Why we shipped a working AI voice agent in 36 hours (and why most agencies can't)
We built a full-stack inbound voice agent for an emergency trades client, qualifying calls and routing to dispatch, in 36 hours of focused work. The reason most agencies can't do this isn't time. It's that they're not engineers.
An emergency plumbing client called us with a problem most trades operators recognize. After-hours calls were going to voicemail, weekend leads were sitting until Monday, and the cost of a missed emergency job was somewhere around $500 to $2,500 a piece. The owner asked if there was a way to have a human answer 24/7. We pitched an AI voice agent instead.
36 hours later it was answering live calls. Here's what made that timeline possible.
The architecture, kept boring on purpose
- A phone provider (Twilio) with a webhook into our app.
- An LLM with a tightly written system prompt and a tool-calling schema for structured intake (address, emergency type, urgency, preferred callback method).
- A streaming TTS for the voice itself (we benchmarked several; the best-sounding option also happened to have the lowest latency).
- A dispatch router that hits the on-call technician's phone with the structured intake the moment the AI hands off.
- A logging layer that records every call, every transcript, every routing decision.
No vector database. No fine-tuned model. No multi-agent orchestration. The boring stack works.
Why this isn't a generic chatbot
A general-purpose chatbot wouldn't have closed any of these jobs. The system prompt was 600 words long and specific. It knew the licensed service area, the emergency categories, the language to use with a panicked homeowner, and the language to NOT use (it doesn't quote a price; it gathers enough information to dispatch).
Most of the engineering value isn't the AI. It's the structured fields we extract, the dispatch rules we encode, and the integration into the team's existing job-management workflow.
Why most agencies can't do this in 36 hours
Not because the LLM tech is hard. Because the work is half engineering and half operations design, and most agencies are 100% one or the other.
- A marketing agency can write the script but can't ship the dispatch routing.
- A dev shop can ship the dispatch routing but can't write a script a panicked customer trusts on the phone.
- A consultancy can do the workflow design but won't write code.
When the three skills live on different teams, integration alone takes 3 weeks. When they're on the same team, the architecture work and the script work happen in parallel and the integration work disappears.
The post-launch reality
The voice agent is now answering most after-hours calls. Some get routed straight to dispatch; some get escalated to a human; some get politely declined (we won't take a call from out-of-area, and the agent knows that). Total runtime cost is cents per call, which is the kind of cost structure that turns previously-impossible 24/7 coverage into table stakes for a small operator.
Full engineering write-up is on the Emergency Plumber MA case study. If you want to talk through whether a voice agent makes sense for your business, book a free growth audit.
Frequently asked questions
- How is custom AI voice agent pricing structured?
- We quote scope-based, fixed-price after a strategy call. Runtime cost is LLM tokens and phone provider fees, typically pennies per call. ROI usually lands fast for any business where a single missed call costs more than a fraction of a typical job. We'll size the engagement for your business on a free growth audit.
- Can the AI voice agent handle complex conversations?
- It can handle a lot more than people assume, but the win is in scope. The 36-hour build above doesn't try to handle billing, complex appointment scheduling, or sentiment-based de-escalation. It does one thing (intake and dispatch) extremely well.
- Will customers know they're talking to AI?
- We don't try to hide it. The current generation of TTS sounds good but not perfect, and the agent introduces itself accurately. We've found that being honest about the AI plus delivering value (24/7 availability, fast response, accurate triage) builds more trust than the cheap deception of pretending.
