The AI Assistant

Your application may include a chat assistant that lives in a panel on top of whatever screen you are looking at. You can ask it questions about your data, ask it to summarize the record you have open, and — if your administrator allowed it — ask it to make changes for you.

This chapter describes the assistant panel as a user of it: how to open it, how to hold a conversation, what it knows about your screen, and where its limits are. It does not cover how assistants are created or configured; that is your administrator's work.

What This Is

If your application has the assistant, you will see a colored button in the bottom-right corner of every screen, marked AI and labeled with the word for chat. Clicking it opens a panel anchored to the same corner.

The panel contains:

This Is An Optional Capability

The assistant is not part of every application. If you do not see the button in the bottom-right corner, your application does not have it — there is no menu entry to switch it on, and nothing is hidden behind a setting you can reach.

The button also stays hidden when:

• the capability exists in your application but your role is not allowed to use the chat, or
• no assistant has been made available to your role or to you personally.

All three cases look identical from where you sit: no button. If colleagues have it and you do not, that is a question for your administrator, and the answer is usually one of the three above. See Why Your Screen Differs for the general pattern.

When You Use This

The assistant is at its best on questions that are awkward to express as a filter, and on work that would otherwise mean reading several records and writing something up.

Good uses:

• "Summarize this claim and tell me what is still missing." (on an open record)
• "Which of these five suppliers has the oldest unpaid invoice?" (with rows selected in a list)
• "Draft a short reply to the customer explaining the delay, in a neutral tone."
• "Explain what the fields on this screen mean in plain language."

Poor uses — a list screen will do it faster, more reliably, and repeatably:

• "How many open tickets are there?" — that is a filter plus a count.
• "Show me everything overdue this month." — that is a saved view you can reuse every day.
• Anything you will need again next week. An assistant answer is a one-off; a saved view is permanent and shared.

Rule of thumb: if the answer is a list, build a list. If the answer is a sentence, ask the assistant. See Filtering Your Data and Saved Views And Queries for the alternatives.

Opening And Sizing The Panel

1. Click the AI button in the bottom-right corner. The panel appears in that corner, and your most recent conversation with the selected assistant opens automatically.
2. Use the close button in the panel header to hide it. Nothing is lost — the conversation is stored and will be there next time.

The panel sits on top of the page but does not lock it behind a backdrop. Wherever the page is still visible, you can keep clicking around, open a record, change a filter, all with the panel open.

It also survives moving between pages. Moving to another screen inside the application normally leaves the panel open and the conversation intact. A full browser reload closes the panel; reopening it brings your conversations back from the list.

Resizing

On a desktop-sized screen the panel has two drag areas:

• The left edge — drag sideways to make the panel wider or narrower.
• The top edge — drag up or down to make it taller or shorter.

Both sizes are remembered in this browser, per application. Sign in on a different computer, or clear your browser data, and the panel returns to its default size. This is the same pattern as the light/dark theme and your column widths — see The Application Window.

On phones and narrow windows the drag handles do nothing: the panel uses the space available and manages its own height.

The Panel Rearranges Itself By Width

• Wide — the conversation list sits alongside the messages, in its own column.
• Narrow — the conversation list slides over the messages as a small overlay; clicking outside it or picking a conversation closes it again.

The button in the header that shows and hides the conversation list works in both arrangements. Making the panel narrow enough also collapses the list on its own.

Choosing Between Assistants

If your application has more than one assistant, the picker labeled Agent in the panel header lists the ones you are allowed to use. Each entry shows the assistant's name followed by the outside AI service and model it runs on.

Two things worth knowing:

• Each assistant was set up for a particular job. One may have been given instructions to behave as a support triage helper, another as a reporting assistant, another as an operations assistant that is permitted to change records. They are not interchangeable, and they do not share instructions or access rights.
• Which assistants you see depends on you. An assistant can be restricted to specific roles or specific people. A colleague may see three where you see one.

Your choice is remembered in this browser, so the next time you open the panel you are back with the same assistant.

Example A dispatch team's application offers two assistants: "Order helper" and "Warehouse operations". The first can only read; it answers questions about orders and drafts customer replies. The second is permitted to update delivery dates. A dispatcher asking "when will order 4471 ship?" should stay on "Order helper"; a dispatcher who wants three deliveries pushed by one day needs "Warehouse operations", and will not see it at all if their role was not granted it.

Conversations Belong To One Assistant

The conversation list always shows only the conversations you held with the currently selected assistant. Switching the picker clears the message area and loads that assistant's own conversations. Nothing is deleted — switching back brings the earlier conversation into view again.

Holding A Conversation

Sending A Message

Type in the box at the bottom and press Enter to send. Shift+Enter inserts a line break instead, which is what you want for a longer, multi-paragraph question. The send button in the bottom-right corner of the box does the same thing as Enter.

While the assistant is answering, the message box is locked and the send button turns into a stop button.

