TL;dr: On October 19, Google gave Gemini in Sheets the ability to query BigQuery in plain English. You connect a BigQuery table through Connected Sheets, open the Gemini side panel, and ask questions like "what were total sales in the Northeast last quarter," and Gemini writes and runs the query against the full dataset in BigQuery, not just the rows visible in your spreadsheet. It can also take a request like "build a quarterly sales review with revenue by region, a monthly trend, and our top 10 products," show you its plan for review, and then build a new tab with native, connected, refreshable objects. Before you touch it, there are three things to check: tie every AI-generated number back to a number you already trust, confirm who owns the refresh and who is allowed to see the data, and understand that every question you ask bills against your BigQuery project, because one curious prompt against a large fact table can burn through the monthly free allowance in a single query.

Today, we will be looking at what actually shipped, how to set it up step by step, and the verification ritual I would run before letting any AI-generated number near a model I sign off on. This matters because the bottleneck it kills is one every finance team lives with, which is the "can you pull this for me" request that sits in the data team's queue for three days while you build a forecast on stale numbers. Google is pitching exactly that: business users run the analysis themselves, and the data team's backlog of one-off pulls gets shorter.

Setting up your connection: first, connect the BigQuery data through Connected Sheets, which is the native connector Sheets has had for years. Go to Data, then Data connectors, then Connect to BigQuery, pick your project, dataset, and table, and the table lands in your spreadsheet as a connected object. Second, open the Gemini side panel in Sheets. Third, ask your question in plain language. Gemini writes the query, runs it against BigQuery itself, and returns the answer. The important detail is that the query runs against the entire table in BigQuery, not against whatever subset happens to be pasted into your sheet, which is what makes this different from asking an AI to analyze a CSV you uploaded. Fourth, keep the conversation going with follow-ups. Asking "how does that compare to the same quarter last year" keeps the thread and the filters without you restating them. Fifth, when you want the dashboard, describe it in one request and Gemini shares its plan for you to review before it builds anything. Approve the plan and it drops a new tab into the spreadsheet with database formulas, pivot tables, and charts that stay connected to BigQuery, so a scheduled refresh keeps them current. You are not exporting a snapshot. The tab stays live.

Here is the verification ritual I would run, because the query being written by an AI does not change the rule that no number enters my model without a tie-out. First, every answer gets checked against a number you already trust. If Gemini tells you Northeast sales were $4.2M last quarter, that number gets reconciled to the prior close, the dashboard finance already publishes, or the last pull the data team ran, before it goes anywhere near a forecast. Second, ask it to show its work. The side panel can tell you what it queried and which filters it applied, and you read that the way you would read a colleague's query: which table, what date grain, what got excluded. Third, be deliberate with follow-ups. The thread keeps your filters, which is convenient, and it is also exactly how a filter gets silently dropped three questions in. When the answer changes shape unexpectedly, restate the full question fresh and compare. Fourth, review the dashboard plan the way you would review a junior analyst's proposal. Gemini shows it before building, and that review is the control point: check the grain, the date ranges, and which objects will refresh on whose credentials before you approve.

The billing trap is the part most announcements skip, so let me state it plainly. The feature itself costs nothing extra on a qualifying Workspace plan, but every question you ask runs a real BigQuery query against your project, and BigQuery bills on-demand at $5 per terabyte scanned after the 1 TB monthly free tier. Ask a curious question against a 2 TB fact table and you have spent real money before lunch. This is manageable with guardrails, and the guardrails are worth setting up before the curious questions start. Connect Gemini to aggregated tables and views rather than raw fact tables wherever you can, because a monthly summary table costs a fraction of a scan to query. Know which dataset the connection points at and who pays for that project, because in most companies that is the data team's budget, not yours. And treat the first month as metered: run your questions, then open the BigQuery console and look at what actually got billed, so the cost of curiosity is a measured number instead of a surprise.

Two more things worth knowing before you roll this out. Permissions behave the way Connected Sheets has always behaved, which means a user with only Sheets access and no BigQuery access cannot manually refresh or schedule refreshes of the data. If you build a dashboard and share it with someone who lacks the underlying BigQuery access, they will see your last refresh and nothing newer, which is fine as long as everyone understands that is the deal. And refresh ownership matters more now that an AI is in the loop. Decide who owns the scheduled refresh, what credentials it runs under, and who gets notified when it fails, because a dashboard that quietly stops refreshing in February will mislead people with a straight face.

What this changes about the FP&A and data team relationship is worth sitting with. The one-off pull requests get shorter, which is good for everyone, but the data team's job does not disappear, it moves up a level. Someone still has to decide which tables are safe to expose, what the definitions are, and what grain the business should be asking questions at. That semantic layer, the shared agreement on what "revenue" and "last quarter" mean, becomes the most valuable thing the data team owns, because every natural-language question Gemini answers is only as good as the definitions underneath it. Finance practitioners should be asking for that layer explicitly instead of just celebrating that they no longer need SQL.

Start with one table you know well, run the tie-out ritual on the first five answers, check the BigQuery bill after a week, and then decide how wide to open the door. The template below has the prompt library, the QA checklist, and the cost estimator ready to copy.

Template: template: BigQuery natural-language prompt library + QA checklist (Google Sheets, make a copy to use it)

- Francois