Chatting With Your Xero Books: What the Answer Covers at Month-End

On 14 October, Xero launched a connected app for Google Gemini. Ask Gemini what your cash position is, which bills are overdue or how last month's profit compares, and it answers from your live Xero records rather than from an exported spreadsheet. Xero lists the same kind of connection for Microsoft 365, Claude and ChatGPT.

For a founder, that is a convenient way to look at the numbers. For whoever closes the books, it raises one question worth settling before the first board pack is built from a chat window: what exactly does the answer cover?

The short version is that a conversational answer reports the ledger as it stands at that moment. It is as complete as your reconciliation, and no more. This post sets out what the app does, why that limit matters most in the first week of a month, and how to tell which questions are safe to ask before the close is finished.

What Xero has released

According to Xero's release, the app runs on JAX, Xero's agentic platform, and can return:

• financial health figures: profit and loss, income and expenses, the balance sheet and cash position;
• sales and receivables: overdue balances, draft invoice summaries, high-revenue clients and recurring subscriptions;
• purchases and payables: outstanding and overdue bills, and what is due to suppliers;
• charts in the chat, such as profit and loss graphs and aged receivables;
• organisation and contact details.

Two design choices in the release are worth noting. Any change to financial data, such as issuing an invoice or sending a payment reminder, needs explicit confirmation in the app before it runs, and Xero says the final review stays with the accountant or bookkeeper and the business owner. Xero also states that shared data is used only for the user's session and that "Xero data is not used to train Google Gemini models." A scheduling feature, Gemini Tasks, is listed as coming soon for eligible plans.

Those are sensible boundaries. The app reads the ledger and asks before it writes. What it cannot do is make the ledger more complete than it is.

The answer is the ledger, not the bank

Xero keeps two views of each bank account. The statement balance is built from the lines that arrive through the bank feed. The balance in Xero is built from transactions that have actually been entered or matched in the ledger. Xero's bank reconciliation report shows both, and between them it lists the outstanding payments and receipts and any unreconciled statement lines. When everything is matched, the two balances agree.

A deposit sitting in the reconcile tab has reached the bank feed but not the ledger. Until someone matches or codes it, it is not in the profit and loss, the balance sheet, or anything built from them, and that includes an answer in a chat window. The answer will be correct for the ledger it read. It will just be describing a ledger that is not finished yet.

For most businesses, the gap is a few small items. For a Shopify brand, it is often the bulk of the month's revenue activity, because sales, fees, refunds and reserves arrive bundled inside processor payouts that someone has to decompose before they can be posted.

A worked example

Take a brand that books revenue when it reconciles each Shopify Payments payout. On 3 November, the founder asks Gemini for October's net profit. The ledger answers $21,400.00.

Three payouts from the last days of October are still unreconciled in the bank feed. Simplified, and leaving GST aside, they contain:

Once those payouts are decomposed and posted, October's net profit becomes $21,400.00 + $14,267.15 = $35,667.15. The first answer was not wrong about the ledger. It was wrong about October, by about 40% of the final figure.

The same gap shows up differently if a connector posts daily Shopify sales into a clearing account. Revenue is then already in the profit and loss, but the payouts that should empty the clearing account have not been matched, so the clearing balance is overstated and the cash position understated. Ask "how much cash do we have" and the answer leaves out money that is in the bank.

Neither case is a fault in the AI. Both are the ordinary state of a ledger partway through a close.

Which questions are safe when

A practical rule is to sort questions by whether they depend on the close being finished.

Payables and receivables questions read documents that exist in Xero as soon as they are entered. Profit, margin and cash questions read balances that only settle once the payouts are posted, accruals are booked and month-end adjustments, such as cost of sales and the refund liability, are in.

Before trusting a month-end figure from any source, chat or report, three checks settle most of the doubt:

1. The bank reconciliation report shows no unreconciled statement lines up to the last day of the month.
2. Each processor clearing account is at zero, or its balance is explained by payouts that cross the month-end.
3. The period has been locked in Xero, so the figure you were given cannot change under you.

Why this matters more as AI spreads

Use of AI in finance is now close to universal. Avalara's 2026 State of Finance report, which surveyed 400 finance, tax and accounting professionals across Australia, Canada, the UK and the US in March and April 2026, found that all of them used at least one AI solution, up from 84% in 2025 and 47% in 2024.

Trust has not kept pace. In a 2026 Datarails survey of 270 US CFOs and finance leaders at organisations with at least $100 million in revenue, only 4% said they trust AI with their month-end close, and 75% named lack of auditability as the biggest reason they hesitate.

A conversational interface over the ledger makes the numbers easier to reach. It does not change where they come from. The audit trail for an October profit figure is still the set of entries behind it, and those entries are only as good as the reconciliation that produced them.

What this means for your close

Asking your books questions in plain language is a real convenience, and Xero's confirmation step for anything that changes data is the right default. Treat chat answers the way you would treat a report run mid-close: accurate for the ledger as it stands, provisional for the month until the bank lines are reconciled, the clearing accounts are at zero and the period is locked.

The slow part of getting there, for most Shopify brands, is turning each payout into posted entries. That is the work Vert is built to do: decompose each payout, prove it sums to the deposit to the cent, and draft a balanced entry for your approval. If that is where your close stalls, you can join the pilot.