How To Use Gemini Notebook for Finance

I have a seven-store coffee company I use to test finance tools. Eighteen months of P&L, a journal-entry-level general ledger, six months of bank transactions, 336 monthly comments from store managers, and a couple of policy documents nobody has opened since the day they were adopted.

I dropped six of those files into Gemini Notebook and asked it to do my job.

Twenty minutes later I had a variance table for every location and account, a chart of actuals against budget, a list of four places where the general ledger and the P&L disagree, and an audit of six months of bank balances against our investment policy. I never opened Excel.

Here is how it went, and where it still needs someone in the chair.

What changed in July, beyond the name

Google renamed NotebookLM to Gemini Notebook on July 16, 2026. If the rebrand were the whole story I wouldn’t be writing this.

The real change is that every notebook now gets its own secure cloud computer. It writes Python and runs it against the files you uploaded, then hands back tables, charts, and CSVs as artifacts you can download. Before, it read your documents and answered questions about them. Now it can calculate.

For finance that’s the difference between “summarize this P&L” and “compute the variance for all seven locations, sort by absolute dollar amount, and show me the code you ran.”

The code part matters more than it sounds. You can read the logic when the Python comes attached to the answer, and when someone in your close meeting asks how you got there, you have something better than “the AI said so.” Most AI tools for finance cannot do that, which is why this one stayed on my desk after the test was over.

What you need before you start

Five minutes of setup, and none of it is technical.

• A Google account with Gemini Notebook access, at notebooklm.google.com. Same URL as before, new name and logo in the corner.
• Code execution switched on. Google’s help pages put running code, charts and downloadable files on the AI Pro and AI Ultra tiers. Ultra and Workspace business accounts got it first in July 2026 and Pro followed a few weeks later. If you don’t see it, check the dropdowns in settings and in the chat window before you give up on it.
• Your files sitting in Google Drive, because Gemini Notebook takes sources rather than workbooks, and adding from Drive is the cleanest path. If your data already lives in Sheets, the Google Sheets side of the stack feeds it directly.
• Files with some structure to them. Clean-ish column headers, one topic per file. It handles messy data fine, but it can’t guess what an unlabeled column means any more than you can.
• Room under the source cap. Free notebooks hold 50 sources and Pro holds 300, and every plan caps one source at 500,000 words or 200MB. Six finance files is nowhere near it.
• Files you are allowed to upload. Check your company’s policy first. This is a Google cloud service and your GL is your GL.

I gave mine six files: a company overview, 18 months of P&L, 336 manager comments, the full general ledger, six months of bank data, and an investment policy document.

How the notebook is laid out

Three panels, and once you see them the whole thing makes sense.

On the left are your sources. Everything the notebook knows comes from here. Ask it a question before you add sources and you get generic knowledge, which is not why you’re here.

In the middle is the chat. It looks like Gemini, ChatGPT, Copilot, or Claude, and it behaves like them with one difference: every answer carries a citation chip pointing back to a source.

On the right is Studio, which holds the artifacts. If you’ve used Claude artifacts, it’s the same idea. Charts, CSV tables, and generated files land here and you can download them.

That’s the tool. Sources in, chat in the middle, artifacts on the right.

Ground it before you ask it anything hard

The temptation is to load everything and ask the biggest question you have. Don’t. Start with three files and one boring question, because the boring question is how you learn whether the citations work.

Step 1: create the notebook and add three sources

I created a notebook called F9 Finance Coffee Shop and added the company overview, the P&L, and the manager comments. You can add sources from Drive, from a URL, from pasted text, from an upload, or by asking it to go research something on the web. Drive is the one you’ll use.

Step 2: ask a question you already know the answer to

Using only the sources I've uploaded, give me a one-page briefing on
F9 Finance Coffee Company: how many locations, when each opened, who
manages each one, and which locations have incomplete history in the
P&L. Cite every claim.

It came back before I finished talking. Seven locations, opening dates, managers, and a note on which stores have gaps in the P&L because they opened partway through the period.

The part worth your attention is the little citation symbol next to each claim. Click it and the source panel jumps to the exact place the answer came from. Click “view source” and you’re looking at the row.

That’s the whole trust model. If you can’t click back to the number, don’t use the number. And I mean click, not glance at the little symbol and feel reassured by it. Researchers at TU Delft and the University of Amsterdam who studied how these systems attach sources found up to 57% of citations were post-rationalized: the model landed on an answer first, then went and found something that backed it up. Princeton’s ALCE benchmark found the best models still left half their statements without full citation support. The chip tells you where to look. That part is worth eight seconds.