Watching The Answer Arrive

Answers stream in as they are written, a few words at a time, with three animated dots while there is more to come. Formatting (headings, bullet lists, tables, bold text) is rendered as you would expect, and any links in the answer open in a new browser tab.

If the assistant needs to look something up before it can answer, a small status label appears on the answer itself, describing what it is doing at that moment. On a broad question this can take a while with nothing but that label changing — that is normal. The status line in the panel header does not track this; it stays as it was until something fails.

Tool Activity

The strip labeled Tool activity, just above the message box, is collapsed by default. Expand it with the chevron button to see a short list of the most recent things the assistant actually did while answering — the lookups it ran, the records it touched.

This is the honest record of the work. If an answer looks surprising, open this strip: it will usually show either that the assistant read something narrower than you assumed, or that it changed something you did not expect.

Only the most recent entries are kept, and only for the conversation you are looking at.

Stopping A Running Answer

Press the stop button (the send button becomes it while an answer is running) to cut an answer short.

• Whatever text had already arrived stays in the conversation, marked as stopped.
• If the assistant had not produced any text yet, the empty answer is removed entirely.
• Your own question stays in the conversation either way.

Stopping is the right move when you realize halfway through that you asked the wrong question, or when the assistant has clearly misunderstood and is writing three paragraphs about it.

Retrying

A Retry control in the bottom-left of the message box, beside the paperclip, re-sends your most recent question as a fresh attempt. Use it when an answer failed with an error rather than when you dislike the answer — for a different answer, ask a better question instead.

If sending failed, any files you had attached are put back into the message box so you do not have to attach them again.

Starting A New Conversation

The New button at the top of the conversation list starts a fresh, empty conversation with the same assistant.

Start a new conversation when you change subject. The assistant carries the earlier messages of a conversation with it, so a long thread that wandered from invoices to holiday planning to invoices again produces worse answers than three short, focused threads.

The Conversation List And Archiving

Each conversation in the list shows a title, a snippet of the last message, and when it was last active. Titles are produced for you — you cannot rename a conversation — and one that has not been given a title yet shows a placeholder instead. Clicking one opens it.

The circular-arrows button in the panel header reloads the list, which is useful if you have been chatting in another browser tab.

The archive button in the header (the archive-box icon) removes the current conversation from your list. You are asked to confirm first.

Archiving is a one-way action from inside the panel. There is no "show archived" switch and no restore button, so treat it as "I am done with this and do not want to see it again", not as a filing system.

Attaching A File

You can send a file along with a message in two ways:

• Click the paperclip in the bottom-left of the message box and pick one or more files.
• Drag files from your desktop and drop them onto the message box. A dashed drop area sits below the box at all times; it highlights while you are dragging over it.

Attached files are listed above the message box with their type and size. Each one has an × to remove it. The send button stays disabled until every attachment has finished uploading; an attachment that fails is marked in place, and you can remove it and try again.

Keep attachments modest in size. Files are prepared inside your browser before they are sent, so a very large document can be slow or can fail outright.

Example A supplier sends a revised price list as a PDF. On the supplier's record, open the assistant, attach the PDF, and ask: "Compare the prices in this file with the agreed prices on this supplier and list only the items that changed." The assistant reads both the attached file and the record you have open.

Note that attaching a file to the chat is not the same as filing it on a record. If the document belongs on the record, upload it there as well — see Uploading Files And Images.

Files The Assistant Produces

When an assistant generates a file — a spreadsheet, a document, an image — it appears under the answer in a Generated files block, showing the file name, its type and size, and a Download file button. Images additionally show a small preview you can click to open full size. Links inside the answer text pointing at the same file work too.

Downloads open in a new browser tab.

A generated file is part of the conversation, not part of your data. If it matters, download it and store it where your team expects it — attach it to the record, or put it wherever your application keeps documents. Do not rely on the conversation being a permanent archive.

It Knows Which Page You Are On

Every message you send carries a short description of where you were when you sent it. That description covers:

• which screen you are on;
• the record you have open — its record type, its title, and its identifying number;
• the records you have selected in a list, if any — the first few individually, plus a count of how many more; they travel as record identifiers, not as names, so the assistant has to look them up to know what they are;
• files and images attached to the record you have open, up to a limit.

The description is short by design and is trimmed if it grows too long, so it is a hint about where you are, not a copy of the screen.

If a pop-up window is open on top of the page, the record in the pop-up is what counts as "the current record".

This is why short instructions work. You do not have to tell the assistant which order you are looking at.

Example On the detail screen of project "Riverside refit", you type "what is late here and who owns it?". The assistant already knows "here" means that project and answers about it. Typing the same question on your dashboard, with no record open, produces a much vaguer answer — or a question back.

Example On a list of applicants, you select five rows and ask "which of these has the strongest logistics experience?". Those five records travel with the question. Selecting rows is only possible on table layouts — see Selecting Records And Bulk Actions.

