AI voice agents
AI voice agent development: phone calls answered in natural speech, with a person on standby
A voice agent lives or dies on timing and on what it does when it does not understand. We design for both, and we are cautious about outbound calling.
In short
An AI voice agent answers or places phone calls using speech recognition, a language model and a synthetic voice. It can book appointments, answer common questions, qualify leads and update your CRM, then transfer to a person when needed. We build voice agents with call recording consent, tested latency, transcripts and a clear fallback path.
Why latency decides whether callers stay
People notice pauses in conversation quickly. A voice agent chains several steps, and each adds delay. Here is the budget we design around. The numbers are illustrative ranges, and we measure real figures on your setup.
| Step | What it does | Typical delay to plan for |
|---|---|---|
| Detect end of speech | Decide the caller has finished | 200 to 500 ms |
| Speech to text | Turn audio into words | 100 to 300 ms |
| Language model | Decide what to say, call tools | 300 to 1,000 ms |
| Text to speech | Start producing audio | 100 to 300 ms |
| Network and telephony | Carry the audio both ways | 50 to 200 ms |
Added up, the total is often between about one and two seconds, which feels slow on the phone. We cut it by streaming each stage, choosing faster models for simple turns, keeping prompts short and placing servers near your callers. We test with real calls, not demos.
Where voice agents work, and where they should not
- Good fit: after-hours answering, appointment booking and rescheduling, order status, simple FAQs, lead qualification with a handful of questions.
- Careful fit: payments, medical or legal triage, anything with strong accents or noisy lines unless tested.
- Poor fit: complaints from upset callers, complex sales calls, anything where trust is the product.
- Not for us: cold-calling lists of people who never agreed to be called.
If a text channel would do, a chatbot is cheaper and easier to test. Voice is worth it when your customers already phone, and calls are missed or queue.
Consent, disclosure and local rules
Phone rules differ by country, and we are not your lawyers. Your counsel should review the call flow. These are the points we raise in the first week.
| Topic | What we plan for |
|---|---|
| Telling callers it is an AI | The agent says so in the greeting |
| Recording and transcripts | Consent notice at the start. Some US states and other countries require all parties to agree to recording |
| Outbound calls in the US | The FCC has ruled that AI-generated voices count as artificial voices under the TCPA, so outbound calls need prior consent. Confirm with counsel |
| UK and EU | GDPR or UK GDPR for voice data and transcripts, and ePrivacy or PECR rules for marketing calls |
| Australia | Privacy Act duties for call data, and the Do Not Call Register for marketing calls |
| Retention | A set period, then deletion, and no use of calls to train models unless agreed |
Want AI voice agents done for you?A short call, a written plan and a fixed quote.
Get a proposalOur process
- Call analysisWe listen to a sample of recent calls, with consent, and write down the five to ten reasons people call.
- Call flow designGreeting, questions, tool actions, and the exact words for each transfer or refusal.
- PrototypeA test number you can ring yourself. Your team tries awkward callers.
- Connect systemsCalendar, CRM, booking or order systems, and your phone provider.
- Shadow launchThe agent handles a share of calls, with transfers to a person and daily transcript review.
- Widen and tuneMore hours, more call types, as the review numbers allow.
What you get and how long it takes
- A voice agent on a phone number or SIP trunk you own.
- Call flows written in plain language, so your team can read and approve them.
- Transcripts and call outcomes saved to your CRM.
- A warm transfer to a person with a short summary of the call.
- A data and consent note for your privacy review.
A focused agent for one call type often reaches shadow launch within a few weeks. Several call types, multiple languages or complex systems take longer. The quote has dates for your scope, available through the contact page. For definitions, see what is an AI voice agent.
How we report and related work
During the shadow launch you receive a weekly note with calls handled, calls transferred, the main reasons for transfers, and a sample of calls to listen to. Later, the same figures come from the call logs we set up.
We have no published voice agent case study, and no real project in our portfolio that matches closely. We would rather say that than borrow a result. Our cost page for AI voice agents explains what drives the price.
Updated 30 October 2026
Questions, answered
Can the voice agent sound like a person?
Which languages and accents does it handle?
What happens if the caller wants a person?
Can it make outbound calls?
Can it use our existing phone system?
Keep exploring
How much does an AI voice agent cost in 2026?
A voice agent is billed by the minute as well as built once. Price it in two parts: the build, and what every minute of talk costs to run.
ServiceAI chatbot development that answers from your content and knows when to stop
A chatbot is judged by the questions it should not answer as much as the ones it should. We design the boundaries first, then build the bot inside…
Tell us why your callers phone
Describe your top five call reasons on the contact page and we will say which a voice agent can take.