Using the AI help assistant

Opening and closing the panel

Whenever you are signed in, a small round "i" button floats in the bottom-right corner of every page. Click it and a help panel opens in the corner, the button turns into an ×, and the text box is focused so you can start typing right away.

The panel stays open on purpose. Clicking elsewhere on the page does not close it, and it stays put as you move between pages and reload, until you close it yourself. You can drag the small grip bar at the panel's top edge up or down to make it taller or shorter (your size is remembered on that device), and double-click the grip bar to return to the default. Closing the panel does not erase your conversation. Reopen it and your messages are still there.

1. 1Look for the round "i" button in the bottom-right corner (you must be signed in to see it).
2. 2Click it to open the panel; the button changes to an ×.
3. 3Start typing in the focused text box.
4. 4Close the panel by clicking the × in its header, clicking the corner button again, or pressing Escape.

Asking a question

When you first open the panel with no conversation yet, you see a short welcome line and a few example question buttons you can click to ask instantly, such as how to file a grievance, send a mass email, add a new staff user, or import a payroll master list. Click any example to send it immediately, or type your own question.

While the assistant is thinking you see a three-dot typing indicator, and the answer then streams in word by word as the panel auto-scrolls to keep the newest text in view. The input box is disabled while an answer is streaming and re-focuses automatically when it finishes, so you can ask a follow-up right away. Answers come back as formatted text with numbered steps, short paragraphs, bullet lists, and tables where useful.

1. 1Type your question in the "Ask a how-to question…" box at the bottom, or click an example button.
2. 2Press Enter to send (use Shift + Enter to add a new line inside your question first).
3. 3Or click the ↑ send button (it is disabled while the box is empty or an answer is still arriving).
4. 4Wait for the answer to stream in, then ask a follow-up if you need one.

What it can and can't answer

The assistant answers only from the user manual. In scope is anything about using the app: finding a screen, what a button does, who can perform an action, and how a workflow runs from start to finish. It remembers the recent back-and-forth, so follow-up questions work naturally.

Three limits are built in, and they are there to keep answers honest. Anything that is not the app (general knowledge, current events, coding help) gets a polite no that steers you back to the app. Your data is hidden from it. It cannot see member records, grievance contents, emails, or files, so if you ask for a specific record it tells you where in the app to look it up yourself. And if something is not in the manual, it says so plainly rather than inventing buttons that do not exist. The assistant also answers using your deployment's display names, so its labels match what you see on screen.

Answers shaped by your exact permissions

Every question is quietly enriched with fresh context the moment you ask: the page you are on, your role and the area you cover, whether you can open Role Management, and the full list of what you are and are not allowed to do. That list is resolved from your role's settings plus any personal exceptions an administrator set for you.

Because of that, the assistant does not guess from a typical role. If you were granted a capability just for you, such as Mass Email, it knows and walks you through using it. If a capability you would need is denied, it tells you up front that you do not have access and that an administrator can grant it in Role Management, instead of giving you steps you cannot complete. Permissions are looked up on every question, so answers always match your access right now. If an administrator just changed something, simply ask again.

On privacy: what leaves the system with each question is your question text, the recent conversation, your current page path, your role, and your capability list. Never member records or any other stored data.

Take-me-there links and conversation history

When an answer mentions another page in the app, the assistant links to it. Inside the panel those internal links render as navigation buttons with a small → arrow, such as Grievances →. Click one and the panel closes and the app navigates straight to that page. Your conversation is kept, so you can reopen the panel on the new page and continue where you left off. Links to anything outside the app open in a new browser tab and leave the app untouched.

Your conversation is kept for the lifetime of the browser tab and stored only in that tab's temporary session storage. Nothing is saved to the server or your account. Closing and reopening the panel, navigating between pages, and refreshing the page all preserve it (the most recent twenty or so messages). Closing the tab or window wipes it completely, and there is no history page to retrieve past conversations. If an answer contains steps you will need later, copy them somewhere before closing the tab. Only the most recent 20 messages are sent with each question, and any single message longer than about 4,000 characters is truncated, so keep questions focused.

1. 1Click a navigation button such as Grievances → in an answer to jump straight to that page.
2. 2Reopen the panel on the new page to continue the same conversation.
3. 3Copy any steps you will need again into your own notes before closing the tab.

Limits, errors, and healthy skepticism

Each user can send up to 20 help requests per minute. Past that you see an italic note in the chat asking you to wait a moment, and you can try again after a few seconds. Other in-chat messages cover an assistant that is not configured for your deployment yet, a service that cannot be reached, and an answer that was interrupted mid-stream (whatever arrived stays on screen, and re-asking usually finishes the job). There is currently no thumbs-up or thumbs-down feedback button. If an answer is wrong or out of date, note the question and the reply and report it to your administrator, who can correct the underlying manual chapter so future answers improve.

Answers are generated by Anthropic's Claude (the Claude Haiku 4.5 model), with the entire user manual supplied as its knowledge base on every request. A permanent reminder at the bottom of the panel notes the guidance is AI-generated from the manual and the assistant cannot see your data. Always sanity-check critical steps, especially anything that deletes, sends, or imports, against the relevant manual chapter or with an administrator.