How it works · Architecture & data flow

Transparent by design — see exactly what happens

What happens when someone talks to a daita agent, the technology behind each step, and where your data lives. No black box.

The data flow

From a question to a completed task

A streaming agentic pipeline — grounded in your knowledge and catalog, bounded by guardrails, observable end to end.

  1. 1
    A visitor reaches outWebsite chat (text or voice) and email — plus phone and SMS on the Growth plan (US and Canadian numbers; European numbers coming soon) — one agent with one shared memory.
  2. 2
    Verified, encrypted ingressThe request enters daita over encrypted channels; phone & SMS webhooks are cryptographically verified before anything runs.
  3. 3
    Claude reasons and plansAnthropic Claude reasons through the task in a streaming agentic loop and decides which knowledge and tools it needs.
  4. 4
    Grounds in your knowledge and catalogTwo-stage retrieval finds candidates, then a neural reranker (EU) orders them by true relevance. Answers are cited, never invented.
  5. 5
    Acts with real toolsSearches your catalog, builds a ready-to-pay checkout, captures leads and, on Shopify, looks up orders and files return requests — real actions, not just talk.
  6. 6
    Responds in text or neural voiceReplies in text and can read its answers aloud — in the premium natural voice on Growth — and can even navigate the page for the visitor.
  7. 7
    Records to an EU database, escalates when neededEvery interaction is written to an EU database with a timestamped audit log; when human judgement is needed, it hands off with the full conversation.
Conversation data is stored in the EU; the app servers, the AI model, voice transcription and file attachments currently run in the US under EU Standard Contractual Clauses · every agent action is timestamped, exportable on request · an EU-region or on-premises deployment can be scoped on Sovereign
The technology

The stack inside, by layer

Best-in-class components, each with a clear role — and we say where each one runs.

US · EU-region / on-prem can be scoped on Sovereign

Anthropic Claude

Reasoning, language understanding, tool use (agentic loop)

EU

BM25 + neural reranker (self-hosted, EU)

Two-stage retrieval over your knowledge, with citations

US · EU-region / on-prem can be scoped on Sovereign

ElevenLabs speech-to-text & neural voice

Transcribes the voice messages shoppers send (website mic and SMS voice notes) and speaks natural replies — the default voice

US

OpenAI gpt-4o-mini-transcribe

Transcription fallback when ElevenLabs fails — part of the service, not optional

Opt-in · may leave the EU

OpenAI gpt-realtime

Live speech-to-speech voice — optional, owner-enabled

US · EU-region / on-prem can be scoped on Sovereign

Twilio · Web · Shopify

Phone & SMS transport, embeddable web widget & Shopify app

EU

Firestore (eur3, Europe)

Leads, bookings, transcripts and timestamped audit logs — leads, bookings and inbox tickets are kept until you delete them, and every web chat is copied in full to a ticket in your inbox (messages, attachments, the customer name and contact details) that stays until you or the customer have it erased; only the conversation records the agent itself uses, shopper profiles and analytics events auto-expire after 90 days

US · EU-region / on-prem can be scoped on Sovereign

Resend (SPF/DKIM/DMARC)

Branded, fully deliverable transactional email and confirmations

EU

Self-hosted observability

Latency, tool use and errors — your data, no third-party analytics

Where your data lives

Where your data lives

Built in Munich. The database is in the EU (Firestore eur3), and the embedding and reranking models behind product and knowledge search run in the EU (europe-west1). Four parts of the core service are not in the EU today: the application servers that every request passes through run on Google Cloud in the US (Cloud Functions and Cloud Run, us-central1); the language model that writes the replies (Anthropic Claude) runs with a US provider under EU Standard Contractual Clauses — an EU path on Claude via Vertex AI (europe-west1) is built but not yet switched on; the voice messages a shopper sends — through the website microphone or as an SMS voice note — are transcribed in the US by ElevenLabs, with OpenAI as the fallback; and customer-uploaded attachments are held in a US storage bucket, with an EU migration planned. None of the four is optional, which is exactly why we name them here. They are not the only US providers: transactional email goes out through Resend, a US provider, and phone and SMS run through Twilio, also in the US, when you switch those channels on — both are in our sub-processor list. The extras that add further non-EU routing — live speech-to-speech voice, for example — stay opt-in and clearly flagged.

Database in the EU (Firestore eur3)
Semantic search in the EU (europe-west1)
The language model runs with a US provider under EU Standard Contractual Clauses — an EU path on Claude via Vertex AI is built but not yet switched on
A data processing agreement (Art. 28 GDPR) on request, on every plan; the assistant always says it is an AI (EU AI Act, Art. 50); timestamped audit trail, exportable on request — every web chat is copied in full to a ticket in your inbox and kept until you or the customer have it erased, leads and bookings are kept until you delete them, and the conversation records the agent itself uses, shopper profiles and analytics events auto-expire after 90 days
Exceptions today: the application servers run in the US (Cloud Functions and Cloud Run, us-central1), voice messages are transcribed in the US (ElevenLabs, with OpenAI as the fallback) and customer-uploaded attachments are held in a US bucket (EU migration planned); transactional email (Resend) and, when you enable them, phone and SMS (Twilio) also go through US providers — see Trust.
daita · München

Want this for your business?

Talk to a live agent on this site right now.