Frequently Asked Questions

Question 1

How much does it cost to build a custom AI voice agent?

Accepted Answer

As a rough guide rather than a quote: a single-purpose agent on an existing platform is typically $4,000 to $12,000, and a custom build on LiveKit or Pipecat with telephony, tool calling and a dashboard usually runs $15,000 to $45,000. Running costs are separate and depend heavily on which speech and voice providers you pick. I quote after a call, because the deciding factor is almost always how many of your systems the agent has to reach rather than the agent itself.

Question 2

Should I build custom or just use Vapi or Retell?

Accepted Answer

Start on Vapi or Retell. They will get a working agent live in days and for most single-use-case agents they are the right long-term answer too. Custom starts paying for itself when one of four things is true: per-minute cost stops making sense at your volume, you need latency the platform cannot reach, you need a model or voice it does not offer, or you have data-handling requirements a shared platform cannot meet. If none of those apply, I will tell you to stay where you are.

Question 3

What latency can you actually achieve?

Accepted Answer

Industry-typical figures, not a promise: speech-to-speech over WebRTC usually lands first audio somewhere around 0.5 to 0.9 seconds, a chained speech-to-text, model and speech synthesis pipeline around 0.8 to 1.4 seconds depending on providers, and a phone call adds a little on top of an in-app connection. What your pipeline actually achieves depends on your providers, regions and conversation design, so I measure it stage by stage and give you real numbers rather than quoting a headline figure up front.

Question 4

LiveKit, Pipecat or the OpenAI Realtime API: which should I use?

Accepted Answer

LiveKit if you need production telephony, scaling and SIP handled by a mature framework. Pipecat if you need to insert custom logic between stages, such as redaction, moderation or routing. The OpenAI Realtime API directly if latency and voice naturalness matter more than model choice and per-layer control. Most production systems I build end up using LiveKit for transport with a chained pipeline inside it, because it keeps a clean transcript at every turn.

Question 5

Can you put an AI agent on a real phone number?

Accepted Answer

Yes. Either through Twilio Media Streams or a direct SIP trunk into LiveKit, covering inbound and outbound calling, warm transfer to a human with context carried across, voicemail detection, and call recording with a retention policy. Twilio is faster to launch; direct SIP is meaningfully cheaper per minute once you are past a few thousand minutes a month.

Question 6

How long does a voice agent project take?

Accepted Answer

A working prototype on your own data takes one to two weeks. A production agent with telephony, tool calling and monitoring is typically six to ten weeks. A migration off Vapi or Retell is usually four to eight weeks including a side-by-side comparison and a gradual cutover.

Question 7

Can you take over an existing voice agent someone else built?

Accepted Answer

Yes, and a good share of my work is exactly that. I start with an audit: per-turn latency, real cost per minute, and where calls actually drop off. You get that written up with a recommendation before any code changes, including the recommendation to leave it alone if that is the honest answer.

Question 8

Do you handle HIPAA or other regulated environments?

Accepted Answer

I have built safety-critical AI for a regulated healthcare practice, working to a hard split between a deterministic, auditable safety layer with no language model in its decision path, and a support layer where a model earns its place with human review. For voice specifically that means BAA-covered vendors, a retention model designed rather than defaulted, and infrastructure owned by your organisation from day one. I am an engineer, not a compliance advisor, so your counsel signs off on the regime and I build to it.

Question 9

Who owns the code and the accounts?

Accepted Answer

You do. All code, infrastructure and vendor accounts are yours, set up in your organisation from the start rather than migrated across at the end. Ending an engagement means revoking my access, not untangling credentials.

Question 10

How does the time zone difference work?

Accepted Answer

I am in Ho Chi Minh City (GMT+7) and work async-first: written updates, architecture documents, and clearly recorded decisions. I keep overlap hours for calls with US and European teams, and typical response time is under four hours.

Tell me what you are building

Tell me what you are building

Questions I get asked most