AI Assistant

The AI assistant answers natural-language questions about the devices on screen and about how the platform works. It lives in the left rail, in AI mode.

Opening the assistant

In the left rail, pick AI mode (magic wand icon). The assistant opens on the list of your conversations.

What it can do

• Analyze your devices: recent consumption, index, hourly profiles. It can search your whole fleet by name or type, not just the devices on screen.
• Answer how-to questions using the public documentation.

It only sees devices you already have access to: a question can never reveal a device outside your permissions.

The assistant's context

Above the input field, the context strip says what the assistant is working on.

• In the rail, context follows the page you're on: the visible devices are passed along automatically.
• In the full tab, context is yours: by default "Whole fleet" (the assistant finds the devices you mention on its own), or pin specific devices ("Pin" button, search by name) to focus its answers. Pins are remembered with the conversation: you'll find them again when you reopen it.

Saved conversations

Your exchanges are saved, one conversation at a time.

• Resume: on open, the assistant lists your past conversations (most recent first). The "Filter…" field below the list header searches by title across all your conversations, including archived ones and those hidden by the list settings. Click one to continue the thread.
• List settings: the settings icon on the right of the list header lets you group (by date, by organization, or not at all), sort (most recent or name), filter by status (active, archived, all) and by last activity (1, 3, 7, 30 days). These settings are remembered on your device. A conversation belongs to the organization where it started; switching organizations mid-thread does not move it.
• Archive all / delete all: at the bottom of the list settings menu, and in My account > Preferences, "My conversations" section (only affects yours).
• New conversation: the row at the top of the list starts a fresh thread. Its title fills in automatically from your first question.
• Each conversation's "…" menu: add it to favorites, open it in a new tab or a new window, rename it, archive it (it leaves the default view without being deleted: find it again via the "Archived" status) or delete it. Deletion and archiving can be undone for a few seconds ("Undo" button).

Each user only sees their own conversations.

Verify where the numbers come from

When the assistant reads your data, each lookup appears above its answer ("Recent metrics", "Active alarms"…). Numeric results render as a mini chart or a readable stats table; expand the step to see the request and the raw result: you can always check a figure at the source. These steps remain visible when you reopen a conversation.

Act on an answer

Under each answer: copy the text (or a code block via its own button), regenerate the latest answer, and rate it (thumbs). On error, a "Retry" button re-runs the question without retyping it. Under your own messages, the pencil edits and resends the question: the conversation resumes from that point.

Talk with the assistant (live voice)

Two entry points, two scopes: inside an open conversation, the composer's main button (filled circle, waveform, empty input) continues that conversation out loud: the voice assistant knows your recent written exchanges and the transcript joins the same thread. The waveform button at the top right, available from every page, starts a new voice conversation in place: the screen does not move, the button becomes the call pill, and you talk while carrying on with what you were doing: for most calls, that is all you need. Want to see the conversation? Click the duration in the pill, or open the rail's AI mode: the call shows up there with the same chrome as a written conversation (back to the list, open in a full tab with live transcript, new conversation). As soon as you type in the composer, the waveform button becomes Send again.

In voice mode: you speak, the assistant answers out loud, and it can look up your devices mid-conversation ("how much did the villa meter use yesterday?"). It can also drive the screen: "show me the villa meter" opens the device page while it comments, "switch to 30 days" changes the displayed period, "in a new tab" opens it alongside. And beyond devices: "open the Consumption dashboard" (or a database, a view, a shared location, a personal board: anything with a name: and "open Settings" opens the displayed organisation's settings without asking which one), "close this tab", "scroll to the Alarms section", "click Export"... the assistant discovers the actions available on the current page and runs them; when a name is ambiguous it asks you to pick, and it asks you to confirm out loud before any destructive step (delete, empty...). Ask it to explain the interface ("show me how to export"): it plays teacher, highlighting each element on screen while it describes it, without clicking anything; the ring stays up for as long as the explanation lasts and moves to the next element with it. The panel shows the conversation as subtle subtitles, a button to mute your microphone and one to hang up. The transcript is saved to your history like any conversation, and shows up there live. Hanging up immediately opens the written thread of the same conversation: re-read, carry on typing, or start the voice again; you can also reopen it later from the list, with the text model of your choice, which reads the whole past voice exchange.

Leaving the panel does not hang up: the call keeps going in the background while you navigate, and the top-bar voice button becomes a call pill: red dot, call duration, a microphone button (mute/unmute) and a red phone to hang up. Clicking the duration takes you back to the conversation. The red phone is the only gesture that ends the call, in the pill as in the panel; the duration also shows under the orb, which reacts to your voice while you speak.

The model selector offers the models the organisation owner enabled (Settings > AI, in the order they chose): the cost-efficient one (default), the maximum-quality one, and, when the platform offers them, "ElevenLabs natural voice" (the most natural voices on the market) and "Grok voice (xAI)" (WebSocket, works on networks that filter WebRTC). The voice timbre is selectable in the panel (ten voices, "Marin" and "Cedar" recommended: pick by ear for your language); your choice is remembered on this device. For ElevenLabs models, the same selector offers the ElevenLabs account voices (when the key allows it; otherwise the agent's configured voice applies), and the assistant switches between French and English on its own based on the language you speak. Owners can also record a custom voice (Settings > AI > Voices): a few seconds of speech, cloned via ElevenLabs after confirming consent, join this same timbre selector for the whole organization. Organizations that plugged in a custom voice model (Settings > AI: OpenAI-compatible endpoint, xAI's Grok, or an ElevenLabs agent, with a button that creates the Drexo agent automatically from the key) find it in the same selector. Owners can also cap voice minutes per month (Settings > AI, with this month's usage shown): at the cap, starting a call is refused until next month. This is a separate channel from dictation below: nothing is written into the input field.

Dictate instead of typing

The microphone button in the input dictates your question: the recognized text is appended to what you already typed. Dictation never submits on its own: review, then press Enter (or the send button). Click the microphone again to stop listening.

Two engines, picked automatically: your browser's live recognition (Chrome, Edge, Safari: text appears as you speak); and otherwise (corporate networks that block it, Firefox), a recording transcribed by the platform: the text appears at once when you stop the microphone (violet button while transcribing).

Choosing the model

At the bottom of the input, a selector picks the conversation's AI model (remembered per conversation). The list, its order and the default model for new conversations are defined by the organisation owner in Settings > Artificial intelligence: up to 10 models enabled from the catalog, with their indicative cost per million tokens. Without any configuration a low-cost default applies: every organisation always has at least one model available.

Dictation and the live voice conversation are configured the same way: each family has its own list of enabled models, reordered by drag-and-drop, with an "Auto" row (the platform default) and an on/off switch for the organisation. You can also plug in custom models (OpenAI-compatible endpoint + your organisation's token) for text, dictation or voice: they show up in the selectors like catalog models. The token stays server-side and is never visible to members.

More room: open in a full tab

Need more space for a long answer (table, list, comparison)? The open in full tab icon (at the right of the top row, in the list and in a conversation) switches to a full-width tab. It's the same thread: you pick up right where you left off.

In the full tab, the sidebar lists your conversations to the left of the exchange (the way settings list Members, Roles…): the open conversation is highlighted, clicking another one opens it in place, and "New conversation" starts one right here.