Skip to main content
cfo.ai brings data from your other systems into your Model in three ways, and each one updates on its own schedule. Which way a source uses answers the common questions: why did my numbers change overnight, is my bank balance live, and does an MCP server’s data land in my Model?

​
Three ways data reaches cfo.ai

Integrations and uploads only read from the source, so a number you edit in cfo.ai changes your Model and leaves your books alone. An MCP server or API can also act in its service, such as creating an issue in Linear, but only through the tools you allow Ari to use. Credential values that an MCP server’s tool result carries, such as an authorization header, a cookie, or an API key that a failed call echoes back, are removed before Ari reads, stores, or shows the result; the error numbers and messages that explain the failure stay. This filter covers MCP tool results only; an API endpoint’s response body is outside it.

​
From a source record to a number on your Page

Take one line of your QuickBooks profit and loss: $4,200 of software subscriptions in September 2026. It travels in three steps.
  1. The sync copies it into the warehouse. cfo.ai keeps a private copy of each source’s tables in its data warehouse, and QuickBooks’ monthly profit and loss lands there as a table, row for row.
  2. A saved query reads it. A saved query is a SQL statement that selects, cleans, and labels the rows your Model needs. cfo.ai adds standard ones for supported sources, such as QuickBooks Profit & Loss, and Ari writes the rest. This one puts the $4,200 in an Expense column, beside the account name, the Entity, and the date.
  3. Each column becomes a Variable or a Dimension. Expense becomes a Variable that adds up, segmented by the account Dimension and the Entity Dimension. Its starting formula sums the source column, so September’s software subscriptions read $4,200 on any Table block that shows Expense by account.
When the next sync brings a corrected amount, the warehouse row changes, the saved query runs again, and every Page that reads Expense updates. Nothing about your formulas or tables has to change. How imported data loads lists the rules for each step. A CSV follows the same path: the upload replaces the warehouse copy, and a saved query brings its rows in. A Google Sheet is different: Ari imports the spreadsheet itself and rebuilds its tabs, formulas, and values in your Model, the same way it imports a spreadsheet model.

​
Connected is not the same as ready

A source becomes useful in three stages, and cfo.ai reports each one separately. First it is linked: you signed in and the connection exists, but no data has moved. Then its tables reach the warehouse, which your Model cannot use until a saved query reads them; that is why a connection can read No queries yet, which someone who can edit the connection must fix. Finally saved queries put rows into your Model, and only then does the sidebar show the connection as Synced. So “I connected QuickBooks” and “my Model has my QuickBooks numbers” can be minutes or hours apart. QuickBooks Online and Xero load your three most recent complete months first, so Ari can start, and fill in the rest of the history in the background; How much history loads lists each source. Ari tells you in the chat when the data is in. Until then, ask Ari where the load stands rather than treating a smaller Model as missing data. Ready also says nothing about whether the numbers match your books. A query can filter rows on purpose, and your Model’s Last close decides which months read as Actuals. Check that your numbers match your books covers both.

​
When data updates

Every integration refreshes once a day, in a shared nightly run that starts at 02:00 UTC. A source that publishes its daily data after that run would always land a day late, so you can give a connection its own daily time, such as 6:30 AM in your time zone. You can also ask Ari to refresh a connection now. A healthy connection has no refresh button; Retry appears when a sync or setup needs another attempt. Connection actions lists when each action is offered. The nightly refresh does not always re-read your whole history. Some sources, such as QuickBooks Online, re-read only recent months, so a correction to an older period arrives only when you ask Ari for a full-history import. How far back each source re-reads is listed in How much history loads and in each source’s entry in the integrations directory. Uploads never refresh, because a file has nothing to sync from; upload the new version. A connected Google Sheet refreshes every night. Every source’s schedule is listed in Refresh schedule. So is your bank balance nightly? Through your accounting system, yes: it is the balance on your synced balance sheet, as of the last sync and as current as your books. Through an MCP server such as Mercury’s, it is live each time Ari looks it up.

​
Integrations and MCP servers do different jobs

An integration answers questions about history: how much revenue came in each month, how payroll grew, where cash went. It brings in past periods, keeps recent ones current, and turns them into Variables your formulas and Scenarios can use. How far back each source loads is listed in How much history loads. An MCP server or an API answers questions about now, or acts in another tool: what is in the bank today, which invoices are open, which issues are blocked. Ari calls it live, so the answer is current, but nothing stays in your Model unless Ari writes a value into a table, and that value is a plain number that does not update. So adding every relevant MCP server does not replace your integrations. If you connect Rippling’s MCP server, Ari can look up an employee or count today’s headcount. Your headcount plan, though, needs a roster in your Model that updates every night, and the Rippling integration brings that. When you do want a server’s data on a schedule, ask Ari for a sync setup. A sync setup copies one read-only tool’s results into your data each night, and a saved query brings them into your Model, like an integration. At that point the server uses a plan slot.

​
Connected data is shared across Scenarios

Every ordinary Scenario reads Main’s imported rows, so a sync updates Main and those Scenarios together, while a Snapshot keeps the numbers it was pinned with. See Scenarios, Snapshots, and version history.

​
When a source stops updating

Most failures fix themselves. A provider’s sync that failed for a moment is retried on the next nightly run, and a table that failed to refresh is retried while the rest of the source stays usable. The sidebar marks these with a steady orange dot. Sometimes the nightly sync of a source that was not connected from a chat fails right after a successful one. No chat is there to hear about it, so cfo.ai starts one background task for Ari, which acts for you with your permissions. Ari checks the connection and repairs what it can, such as a saved query that a change in the source’s layout broke. A source counts as restored only when its queries return values again for the records they returned before. If the repair restored the source, it says nothing. If it needs your decision, could not finish, or changed something you should know about, it opens a chat titled [Connection name] stopped updating in your chat list. A source that keeps failing starts no further task. Problems only you can fix, such as an expired sign-in, wear a red triangle until you act. Fix a connection that needs attention walks through each one.

​
Connection slots

Your plan limits how many sources you can sync, so each integration uses one of its connection slots until you remove it. All connected Google Sheets together use one slot. Uploads use none, and an MCP server or API uses none until Ari saves a sync setup for it; the setup creates a connection that counts. Removing a connection stops its syncs, and the data it already loaded stays in your Model. The exact counting rules are in the integrations directory, and how plans set the number of slots is in Plans, AI usage, and connection slots.