The AI Facilitator and setup plans

Every workspace has an AI Facilitator: a built-in agent that runs LEMA Chat itself for the people in it. It answers how-to questions, sets things up for you, and takes general requests that no other agent clearly owns.

What it does

• How-to questions: "how do I...", "where is...", "what is...". It answers from this Handbook and cites the section. If the Handbook does not cover something, it says so.
• Setting things up: Topics, interviews and checklists for people, courses, onboarding a person, links and support tickets.
• People, invites and access: it is the only agent that may propose adding someone to a group, inviting someone, assigning a checklist to people, or widening who can see something. Any other agent asked to do these refuses and suggests handing off to the AI Facilitator.
• Undoing a setup it made, and saying what cannot be undone.

Managing the workspace for you

Ask it in the chat and it can do what the app's buttons do:

• Agents: create, edit, delete and restore an agent.
• Invites: list pending invites, re-send one or cancel one.
• Members: change a member, or remove one from the workspace.
• Topics: create, edit, delete and restore a Topic. A new Topic is top level unless you name a group.
• Interviews: create one, change its Setup, add checklist items, add participants (someone not yet in the workspace is sent an invite by email), remove a participant, or reset a person's session.

Each one runs through the same route as the button, as you: it can only do what you may do yourself, and each change is audited with the agent and the person it acted for. Other agents can do this only when an admin turns on Can manage this workspace for them (see Agents and models); asked without it, an agent says it can't and points you to the AI Facilitator.

You do not have to mention it by name. The Dispatcher sends setup and how-to requests to it. In a solo room it also picks up any request that expects a reply and that no other agent took; the reason is shown as Routed to AI Facilitator: general task.

It acts with your permissions and never more. If a step needs rights you do not have, it writes the step as "ask an admin to ..." instead of doing it.

Setup plans

For anything that touches people, invites or access, the AI Facilitator puts the steps on one plan and asks you to approve it. The plan appears in the chat as an approval card:

• Each step is a line with a checkbox. Optional lines start unticked.
• Lines that invite someone or add them to a group are marked (sends an email or grants access now). Only the person who asked can approve those lines; for anyone else they are locked.
• Ask an admin to: lists steps you cannot do yourself.
• Assumed: lists the defaults it filled in for you.
• Press Approve N of M to run the ticked lines, or Reject.

Approved lines run in order, as you. The run stops at the first failure, and the card then shows each line as Done, Failed, Skipped or Not selected.

A plan can include:

• inviting a person (they join as a member),
• adding a member to a group,
• assigning a checklist, or single items on it, to people, with due dates (you must be a Facilitator or above, and everyone named must already be able to see the chat),
• creating a chat,
• scheduling a reminder in this chat or in a chat the plan creates.

A later line can use something an earlier line creates, for example a reminder in the chat the plan just made. If you untick a line that a later line depends on, the plan is refused before anything runs.

Safety rules

• Plans that touch people are never approved while someone is masquerading as you.
• If the agent read untrusted content in that turn, it cannot propose people steps.
• Agents may send at most 5 invites per plan and 20 per workspace per day by default (platform settings). Invites you send yourself from Settings are not counted.

Undo

Ask the AI Facilitator to undo a setup it made. Undo reverses what the plan created, newest first, once. It cannot unsend an email, reverse an invite someone already accepted, or un-read history someone already saw. If people have edited an assigned checklist since, their edits are kept.

For admins

The workspace setting AI Facilitator setup actions lists which risky actions the AI Facilitator may propose: add_group_member, invite_member, grant_access, open_access, assign_checklist, interview_access, assign_role and manage_workspace (all eight by default). An admin may remove actions but never add new ones. A removed action is refused, and the AI Facilitator asks an admin to do it instead.

Whispers from people

Workspace facilitators, and platform staff when supporting the workspace, can see conversations in groups they facilitate and may privately whisper to the AI agents in them. The AI Facilitator follows those whispers alongside its own instructions, and may tell you that facilitators can whisper to it.