The AI panel

The AI panel is the everyday way to talk to the assistant. It opens from the AI bubble in the top bar of the backend and of the website preview, from AI ‣ Conversations, or from the command palette (see AI search and command palette). When you open it on a record, the assistant sees that record and offers quick prompts that fit it.

Everything the assistant reads or changes goes through the same access gate as your own Odoo rights: it can never see or do more than you can, and an administrator can narrow it further (see Who can use what).

Quick prompts

When a record form is open, a row of small buttons appears above the message box: the quick prompts. One click sends the stored instruction, for example “Summarise” or “Next best action”. Hover a button to read the instruction it sends. Prompts made for the current model come first; generic prompts come last, and a specific prompt can replace a generic one.

Administrators manage them under AI ‣ Configuration ‣ Behavior ‣ Quick Prompts: name, instruction, the model they belong to, whether they need a single open record, the groups that see them and the generic prompt they replace. Many app modules ship their own prompts (see AI in the business apps); you can edit or archive them.

Conversation options

The sliders button (Conversation options) opens a small drawer. Your choices are remembered for the next conversations.

  • Think longer asks the model for extended reasoning. Answers are slower and cost more, but hard questions get better results. It only has an effect on models flagged as reasoning-capable (see Providers and models).

  • Show the steps the assistant takes adds a collapsed “N steps” line under each answer. Open it to see which tools were called, whether each call was allowed and how long it took. Use it to understand or check an answer.

  • Changes to data: chooses how this conversation handles changes. Follow the AI settings uses the global write mode; Always ask me first turns every change into a proposal you confirm; Apply without asking skips the confirmation step. The last choice only appears when an administrator has enabled it, and rules that force confirmation still win.

Confirming changes

When the assistant wants to create, update or delete something and confirmation is required, it does not change the data. A card with a plain-language summary appears under Pending changes. Click Confirm to apply the change in your name, or Cancel to drop it. All open proposals are also listed under AI ‣ Write Proposals, and a proposal that nobody answers expires after the configured lifetime.

The Records created or updated button lists the records the assistant touched in this conversation, with links.

What the assistant can show

  • Record cards: for a list of records (up to 20) it shows compact clickable cards with the name and a few key fields instead of a long text list.

  • Pictures: ask for an image and, if an image model is configured, the result appears under the answer and is stored with the conversation.

  • Navigation: it can open the list, pivot or graph you ask for, with the filters, grouping and measures it understood, and it can change the search of the view you are looking at (add or remove filters, grouping, ordering).

  • Dates: “last quarter” or “three working days from now” are computed with your time zone and company calendar instead of being guessed.

  • Menus: it can tell you where in the menu tree a feature lives, limited to the menus you can see.

Files

Use Attach files to add images and documents to your message. The assistant can read the text of a PDF or office document you are allowed to read, and look at pictures. Open Conversation files to see everything attached to the conversation. Allowed types and the maximum size are set in the AI settings.

When something goes wrong

If a tool the assistant calls fails, the assistant is told why, so that it can correct itself: when it used a field that does not exist, for example, it sees the message of the error and tries again with a valid one. The message is shortened to its first lines and never contains technical details such as SQL statements, tracebacks or server paths. The exception is a refusal by the access rules: in general the assistant only learns that access was denied, not why, so that it cannot be used to find out which records or fields exist. Changes made by the failed tool are rolled back.

The assistant answers in plain words and in the language of your message. It uses the labels you see in your interface and does not name its internal tools, not even when it explains what it cannot do.

Delegating to specialist agents

An agent can hand a sub-task to a specialist agent, and the everyday assistant can consult agents that an administrator made available to it (“ask the accounting agent”). The sub-task runs as you, in its own conversation, and the specialist can use only the capabilities that both agents and you have: delegation never widens rights. You see the sub-task in the panel with a Back link to return to the main conversation. Delegation goes at most two levels deep and cannot loop.

Administrators set this up on the agent form: Can delegate to lists the specialists an agent may use, and Available to the assistant makes an agent consultable from the everyday chat. Only agents without their own user can be delegates.