Home / How I Can Help / Voice AI Development
Service · Voice and contact center
Voice AI development for real customer calls
A voice agent is not a chatbot with speech bolted on. It has about a second to respond, no screen to hide behind, and a caller who will talk over it. That changes almost every design decision.
In a chat interface a two-second pause reads as thinking. On a phone call it reads as a dropped line, and the caller starts talking again. The whole turn, from the caller finishing a word to the agent making a sound back, has to fit inside roughly a second, and that budget has to cover speech recognition, whatever reasoning you are doing, any tool call you make, and speech synthesis.
That constraint is what breaks ports of working chat agents. A retrieval step that was fine at 800 milliseconds is now most of the budget. A reasoning loop that retries is now silence. A human interrupting mid-sentence, which is normal on a phone call and rare in chat, has to cancel audio that is already playing and rewind what the agent believed it had said.
This is also the work that sits behind the phrase most contact center teams are searching for now, which is replacing an IVR with an intelligent virtual agent. An IVA is not a better phone menu. It is an agent that can hold an open-ended conversation, reach into the systems where the answer lives, and finish the job or hand it over cleanly.
Then there is the thing nobody measures until it embarrasses them. A call that ends without a human is contained. It is not necessarily resolved. Those two numbers can move in opposite directions, and if the dashboard only carries the first one, a system can look like it is working while the same customer calls back three times.
This is for you if
- → You run a contact center or customer operation and an IVR that people press zero to escape, and you are weighing an intelligent virtual agent (IVA) to replace it.
- → You have a working chat assistant and a mandate to put it on the phone.
- → You have a voice pilot that demos well and falls apart on accents, background noise or interruption.
- → You need containment and resolution measured separately, and reported honestly.
Where I am not the right fit
- × You want an outbound calling system for cold sales. Not the work I do, and not the work I want to be good at.
- × You want a voice clone of a specific person. Different problem, different ethics review, different vendor.
- × You want the cheapest possible per-minute cost as the only constraint. That decision usually gets made again six months later.
- × You have not decided who the call escalates to when the agent should stop. That is a prerequisite, not a detail.
- 01
Listen to the calls
Real recordings and transcripts before anything else. What people actually ask, where the current system loses them, which intents are worth automating and which are a trap. Most of the design falls out of this.
- 02
Set the latency budget
Write down the milliseconds available to each stage before choosing anything. Speech in, reasoning, tool calls, speech out. Most architecture decisions in voice are really budget decisions, and they are much cheaper to make on paper.
- 03
Build the conversation, not the transcript
Turn-taking, barge-in, backchannel, what happens when the caller goes quiet, and what the agent says when it does not know. Plus the boundary: which tasks it may complete, which it may only draft, and which it must hand off.
- 04
Wire the telephony and the handoff
The carrier side, the transfer path, and the context that travels with the caller so a human does not open with "can you tell me what this is about". A handoff that loses context is a worse experience than no agent at all.
- 05
Measure both numbers
Containment and resolution, separately, with repeat contacts attributed to the issue rather than the session. Plus cost per resolved issue, which is usually not the number the pilot has been reporting.
- → A working voice agent on your telephony, your systems and your data, not a sandbox demo.
- → A written latency budget per stage, and the measurements showing what the built system actually hits.
- → An evaluation set built from your own call recordings, with the answer key reviewed by a person rather than assumed.
- → A defined escalation path with context carried across the handoff.
- → Containment, resolution and cost reported separately, with the method written down so the numbers survive a challenge.
- → Knowledge transfer to your engineers, so the system does not depend on me after go-live.
Why Voice AI Agents Are Harder Than Chatbots
The constraints that make a working chat agent fail on the phone.
Read it → ArticleFrom Deflection to Resolution
Why containment and resolution are different numbers, and what the contact center market is actually buying.
Read it → ArticleLangGraph for Voice AI Agents
Where a control-flow framework helps in voice, and where its idea of an interrupt is not the same as barge-in.
Read it →Frequently asked
Quick answers
- How is a voice AI agent different from a chatbot?
- The budget and the interruption. A voice agent has roughly a second from the caller finishing a word to making a sound back, and that has to cover speech recognition, reasoning, any tool call and speech synthesis. A chat agent can take several seconds and the user reads it as thinking. On top of that, callers talk over the agent, which means cancelling audio that is already playing and correcting what the agent believed it had said. Those two constraints reshape retrieval, tool design, error handling and the conversation itself.
- Can you build on our existing contact center platform?
- Usually, yes. The voice agent sits alongside the platform rather than replacing it, and the integration work is telephony, transfer paths and passing context across the handoff. It matters a great deal which platform, so that is one of the first things we establish. If a platform genuinely cannot support what you want, you should hear that early rather than after a build.
- What is a realistic containment rate to expect?
- I am not going to give you a number here, and you should be wary of anyone who does before looking at your calls. Containment depends almost entirely on your intent mix, how much of the resolution path is automatable in your systems, and where you have decided the agent must stop. The honest answer is that we listen to your recordings first, then set a target you can defend. I would rather discuss resolution anyway, because containment on its own can rise while the customer experience gets worse.
- Do you work with a specific voice stack?
- No. I have worked across the common speech, orchestration and telephony pieces, and the right combination depends on your latency budget, your existing platform and your compliance constraints. The stack is a consequence of those, not a starting position. If you have already committed to one, that becomes a constraint to design inside.
Start a conversation
Send me twenty call recordings and we will both learn something.
A 30-minute call to start. Tell me what your callers ask for, what the current system does with it, and where it goes wrong. If there is a fit we will scope it tightly in writing.