The Boundaries, In Plain Words

This section is the important one. Read it once even if you skip the rest.

The Assistant Does Not Use Your Access — It Uses Its Own

Each assistant works through its own account, created by your administrator, with its own permissions. It is not looking at your data through your eyes.

The practical consequences:

• The assistant may be able to read records you cannot see. Do not treat it as a safe way to check what you personally are allowed to know.
• The assistant may be unable to read records you can see plainly on your screen. When it says it cannot find something you are looking at, this is usually why — it is not being difficult, and repeating the question will not help.
• What it is allowed to do is a fixed part of how it was set up. You cannot widen or narrow it from the chat, and neither can the assistant.

Which assistants appear in your picker is a separate matter, decided per role or per person. Being able to talk to an assistant tells you nothing about how much that assistant can see.

If you handle data with confidentiality rules attached, ask your administrator what each assistant's account is allowed to reach before you use it on that data.

It May Be Allowed To Change Your Data

An assistant can be set up in one of two ways:

When an assistant is set up to require agreement, an attempt to change something is held back until you approve it, and the assistant should come back and ask you in the conversation. There is no separate approve button anywhere in the panel — you answer in the chat like any other message. When it is set up to act directly, a clear instruction can be carried out without a second question; how careful it is about asking first depends on the instructions your administrator wrote for it.

You cannot tell which of these you have from the panel. Ask your administrator which mode your assistant is in before you use it on live data, and assume the more permissive one until you know.

Two facts about the application make this worth taking seriously:

• There is no undo anywhere in this application. Nothing the assistant changes can be reversed with a button.
• Deleting is usually permanent. Unless your administrator turned on the recoverable-delete behavior, a deleted record is gone. See Deleting Records And The Trash.

So treat "please tidy up these records" as an instruction you would give a new colleague on their first day: be specific, be small, and check the result.

It Can Be Wrong

The assistant produces confident, fluent text whether or not it is right. It can misread a field, misinterpret your question, invent a plausible-sounding detail, or summarize five records and quietly leave out the sixth.

Check anything consequential against the record or list itself before acting on it — anything that goes to a customer, anything involving money or dates, and anything you would have to undo.

The Tool activity strip is your quickest check: it shows what was actually read, so you can see whether the answer rests on the records you meant.

Your Messages Leave The Application

Conversations are processed by an external AI service that your administrator chose and connected. Your messages, your attachments and the description of the page you were on are sent to that service.

Apply the same judgment you would to any other outside service: if a document is not something you would email to a supplier, think before you attach it. If your organization has rules about which data may leave its systems, those rules apply here.

The conversation list you see is your own — you are not browsing colleagues' conversations, and they are not browsing yours.

The Page Behind The Panel Does Not Update Itself

If the assistant changes a record, the list or record screen sitting behind the panel keeps showing what it fetched earlier. Nothing in this application pushes data changes to an open screen.

Use the refresh control on the list (the circular-arrows button; it re-runs the same view with your filters and page intact) or reopen the record to see the new state. See Saved Views And Queries.

Tips And Gotchas

• Enter sends, Shift+Enter makes a new line. Do not go by the tooltip on the send button — it names Shift+Enter as the send shortcut, which is the opposite of what happens.
• Refreshing the whole browser page mid-answer cuts the stream off. Reopen the conversation from the list to see what was stored.
• A long silence usually means work, not a hang. Expand Tool activity to see what is happening. If the status line turns red, something genuinely failed — read it, then retry.
• One subject per conversation. Long, wandering threads make answers worse and slower.
• Ask for the shape you want. "As a five-bullet summary", "as a table of item and price", "in two sentences" all work and save you editing.
• Say which records you mean. "This one", with a record open, is precise. "The Novak order", typed on a dashboard with nothing open, is not.
• The panel is per-browser. Its size and the assistant you picked do not follow you to another computer. Your conversations do, because they are stored on the server rather than in your browser.
• The panel does not lock the screen. You can keep working on the page behind it; drag its left edge to narrow it if you need more room for the record.
• Nothing you type reaches a colleague. The assistant is not a way to message someone. For that, use comments and mentions — see Comments And Mentions.

What To Ask Your Administrator For

Everything about the assistant is set up centrally. Useful, concrete requests:

• "Please give my role access to the assistant." (when you have no AI button and colleagues do)
• "Please make the reporting assistant available to my team as well."
• "Is the assistant we use allowed to change records, and does it have to ask me first?"
• "The assistant cannot see our supplier contracts. Please widen its access, or tell me it is deliberate."
• "Please set up an assistant that knows how our complaints process works, so it stops answering generically."
• "The assistant keeps getting X wrong." — worth reporting; the instructions an assistant follows can be adjusted.

See What To Ask Your Administrator For for how to phrase requests so they can be acted on.

Related Chapters

• The Application Window
• Roles, Groups And Permissions
• Uploading Files And Images
• Saved Views And Queries
• What To Ask Your Administrator For