Why we built a vertical AI receptionist (and why "horizontal" is a marketing word)

I've called a lot of AI receptionists in the last six months. Most of them sound like a startup founder demoing voice AI in 2023. They're polite, they're grammatically perfect, and they have absolutely no idea what a green pool is, what zone-1 versus zone-2 means on an HVAC service ticket, or whether you'd charge differently for a residential or commercial roof.

That's the “horizontal” AI receptionist problem. One assistant, every business, every vertical. It can answer the phone but it can't do the receptionist's actual job: triage by service type, ask the qualifying questions a real office person would ask, route the urgent stuff to the owner's cell.

What we did instead

Twelve assistants. Each one with its own voice character, its own first-message script, its own qualifying-question tree, its own knowledge of the trade's jargon. Cindy answers the pool line in a tone that sounds like the office manager at a 4-truck pool service. Wyatt picks up the HVAC line with a slightly more dispatcher-y feel because that's how successful HVAC offices actually sound. Emma at the salon knows that “walk-in or chair” is a real distinction that matters to scheduling.

The technical version: each vertical gets its own assistant prompt, its own intent taxonomy in our insights pipeline, and its own list of out-of-scope phrases that route to the owner instead of trying to answer. The pool assistant doesn't try to answer questions about HVAC zoning. It says “that sounds like an HVAC question, not us — let me see if your owner has someone they'd recommend?” and writes a Slack message.

Why this isn't just “configure your own prompt”

Plenty of horizontal services let you tweak the system prompt, upload an FAQ, paste in your service hours. That's a marketing checkbox, not a real solution. Three reasons:

1. Most owners won't do it well. Writing a good system prompt is a skill. The owner of a pool service is the best person in the world at running a pool service, not at prompt engineering. We do the prompt work once, per vertical, with care.
2. The intent taxonomy matters as much as the prompt. Our voice-AI insights pipeline tags every call with structured fields — intent, urgency, outcome — that drive the dashboard. Those tags only work if the taxonomy matches the trade. A horizontal service either gives you a generic taxonomy (low signal) or asks you to define your own (low adoption).
3. The voice itself is part of the brand. A pool service whose AI receptionist sounds like a chipper Bay Area startup voice has a brand mismatch the customer hears in three seconds. We picked twelve voices from Telnyx's Ultra catalog that sound like they belong on the trade they're assigned to.

The cost of doing it this way

Twelve different production models means twelve times the prompt-tuning load, twelve times the QA work when models update, and twelve times the surface area for vertical-specific bugs. We're betting that's worth it because the alternative is owners deflecting customers with a robotic “I'm sorry, I can't help with that” on every fifth call.

Also, we're only doing twelve trades — the home-services verticals where there's an obvious booking-pattern playbook. We're not going to ship an AI receptionist for criminal defense law or hospital admissions. Those are real markets but the voice and intent surface is so different that a vertical-tuned assistant is the wrong shape entirely. Someone else should build those.

Want to hear the difference? Call the pool-service assistant at +1 (407) 598-8144. Pretend you have a green pool. See if you'd guess you're talking to AI.