Ask the assistant

The assistant is one conversation that follows you across the product. It answers questions from your own data, read under your own permissions, and it can make changes — but only ones you confirm on screen first.

What this does

Before the assistant, every AI entry point owned its own private conversation: what you asked on a program had nothing to do with what you asked on an assessment. Now they all hand their question to the same door, pre-scoped. One question, one conversation, one history — whether you asked it from a program, from the search palette, from your home page, or from the assistant’s own page.

Three rules hold everywhere:

It says what it’s looking at. Every conversation shows a scope line — “Looking at: Everything in Meridian Academy”, “Looking at: The program ‘Onboarding’” — so an answer is never silently about something else.
It shows where a number came from. Counted figures are clickable, and it refuses to state a figure the evidence it read doesn’t support. If it could only read part of the data, it says so rather than rounding the gap away.
It never changes anything on its own. A change arrives as a proposal on a card. You press Confirm. Riskier ones ask twice; a few need a second person entirely.

It mostly waits to be asked. Where it does offer something unprompted — a chip, or at most one banner a day — that’s bounded and switchable off; see Suggestions and interruptions.

Before you start

You need permission to use AI in the organization you’re working in. Without it the Ask button renders disabled with “AI features are turned off for this organisation.”
AI must be switched on for your organization, and your organization must be in the assistant rollout.
Nothing else. The assistant reads under your own permissions — it can’t show you anything you couldn’t open yourself.

Open it

There are five ways in, and they all land in the same conversation.

Where	How
Any staff page	The Ask button in the header, beside the notification bell
Any staff page	Press ⌘J (Ctrl+J on Windows)
The search palette	Press ⌘K, then choose Ask the assistant
Your home page	Type straight into the composer under the greeting
A program, an assessment, a document	Ask AI about this program, Ask AI, Ask this document

⌘J opens the assistant; ⌘K opens search. They’re deliberately different keys — one is a place to type a question, the other a place to jump. ⌘J puts your cursor in the composer, and ⌘K won’t steal focus from a field you’re already typing in.

The assistant shares one overlay slot with the help panel, so opening one closes the other and closing it returns your focus to whatever you were on.

The panel, and the full page

Opening from anywhere gives you a panel on the right of the current page, with two tabs: Conversation and Recent changes. It’s scoped to the page you opened it from.

Open full page expands the same conversation to /dashboard/assistant — three panes:

Your conversations — everything you’ve asked, grouped Today / This week / Earlier. New starts a fresh one. Conversations title themselves from your first question; you can rename or delete them.
Conversation — the same panel, full width.
Sources and Recent changes — what this conversation can read from, and what it has changed.

Below a laptop width the conversations rail becomes a slide-over sheet and the right column becomes tabs under the composer. Nothing is dropped.

It is genuinely one conversation: expanding to the full page doesn’t start a new thread, and neither does walking to another part of the product mid-conversation — you’ll see “Continuing your conversation from another page.”

Ask something

Type and press Enter (Shift+Enter for a new line). While it works you get honest progress — Understanding your question…, Searching your learning materials…, Composing your answer… — an elapsed counter, and a Stop button that actually aborts the request rather than just hiding it.

An empty conversation offers starters appropriate to where you are:

On an organization scope — “How many learners are enrolled right now?”, “Which courses have no teacher assigned?”, “Summarise what changed this week.”, “What is waiting for me to mark?”
On a program or assessment — “Summarise this assessment for me.”, “Which questions look too easy or too hard?”
Contextual chips also appear on the page you’re on: Who is on this?, How is it going?, What’s missing?, What needs me?

Answers end with Looked at: and the things it read. Counted figures are clickable — following one takes you to the rows the number came from. Where it could only read part of a large set, the figure is marked “partial — narrow the question for an exact count” instead of being presented as exact.

If the answer isn’t what you meant, the reply offers Not what you meant? with Look at the reports instead, Search the documents instead, and Reword it — each of which re-asks your own question against a different body of data.

Find out what it can do here

Choose What can you do? in an empty conversation. It lists what you can ask for on this page, grouped as Content, Publishing, People on courses, Timing and Marking, with each entry marked where relevant:

asks you first — a proposal you confirm.
needs a second approver — it queues rather than runs. See Approvals.

The list is filtered by your actual permissions, resolved on the server, so it never offers something that would be refused when you asked for it. Without permission to change anything you get: “You can ask me about anything on this page, but you don’t have permission to change things from here. An admin in your organisation can widen that.”

