AI overview Planned

Our AI is open-weights end to end. No closed proprietary frontier model in the default path. Inference, embeddings, retrieval, and voice synthesis all run on infrastructure we control, so the promise that customer data stays on our infrastructure actually holds.

The stack

Every component is chosen so we can self-host it when scale economics demand it. Phase A serves through hosted endpoints we have vetted contractually (no-training, no-logging) to validate; Phase C runs entirely on GPUs we own.

Inference

An open-weight large language model running on hardware we own and operate. Strong on tool-calling, grounding, and instruction following. No closed proprietary model sits on the default inference path — the data-sovereignty promise requires it.

Embeddings

Self-hosted from day one. All embedding compute runs on our own infrastructure, so no customer content is sent to a third-party model for retrieval. We use an open-weight embeddings model sized to run cheaply on commodity GPUs.

Vector store

Postgres with vector search. One database for transactional state and retrieval — no separate vector DB to operate. Row-level security enforces per-workspace isolation in the same query, so there are no separate tenant boundaries to police.

Voice TTS and STT

TTS. Neural text-to-speech with voice cloning from a fifteen-second sample. We ship the cloned brand voice as a free feature; it is a meaningful differentiator versus closed-stack vendors. See voice cloning.

STT. Streaming transcription with sub-second latency and strong noun handling, which matters on customer-support calls full of brand names and order numbers.

Product surfaces

The substrate powers four customer-visible features. They share the same retrieval and tool-calling primitives.

• Knowledge base. Your site crawled, chunked, and indexed in our vector store; the source of grounded answers.
• Smart Hold. AI engages callers who are waiting in the phone queue.
• AI Receptionist. Opt-in press-1 flow that lets AI handle a full call.
• Agent assist. Drafts, classifications, suggested replies inside the dashboard.

Grounding

The model never answers from parametric knowledge alone. Every reply is grounded in retrieved context: KB chunks, conversation history, visitor attributes, and (where configured) live tool calls into customer systems like Shopify Admin or your own order database. See custom tools for the function-calling surface.

Billing: AI engagements

One AI engagement is one AI interaction across any channel: a chat where AI drafted at least one reply, a Smart Hold session that engaged the caller, an AI Receptionist call, or an agent-assist suggestion an agent accepted. Starter plans include 100 per month, overage at $0.10 each. See AI engagements API.

What leaves your workspace

We are explicit about what crosses the boundary at every step. The full list of third-party sub-processors and what they receive lives on the sub-processors page.

• Embeddings. Never leave our infrastructure. All embedding compute runs on GPUs we own.
• Vectors. Stored in our own database, never sent to a third-party vector store.
• Inference (during validation). Prompts and completions transit a hosted endpoint we have vetted contractually with no-training, no-logging terms.
• Inference (steady state). Stays entirely on GPUs we own and operate.
• TTS. Voice prompts transit a third-party synthesis provider; output audio is cached on our CDN.
• STT. Audio chunks stream to a third-party transcription provider for transcription only; we retain the resulting text.

Common pitfalls

• Do not point the AI at unauthenticated public endpoints as a tool. AI inherits the workspace's permission scope, but tools that hit your systems must use auth you control.
• KB freshness matters. If your help centre changes weekly, schedule recrawls weekly. The default recrawl cadence is monthly.
• Smart Hold is opt-in per phone number, not workspace-global. New DIDs start with it disabled.