Datarails includes an AI assistant that follows you across the whole product. It sits in a bubble in the corner of every screen, it knows which area of FinanceOS you are working in and what you have selected there, and it is where the output of every AI agent in the product is delivered. This article explains how to use it, how the agents built into individual screens hand their work back to it, and how your data is handled along the way.

Prerequisites

• The assistant must be enabled for your account — Availability is switched on per account. If you do not see the bubble, contact your Datarails representative.
• Agent availability is set per account — The agents built into individual screens (Ask AI on widgets, the AI tools menus) are switched on separately, so you may not see all of them.
• Use the Datarails web application — The bubble is part of the web app. It does not appear inside the Excel add-in, the desktop application, or the Desk module, which has its own embedded assistant.
• Your own Datarails permissions apply — The assistant can only read data you are already allowed to see.

Where to find the assistant

The assistant lives in a bubble in the bottom-right corner of the screen, and it is on every screen of the web app: dashboards, reporting, the workspace, Budget and Forecast management, Cash, Spend Control, workflows and Month-End Close. There is nowhere special to navigate to in order to ask a question.

Step 1: Open the assistant

1. Click the bubble in the bottom-right corner of any screen.
2. The panel opens with a greeting (Hi <your first name>,) and a short line about the area you are currently in, for example, How can I help you move forward with this dashboard?
3. Under Action suggestions, review the suggested starting points for that screen.
4. Click a suggestion, or type your own question in the Ask me anything… box and send it.

The suggestions are not a fixed menu. Each product area provides its own, so what you see depends on where you are when you open the assistant.

How the assistant knows where you are

Every question you send includes two things alongside your words: the product area you're in and the selections currently on that screen. This is what separates the assistant from a general chat window. Because it already knows where you are, a short question still lands. You can ask Why is this number off? without pasting data in or explaining what “this” means, and the assistant will not ask you which module or page you are looking at.

What the assistant is told about your screen depends on the area:

Important: The assistant is handed the names of your selections, never the numbers on screen. It fetches the underlying data itself, server-side, under your permissions. Where a selection changes the answer, the assistant uses it and says that it did. If your question depends on a period, scenario or entity that is not listed above for your screen, include it in the question, for example What moved in OPEX in Q2 2026?

Moving between screens with the panel open

If the panel is open and empty, the greeting and suggestions change as you move between areas. If a conversation is already in progress, the thread stays where it is and the assistant simply notes Now looking at: <area>, so you can keep the same conversation going while you navigate. Start a new chat if you want a clean slate for the new screen.

Step 2: Ask a question

1. Type into the box and send.
2. While the assistant works, a status line tells you what it is doing (Looking through your data…, Crunching the numbers…, or the specific step it is on), so you are never left watching a blank spinner.
3. The answer streams back a piece at a time. Click Stop at any point to keep what has arrived so far.

What comes back matches the question rather than a fixed output type:

For named financial figures such as revenue, margin, or headcount, the assistant reads your organization's governed business metrics where they are defined, and tells you whether a figure came from a governed metric or from an aggregation of the underlying table. It never estimates or fabricates a number. If it cannot answer accurately, it says so and asks one clarifying question.

You can also attach a file to a question, and you can bring a specific agent into the conversation by typing @ and choosing it from the list.

Step 3: Choose a window mode

The panel is not fixed in the corner. Open the Change window mode control and choose one of the options under Switch to:

• Floating — An overlay sitting on top of the page. You can move and resize it.
• Sidebar — Docks the panel to the right and pushes the page across, so nothing is hidden behind it. On screens at least 1024 pixels wide, this is where the assistant opens by default.
• Full screen — Turns the assistant into a workspace of its own, with New chat, Chat history and a Files library down the side.

A few things to know about window modes:

• Your choice of Floating or Sidebar, and the panel's position and size, are remembered per area on this browser. Full screen is not remembered; the assistant reopens in Floating or Sidebar.
• Sidebar is unavailable while another side panel, such as comments, is open. The tooltip Close the other side panel to dock the assistant tells you why.
• Navigating to a different area while in full screen collapses the assistant back to a floating panel.

Step 4: Find an earlier conversation or file

Every conversation you have with the assistant is saved automatically, across all areas. To find an earlier one:

1. Open the More menu in the assistant panel and choose History, or switch to full screen and open Chat history.
2. Browse the list, which is grouped by when you last touched the conversation: Today, Past week and Older. Each row shows the date and the area the conversation belongs to.
3. Use the search box to search the text of your conversations, or click Load more to page further back.
4. Click a conversation to open it again exactly where you left off. You can also rename or delete a conversation from the list.

Two things to know about history:

• It is yours — Conversation history is kept per user. Nobody else sees your conversations.
• Each screen reopens its own latest thread — When you open the bubble on a screen, the assistant restores your most recent conversation from that area. Conversations from other areas are still there; find them in History.

In full screen mode, the Files library collects every file the assistant has produced for you, grouped the same way and then by conversation, so a file you generated last week is still available to download. Generated files are kept for a limited time; a file that is no longer available is marked This file has expired.

