Most agencies will not tell you what is inside the box. Here is ours: six model classes, a documented role for each, a per-system breakdown of how a request flows through them, and the honest answer to the training question — we ground and guardrail models rather than fine-tune them.
Model names below describe the class and capability tier we deploy, not a fixed contractual version. Models rotate as better ones ship; the evaluation rules that govern when we rotate do not.
No system runs on a single model. Each one is a composition of these six layers, chosen so that latency, cost, and consequence are matched to the task.
Drafting, classification, qualification, and short-form replies where a response has to land in under two seconds.
Typical models: Gemini 2.5 Flash class
Most revenue-critical steps — qualify a lead, write a text-back, tag an inbound — are short, well-specified tasks. A fast model with a tight prompt beats a slow model with a vague one, and the cost per action stays low enough to run on every inbound.
Multi-step work: proposal drafting, objection handling, long-context summarisation, and anything where a wrong answer is expensive.
Typical models: Gemini 2.5 Pro / GPT-5 class
Used only where the extra thinking pays for itself. A quote follow-up that reasons about the job scope converts better than a template; a missed-call text does not need to reason at all.
Turning your documents, past jobs, price lists, and won deals into vectors so the system can retrieve the right context before it writes.
Typical models: Text embedding models, 768–3072 dimensions
This is what stops the AI inventing your prices. Retrieval grounds each response in your own approved material rather than the model's memory of the internet.
Transcribing inbound calls and voicemails in real time so a voice agent can act on what was actually said.
Typical models: Whisper-class streaming ASR
Trade accuracy against latency deliberately: a live voice agent needs streaming partial transcripts, while an after-hours voicemail can use the slower, more accurate batch pass.
Speaking the agent's side of a phone call in a consistent, named voice.
Typical models: Neural TTS with cloned or stock voices
Time-to-first-audio matters more than raw fidelity on the phone. Silence over 800ms reads as a dropped call and callers hang up.
Business logic that must never vary: budget thresholds, service areas, compliance blocks, routing, and escalation.
Typical models: Plain code, no model
Anything with a legal, financial, or safety consequence is coded, not prompted. A model decides how to phrase it; code decides whether it is allowed to happen.
The same six classes are weighted very differently depending on how your customers reach you. Phone-first trades lean on the voice stack; content-led practices lean on retrieval and reasoning. Three dots means that class is the workhorse for that industry.
| Industry | Fast general LLM | Reasoning LLM | Embedding model | Speech-to-text | Text-to-speech | Deterministic rules & scoring | Highest-leverage system |
|---|---|---|---|---|---|---|---|
| Real Estate AgentsSpeed-to-lead dominates, so the fast model carries inbound while retrieval keeps listing facts accurate. | Primary workhorse | Moderate | Primary workhorse | Light | Light | Moderate | AI Lead Generation System |
| ContractorsQuote follow-up needs reasoning over job scope; dispatch and service-area logic stays in code. | Primary workhorse | Moderate | Moderate | Moderate | Moderate | Primary workhorse | AI Quote Follow-up |
| PlumbersEmergency work is phone-first, so voice models carry the load and safety keywords bypass the model entirely. | Primary workhorse | Light | Light | Primary workhorse | Primary workhorse | Primary workhorse | AI Missed-Call Text-Back |
| Coaches & ConsultantsAuthority content and nurture sequences reward the reasoning model and heavy retrieval over your own material. | Moderate | Primary workhorse | Primary workhorse | Light | Not used | Light | AI Content Engine |
| Airbnb HostsHigh-volume, low-variance guest messaging: fast model plus strict rules around check-in and access codes. | Primary workhorse | Light | Moderate | Not used | Not used | Primary workhorse | AI Automation System |
| DentistsScheduling and recall drive value; clinical questions are hard-blocked and routed to staff. | Primary workhorse | Light | Moderate | Moderate | Moderate | Primary workhorse | AI Appointment Setter |
| ChiropractorsSame shape as dental: booking and no-show recovery, with clinical advice out of scope by rule. | Primary workhorse | Light | Moderate | Moderate | Moderate | Primary workhorse | AI Appointment Setter |
| HVACSeasonal call spikes make the voice stack the highest-leverage layer; maintenance plans use retrieval. | Primary workhorse | Moderate | Moderate | Primary workhorse | Primary workhorse | Primary workhorse | AI Receptionist |
| RoofersLong sales cycles and large tickets justify the reasoning model on every follow-up. | Primary workhorse | Primary workhorse | Moderate | Moderate | Light | Moderate | AI Quote Follow-up |
| RestaurantsPhone orders and reviews dominate; negative reviews always route to a human before posting. | Primary workhorse | Light | Light | Primary workhorse | Primary workhorse | Moderate | AI Review Responder |
Weightings reflect where the work actually lands in our builds for each sector. They are starting points for scoping, not guarantees — the mix is confirmed against your own call, form, and CRM volumes during onboarding. See methodology and metric definitions for how we model the numbers around them.
First 50 replies are reviewed and approved before the system runs unattended.
Every long-form piece is edited by a human before publication.
Any deal above your set value threshold requires rep approval to send.
Weekly audit of the exception queue.
No-show and reschedule exceptions surface to your inbox.
Live barge-in: a human can take the call at any point.
Sequences are approved before the first send of each campaign.
You confirm stage changes on deals above your threshold.
Escalation queue reviewed daily during onboarding.
All 1–3 star responses require sign-off.
Price changes always require a human-issued revised quote.
Conversations hand off to a human once a booking intent is detected.
Plan review before each cycle.
This is the most common misconception about systems like ours, so it is worth stating plainly: we do not retrain a large language model on your data, and we do not need to. Fine-tuning is slow, expensive, and locks you into a model generation that will be obsolete within a year. Every improvement below happens outside the model weights.
Before any model writes a word, we retrieve the relevant facts from material you approved — price lists, service areas, past jobs, policies, FAQs, and won deals. The model composes language around retrieved facts rather than recalling them.
Each system runs a versioned instruction set plus a small library of real, approved examples of the outcome you want. Ten good examples of your best reply outperform a paragraph of adjectives describing your tone.
Anything with a financial, legal, or safety consequence is enforced in code that the model cannot override. The model chooses phrasing; code chooses permission.
Every build runs supervised before it runs unattended. Responses are reviewed, corrections are captured as new exemplars, and the guardrails are tightened around the failures we actually see rather than the ones we imagined.
Each system carries a held-out set of real cases with known-good outcomes. When a new model generation ships, we run that set against it and only switch when it wins on both quality and latency.
The same disclosure discipline applies to every number we publish. Formulas, assumptions, and cited research live on the methodology page, and the terms used here are defined in the AI systems glossary.
Pick your industry, read the model mix, and check the arithmetic before you buy anything.
