Product
AI Assistant
The assistant drafts, a human publishes. It only answers from what your knowledge base already contains, it shows where every sentence comes from, and it runs on an open model hosted in France.
It drafts, you publish
Every output is a draft: the agent reads it over, corrects it, sends it. The draft keeps its label until the first key is typed — a suggested text is never mistaken for yours.
No answer without a source
The assistant starts from your articles, your macros, and your resolved tickets. When nothing answers, it says so instead of inventing — and you know which article to write next.
An open model, hosted in France
Gemma on Scaleway infrastructure in Paris. Your conversations train no model and do not leave the Union. Prefer your own? Give us its address.
#4821 · Duplicate invoice for June · Marie Duval
Public reply
AI draftHello Marie, The June invoice appears twice because the June 3 direct debit was retried after a technical failure. The duplicate was canceled the same day; the refund goes out within 48 hours. Best regards,
Thread summary
Marie Duval reports a June invoice counted twice. The previous exchange confirmed the double debit and the cancellation is done; the refund remains to be confirmed.
Written at 14:12
Feature details
Triaged before it is read
On arrival, every ticket gets a category, a priority, and a language. The agent sees what the assistant proposed: a field filled in silently is not audited.
Picking up a thread of thirty messages
A summary on demand, timestamped, in the side panel. On demand, because a summary computed in advance on a thread that has moved inspires confidence and is wrong.
An answer, and where it comes from
The draft arrives with the articles it is drawn from, in plain view and not in a tooltip: a check outside the field of vision does not happen.
What it is allowed to read
Four switches — published articles, macros, resolved tickets, internal notes, the last of these off by default. And before every send to the model, emails, card numbers, IBANs, and tokens are masked.
It answers the customer, and says so
On the portal and the widget, the assistant can answer on its own in the languages you open. It announces itself as a machine from the first message, as the European regulation requires, and hands over as soon as it is asked to.
The cost, call by call
Every call is logged with its model, its tokens, and its cost. Quota reached, the assistant stops and offers you credits: never an invoice discovered at the end of the month.
What the log keeps from every call
{
"capability": "reply_draft",
"provider": "scaleway (fr-par)",
"model": "gemma-4-26b-a4b-it",
"actor_kind": "agent",
"actor_name": "Lea Marchand",
"ticket_id": "4821",
"input_tokens": 1834,
"output_tokens": 213,
"cost_micros": 565,
"duration_ms": 1420,
"redactions": { "email": 2, "iban": 1 },
"status": "ok"
}Other modules
Five pages for five parts of the tool
Shared Inbox
All emails, forms, and chat messages in one queue.
SLA and Rules
First response times, escalations, triggers, and macros.
Knowledge Base
Public or internal articles, customer portal, and widget.
Channels and Integrations
Email, chat, WhatsApp, API, signed webhooks.
Mobile app
Two apps for iOS and Android: the agent workspace, and the portal for your customers.