Opening the Copilot
Three entry points reach the same assistant, reading the same ledger data under the same permissions. Which one you use is a question of surface rather than capability. Press Cmd+K (Mac) or Ctrl+K (Windows) from any screen to open a centered command bar with the cursor ready for input. This is the fastest path for a single question. Select the Copilot button in the top bar to open the Agents panel, which stays alongside the main content area and keeps your conversation history. It suits multi-turn work such as iterating on a journal entry draft, and it is the only surface with navigation, so skills, connectors, and scheduled agents are managed there. The panel also supports copy-and-paste of tabular data, so you can paste a snippet from a spreadsheet and ask the copilot to analyze it against your GL. On an individual record or report, the embedded copilot appears with that record already in scope, so you do not have to name it in your question. The difference that matters is scope, not power. A question typed into the command bar reaches the same ledger under the same permissions as the same question typed into the panel, and gets the same answer.Scoping a conversation
An entity selector and an accounting-basis selector sit below the composer, and between them they decide which numbers come back. Both scope the whole conversation rather than the question in front of you, so set them before you ask anything that returns a balance. The Agents Panel Reference lists their options and defaults. Scoping is a conversation-level setting because the copilot is answering from your ledger rather than from the text of your question. It has to know which books to read before it can read them, and nothing in the wording of a question tells it that. A question that names a customer does not imply the entity that invoiced them. Getting these wrong is the most common cause of an answer that looks precise and is not. A consolidated balance quoted to a single-entity customer, or a GAAP figure used in an IFRS workpaper, is wrong in a way the wording of the answer will not reveal. The copilot reports the balance it was asked for, correctly, from the wrong set of books. There is no hedge in the response to warn you, because from the copilot’s side nothing went wrong. This is why the selectors are worth a glance at the start of any conversation that will end in a number someone else relies on. It is also why a scheduled agent, which has nobody at the composer, carries its own entity scope in its definition instead.Querying your ledger
Type a question in natural language and the copilot returns an answer grounded in your GL data. Examples of supported queries:- “What is the AR balance for Acme Corp as of March 31?”
- “Show me all journal entries posted to account 6010 this month.”
- “Which vendors had the highest spend in Q1 2026?”
- “Summarize trial balance movements for February.”
Drafting journal entries
Describe a journal entry in plain text and the copilot proposes the debit and credit lines for you. For example, type “Accrue $12,000 of January rent to prepaid expenses” and the copilot returns a two-line entry with the accounts, amounts, and a suggested memo. You review the proposed lines in a structured preview, adjust any field, and post directly from the copilot panel. The draft follows your chart of accounts and respects any dimension requirements configured on the target accounts. If the copilot cannot resolve an account reference, it asks you to clarify before generating lines. You can also provide multi-line descriptions to create entries with more than two lines. For example, “Split $30,000 of consulting fees across departments Engineering (60%), Sales (25%), and G&A (15%)” produces a six-line entry with the correct amounts pre-calculated. Nothing reaches the ledger until you post it. The preview is the point at which a proposed entry becomes your entry, so review the accounts and the period as carefully as you would a manual draft.Explaining flux variances
During Close Management, you can highlight an unusual variance in the flux analysis view. Ask the copilot “Why did this move?” and it examines the underlying transactions for the account and period. It identifies the largest contributors to the change and returns a narrative summary. You can paste this explanation directly into a close note or workpaper. This workflow pairs with the flux analysis page. You review the numbers there, then use the copilot to generate the qualitative commentary. The copilot attributes each driver to specific transactions, so the narrative is traceable rather than generic. If you need to adjust the explanation, for example to exclude a known one-time item, you can ask the copilot to regenerate the summary with that context, and it produces a revised version in the same conversation thread. Because the commentary goes into a workpaper an auditor may read, wording that gets reused month after month belongs in a skill rather than in a prompt someone retypes.Going beyond a single question
The Agents panel adds three things to the assistant described above: saved instructions, outside systems, and a schedule. Each has its own reference. A skill is a saved set of instructions invoked by typing/ as the first character of a message. Skills keep recurring work worded the same way regardless of who runs it, which matters for close narratives and anything an auditor reads later. Six ship with every organization, including /ar-aging-digest and /period-close-readiness.
A connector lets the copilot act in a system outside DualEntry, such as drafting mail in Gmail. A connector is installed once for the organization but authorized by each person separately, so what the copilot can reach through it depends on who is asking.
A scheduled agent runs an instruction on a cadence with nobody in the composer, which suits recurring review work such as a Monday list of newly overdue invoices. For a worked example combining all three, see how to send dunning emails with Copilot and Gmail.
Data grounding and compliance
Every copilot response is generated from your ledger data, not from generic training knowledge. The assistant does not fabricate balances or invent transactions. If the data needed to answer a question is outside your permission scope or does not exist, the copilot tells you it cannot answer rather than guessing. Each interaction, meaning your prompt and the copilot’s response, is logged with a timestamp, your user ID, and the data objects referenced. You can review these logs under Settings → Audit Trail or export them for external compliance review, and they are included in any audit-trail export you generate for external auditors. Calls the copilot makes through a connector are logged separately, and those records hold metadata only: which tool ran, who ran it, and whether it succeeded. The contents of what was sent are not stored, so a connector-based workflow leaves a thinner trail than a native DualEntry action.Copilot interactions are read-only until you explicitly post a drafted entry. Asking a question never modifies your ledger.