Agents built into your screens

Alongside the bubble, Datarails has task-specific agents built into the exact screens where the work happens. These are known as in-flow agents. They are not separate chatbots. Each one opens the same assistant, already loaded with the context of that screen and, in most cases, with the request written and sent for you, so you never start from a blank prompt box.

Note: In-flow agents might not be active in your organization. Having the assistant bubble does not mean you have every agent below. Each in-flow agent is activated separately for your account, so the AI tools menus and the Ask AI button may show fewer options than this table, or none at all. To activate an agent, contact your Datarails representative.

Note that Report Builder, Month-End Close and the planning screens each host more than one agent behind the same AI tools button. If you have found one, check the menu for the others. Everything an agent produces is delivered back into the assistant, so closing a tab never loses your work.

Example: Asking about a dashboard widget

The clearest way to see how these agents work is on a dashboard.

1. Open a dashboard.
2. Hover over the widget you have a question about. An action row appears on the widget.
3. Click Ask AI.
4. The assistant opens with the question already written (Give me insights about the "<widget>" widget on the "<dashboard>" dashboard) and sends it for you.
5. Read the two- or three-sentence explanation of what that widget is showing. The assistant then offers up to two follow-ups, such as explaining what is driving the trend or comparing against budget. Pick one, or ask your own.

The important part is what you did not have to do. You never told the assistant which numbers you meant. Because you started from the widget, the assistant was handed the widget's definition and its current data, computed for the dashboard's reference date, so a question like “what is driving this?” is unambiguous from the first word.

This is also the fastest surface in the product. The first answer is written from the widget's data alone, before any further lookups, and the assistant only goes back to your tables when you ask a follow-up. Every other agent in the table above works the same way in its own place: you click a button on the screen you are already on, the request is sent for you, and the result arrives in the assistant.

Note: Ask AI on widgets is not shown on mobile screen sizes.

Data, permissions and control

Three principles govern every agent in the product:

• It reads your data for analysis — The assistant never changes your financial data. The only things it can create are outputs you asked for, such as a chart, a file, or a widget on a dashboard, and before creating or modifying anything, it states exactly what will change and waits for your OK.
• You stay in control — Every draft is yours to edit or discard. Nothing is posted to your books: a journal entry arrives in the conversation as a draft for you to check, and a person posts it, not the agent.
• Your data is not used to train AI models — See the next section for how this is guaranteed.

Underneath those principles:

• The assistant runs with your permissions — It inherits the permission scope of whoever is using it, so it cannot read anything you cannot read. Authorization is re-checked on every request, not once when you log in, and access to each agent is re-checked on every message, so a change to your access takes effect mid-conversation.
• Conversation history is per user — Your conversations are not visible to colleagues.
• The assistant does not give financial or investment advice — It supports your analysis; the decisions remain yours.

How Datarails uses Anthropic

The assistant and the agents built into your screens are powered by Claude models from Anthropic, running on Anthropic's Claude Managed Agents platform. This means that some of your data leaves Datarails and is processed on Anthropic's infrastructure while the assistant works. Here is what that involves.

What is sent to Anthropic

• Your questions and the assistant's answers, including the area and on-screen selections described above.
• The data the assistant retrieves to answer you. When the assistant queries a table or a metric under your permissions, the results are returned to the model so it can reason over them. This is how it answers with your real numbers.
• Files you attach to a question, files generated for you, and the working data an agent is handed when you start it from a screen, such as the definition and current data of the widget behind Ask AI.

How Anthropic handles it

• Not used for training — Datarails uses Anthropic's commercial API. Under Anthropic's commercial terms, inputs and outputs are not used to train Anthropic's models.
• Limited retention — Each conversation runs as a session on Anthropic's infrastructure, which holds the conversation history, files, and outputs for that session. Anthropic's default for commercial API data is deletion within 30 days, and Datarails can delete sessions and uploaded files on request. Because sessions are stateful, this product is not covered by Anthropic's zero-data-retention option.
• No standing access — Anthropic's platform has no connection of its own to your Datarails account. Every lookup the model requests is executed by Datarails, as you, against your permissions, and only the result of that lookup is returned to the model.

Questions about data residency, retention, sub-processors, or Anthropic's security documentation are handled by your Datarails representative.

If you previously used the Planning, Reporting and Strategy agents

Those three agents, described in The Three Datarails AI Agents, have merged into a single general assistant. Nothing was taken away, and the assistant does everything the three did and more, but you no longer choose which agent to open first. Open the bubble on whatever screen you are on and ask your question. If you still use the AI Agents page, you will find the same single assistant there.

Agents built into your screens

Alongside the bubble, Datarails has task-specific agents built into the exact screens where the work happens. These are known as in-flow agents. They are not separate chatbots. Each one opens the same assistant, already loaded with the context of that screen and, in most cases, with the request written and sent for you, so you never start from a blank prompt box.

Note: In-flow agents might not be active in your organization. Having the assistant bubble does not mean you have every agent below. Each in-flow agent is activated separately for your account, so the AI tools menus and the Ask AI button may show fewer options than this table, or none at all. To activate an agent, contact your Datarails representative.

Updated September 22, 2026 13:54