Why we ship Voice AI in 4-6 weeks, not 4-6 months.

Every Voice AI proposal I read in 2026 says the same three things. Real-time turn-taking. Warm human hand-off. Domain-specific RAG. Then the timeline says four to six months. That gap — between what everyone claims and how long it takes to ship — is where most Voice AI projects die.

Here is what we do differently.

1. Scope the win, not the surface area.

Week one is not "gather requirements". Week one is which single conversation will we prove works. One caller intent, one domain, one hand-off condition. If you cannot state the win in a sentence, you have not scoped it. If your scoping document is longer than one page, you have not scoped it.

The fastest ship-teams write the smallest scoping documents.

2. Build with an eval loop on day two.

Every voice agent regresses. Latency creeps. The model gets fooled by a new phrasing. The RAG index goes stale. If you do not have an evaluation loop wired in on the second day of the build — before you have even hit the happy path — you will spend weeks four and five debugging things that regressed in week three.

We use a very small eval suite. Twenty conversations, hand-labelled. Every commit runs against them. If the pass rate drops, the commit does not merge. That is the whole discipline.

3. Latency budget as a first-class number.

Turn-taking under 350ms sounds like an engineering goal. It is really a UX decision. The caller decides in the first turn whether they are talking to a human. If the agent hesitates once, they treat every subsequent turn as suspect. So we budget latency the same way we budget cost: end-to-end, per hop, with a hard ceiling.

• ASR — 80ms budget
• LLM inference — 180ms budget
• TTS — 90ms budget
• Network + jitter — the rest

If any hop breaks its budget, we do not tune it. We route around it. That means the ASR is Speechmatics or Deepgram, the model is a fast one from Anthropic or a distilled Llama, and the TTS is ElevenLabs Turbo or Cartesia. Choose fast, then optimise.

4. Warm hand-off is a design decision, not a fallback.

Most Voice AI projects treat human hand-off as the failure case. We treat it as the trust case. The agent tells the caller when it is uncertain. The caller stays in control. The escalation is one line: "I am going to hand you to a human colleague who can help with this, one moment."

The result is not just fewer bad calls. It is more callers who trust the agent on the calls it does handle.

5. Ship at the end of week 4, not week 24.

By the end of week four we have a working agent handling one conversation at production quality, with the eval loop green, the latency budget hit, and the hand-off condition tested. Weeks five and six are for the second conversation, the second hand-off, and the runbook.

Then we hand over. Your team owns the code, the evals, the prompts, the runbook. We stay on retainer if you want us to; if not, everything transfers.

The trade.

You get a working Voice AI agent in production in six weeks. In exchange, you agree to scope down. One conversation, not five. One intent, not fifteen. One hand-off, not a decision tree.

If the answer to that trade is yes, the rest is engineering.