Step 3: make it find a figure and check it yourself

Which store month has the single largest sales revenue in the entire
P&L file? Give me the store, the month, the dollar amount, and then tell
me what the manager commentary says about that month.

Astoria, July 2025, $952,412.19. I clicked the citation, landed on the sales revenue rows, scrolled to July, and read the same number off the source. Then it pulled the manager’s comment for that month, which explained the spike and mentioned the operational strain that came with it.

Two files, one answer, both verifiable in about eight seconds.

Step 4: read the file nobody reads

Every finance team has a commentary file. Managers type into it every month, it goes into a folder, and it is functionally write-only, because 336 rows of prose has no formula.

Read all 336 manager comments. Find every comment where a manager
reports a payment error, a duplicate charge, a missing record, or
something they were unsure how to book. List each one with the
location, the month, and a direct quote.

Upper West Side paid rent twice in January. The autopay ran, then a physical check cleared for the same amount. The store manager wrote it down in her monthly comment, nobody read the comment, and the month closed. It had been sitting in a file called CoffeeShop_LocationManagerComments, in the same folder as the P&L, since January.

It also came back with missing records, non-standard booking practices, and cash activity happening outside the point-of-sale system. All of it useful in a reconciliation, none of it visible when the commentary lives on seven different people’s laptops.

Now make it do the math

Grounding is the setup. This is the part that’s new.

Building a variance table with code

Using the P&L file, build a variance table for the full period, for
every location and account. Calculate actuals, budget, the variance,
and the percent variance. Show me the code you run. Return the table
sorted by absolute dollar variance, largest first.

It ran for about five minutes, which is slow by chatbot standards and fast by finance standards. Long enough that I tabbed away and forgot about it. What came back: the Python it executed, the finished variance table, and a downloadable CSV sitting in Studio. It also volunteered a set of observations I hadn’t asked for, which were mostly right.

Five minutes for a full variance pack. I’ve spent a lot of first-week-of-the-month days building that by hand.

Charting without a chart wizard

Now chart Astoria's monthly sales revenue from January 2025 through
June 2026 as a line chart, with the budget line overlaid. Label the
peak month with its dollar value.

Clean chart, peak labeled, budget line where it should be. Google’s chart output is better than I expected, which is not a sentence I planned to write. I use Claude for most of my work and I’d still hand this particular job to Gemini.

The part where it was smarter than I planned for

I set a trap. I asked it to rank all seven locations by how badly they missed their opening-quarter sales budget, with no guardrails, expecting it to confidently tell me our newest store was the worst performer in the company.

Williamsburg opened mid-March. A half month of revenue against a full month of budget will always look like a catastrophe, and it isn’t one.

It caught it. Without being told, it noted that some locations opened partway through the period, offered two ways to interpret the ranking, and then worked both of them rather than making me choose.

So I gave it the correction anyway:

Williamsburg opened mid-March 2025 and Long Island City opened in
September 2025, so their first month is a partial period being compared
against a full month of budget. Exclude each store's first partial
month, then redo the ranking. Tell me what changed.

It came back fast, collapsed the two interpretations into one, and named Astoria as the worst normalized miss at -$161,000, down 17%.

So it dodged the trap on its own, and my correction still earned its keep by turning two hedged answers into one number I could take into a meeting. You are the analyst either way. The judgment just moves off the assembly and onto the framing.

Three systems, one question

Everything so far ran on files the notebook already had. This is where I handed it two it had never seen.

I added the full general ledger, at journal-entry level, and six months of bank transactions. Different systems, different grain, different shape. This is the same tie-out I ran when I turned Claude into my accountant, so I knew what a right answer looked like before I asked.

Use code execution. Roll the general ledger up to the P&L level using
the PnL_Group column, excluding the balance sheet and other groups.
Compare that rollup to the P&L file by location, PnL_Group, and month.
Return only the combinations where GL and P&L do not tie, with the
dollar difference on each. Show your code.

Four breaks came back:

Location	What broke
Upper West Side	Rent, which is the duplicate payment showing up again from a completely different direction
Upper West Side	Utilities out of step
Park Slope	A sales revenue difference on the catering line
Long Island City	Internal wholesale revenue, which is a consolidation rollup problem

Then it offered to help correct them, which I did not take it up on.