Attach documents to a conversation

Sources on the full page — or Add a document in the composer — attaches documents so the assistant answers from them.

Choose Attach. Pick from your organization’s Sources library, or Upload a new document from here.
Wait for its status to reach Ready. You’ll see it move through Queued → Reading → Read → Preparing to search → Ready. A file that can’t be read says Couldn’t be read and offers Try again.
Ask your question. The answer cites the document.

Selecting an attached document opens a preview. It opens on the Document — the original file you attached, rendered inline for PDFs and images and painted from the file itself for Word and PowerPoint — because that’s what you recognise. Extracted text is one click away and shows what the assistant actually reads and grounds its answers on; Open original downloads the file. If a format can’t render inline, the extracted text is shown instead.

A document you upload from inside a conversation is private to you. It is marked private to you and nobody else in your organization can read it, its text, or search it — not even someone who can manage the shared document library. Documents you attach from the shared library keep whatever visibility they already had.

The assistant can only search the documents attached to the conversation you’re in. There’s no way to ask it to reach a document you haven’t attached.

Make a change

Ask for a change in plain words — “Create a new module called Introduction”, “Rename this program to…” — and you get a proposal card, not a change.

The card names the specific things it would touch (“Names 14 items”, expandable), and for bulk enrollment work it goes further: the server runs the change without committing it and reports the outcome each person would actually get — Would be enrolled, Already enrolled — no change, Would go to the waitlist (full), Needs consent first, Would fail. Both the preview and the real run are built from the same request, so they can’t describe different work.

Then:

Confirm runs it, through exactly the same gated path a manual edit would take.
Not now discards it.
Mass or irreversible changes — bulk withdrawals, enrollment transfers, prerequisite gates — arm a second button: “This is a mass, irreversible change. Click ‘Confirm again’ to proceed.”
Publishing and archiving a program don’t run on your confirmation at all. They queue for a second person. See Approvals.

Everything that ran is listed under Recent changes, most of it undoable for 24 hours. See What the assistant changed.

Privacy of your conversations

Your conversations are yours. The database rule is self-read-only: nobody else in your organization can read your transcripts — not someone who manages people, not someone who administers the organization. Oversight runs through the change receipts and the audit trail instead, which record what was done without exposing what you typed.

There is one deliberate exception, for legal hold and safeguarding. Someone holding permission to update the organization can export a specific conversation, and only under three conditions at once: a second person must have approved that specific request beforehand, the export is written to the audit trail before the transcript is returned, and a permanent disclosure is added to the conversation itself — you will see “This conversation has been read by an administrator” with the date and the stated reason. It cannot be dismissed, missed, or switched off.

Conversations and receipts are kept for 180 days and then deleted automatically.

Reference

Thing	Behaviour
Open	Ask in the header · ⌘J · ⌘K → Ask the assistant · home composer · in-context Ask AI buttons
Scope	Shown on every conversation as “Looking at: …”, plus any attached documents
Reads	Under your own permissions, re-checked on the server
Writes	Always a proposal you confirm; some ask twice; some need a second approver
Attached documents	Uploaded from a conversation are private to you; from the library, unchanged
Undo	24 hours, by the person who made the change
Transcript privacy	Self-read-only; break-glass export is four-eyes gated, audited and disclosed
Retention	180 days, purged weekly
Availability	Rollout opt-in by Lurno + AI enabled for the organization + permission to use AI

Troubleshooting

There’s no Ask button anywhere. Your organization isn’t in the assistant rollout yet. Nothing you or your administrator can grant changes that — contact Lurno.

The Ask button is there but greyed out. It reads “AI features are turned off for this organisation.” Either AI hasn’t been switched on for the organization, or you don’t hold permission to use AI. The first is a Lurno setting; the second your administrator can grant.

It says it can’t answer, or offers the reports instead. It read something and the evidence didn’t support an answer. That’s deliberate — it abstains rather than guessing. Take the Look at the reports instead or Search the documents instead offer, or narrow the question.

A number looks wrong, or says “partial”. Large result sets are read in pages; when it couldn’t read all of them it marks the figure partial rather than presenting it as exact. Narrow the question — one program, one cohort, one date range.

It won’t answer from my document. Check the document reached Ready in the conversation’s Sources. It can only search documents attached to the conversation you’re in.

“Too many tries — wait a moment and try again.” There’s a per-person rate limit on AI requests. Wait and retry.

Requests start failing after heavy use. Your organization’s AI budget may be exhausted. Budgets are set per organization by Lurno.