFC 2833 and SIP INFO, and keep media paths symmetric so turn detection is reading the caller rather than fighting the network.
How many agent calls can run at once?
Capacity is sized for fleets of parallel sessions rather than a handful of demo lines, and concurrency is raised as your volume grows. The same multi-carrier failover that protects human traffic sits underneath agent traffic, so a degraded carrier path costs a few seconds of setup time rather than an afternoon of calls.
What do recordings and transcripts of agent calls cost?
The same as any other call on the network: recording at $0.0025 per minute each way, transcription at $0.05 per minute, summary at $0.005, and sentiment at $0.004. Redaction and encryption at rest are included at no additional charge, and every leg lands in a CDR with the rate that was applied.
Something else? Contact us
Hear an agent on your own call flow.
Bring the script and the handoff rule. We will run it on live rails and you pay for the minutes.