Does the ledger balance

Check whether the general ledger balances. Sum all debits and all
credits across every journal entry line. If they don't match, find what
the issue is, and show me the memo and who posted it.

The answer was more interesting than a yes or no. At the corporate level everything comes out in the wash, so a consolidated check would have passed. At the location level, individual books are out of balance, and Williamsburg is the worst offender.

That’s a real finding. Corporate accounting has been absorbing the difference, so nothing ever escalated, and the location books have been wrong the whole time and nobody knew.

Auditing a Word document against a bank file

Last one, and the one I’d point a controller at.

Our investment policy is a Word document. It has almost no numbers in it. It has rules: minimum operating cash, what counts as a reportable breach, what instruments are permitted, when a capital expenditure needs a funding plan.

I asked it to review that policy against every bank transaction and list each breach with the section it violates. The reusable version of that prompt:

Using the Investment Policy document and the bank Balances sheet:
audit every account against the policy. List every reportable breach
with the account, the dates, the lowest balance reached, and the exact
policy section it violates.

It cross-referenced prose against six months of daily balances and came back with a liquidity breach at Astoria, the rent overpayment again, and a suggested framework for auditing the transaction log so it stops happening.

Nobody has time for that audit. That’s why it never gets done.

What it caught, all in

Six files, roughly 20 minutes of my time:

• Four breaks between the general ledger and the P&L
• A duplicate rent payment, found twice, from two different files
• Location-level books out of balance while corporate looked clean
• Reportable policy breaches nobody had checked for

None of that is a close. It’s the review pass that sits on top of one, and it’s the first thing to go when the deadline is Thursday. If you want the close itself run end to end by an AI, that’s a different workflow with a different tool.

Where it still needs you

I want to be straight about the limits, because a tool review that only lists wins is an advertisement.

Start with what it doesn’t know, which is your business. It caught the partial-period problem in my data, but it caught it from a pattern in the file, not from knowing Williamsburg opened on a Tuesday in March. Give it a period comparison where the context lives in somebody’s head and it will compute exactly what you asked for and hand it back with a straight face.

Then there’s the answer itself, which is yours to own. Every number I put in front of a CEO out of this workflow, I clicked back to the source and read with my own eyes first. The ICAEW puts it in stronger language than I would, in its guidance on managing AI risk in audit work: human-in-the-loop is a safeguard, and the person doing the reviewing has to understand what the tool can and can’t do.

It replaces the four to six hours of assembly around the judgment. That’s the trade, and it’s the same trade behind every piece of finance automation worth running.

It reads files. It does not model them. Scenario toggles and a live-linked deck are a different tool, and you already know which one. This is for interrogating what you have.

The big-file question is still open. I deliberately left my 7.3-million-row point-of-sale file out of this. That’s a different test and a different video.

Common issues

You don’t see a code execution or analysis option

The feature landed on Google AI Ultra and eligible Workspace business accounts first and rolled out to Pro from there. Check the dropdowns in settings and in the chat window before assuming your account doesn’t have it. If it really isn’t there, the notebook still reads and cites, it just won’t compute.

It answers from general knowledge instead of your file

You asked before the sources finished processing, or the question didn’t force it to look. Add “using only the sources I’ve uploaded” and “cite every claim” to the prompt. Both phrases do real work.

Your Excel file won’t upload cleanly

Put it in Google Drive first and add it from there. Multi-tab workbooks are fine, but give each tab a clear name, because the notebook will refer to them by that name.

The numbers look right but you can’t trace them

If a claim has no citation chip, treat it as a guess. Ask it to redo the answer with citations before you use anything in it.

What it costs

Gemini Notebook has a free tier, and the reading and citing features work there. Code execution rides on the paid plans. As of August 2026, Google AI Pro sits at $19.99 a month and Ultra starts at $99.99, with Workspace business plans on their own path. Check Google’s current pricing page before you buy, because these tiers have moved twice in a year and any number I print here has a short shelf life.

For most finance pros reading this, Pro is the tier that matters. And if your company is a Microsoft shop, the comparison worth running is against Copilot Cowork, which your M365 license may already cover.

So where does Gemini Notebook for finance fit

Anywhere you already have the files and already know the question. The review pass, the reconciliation, the policy check, the commentary file nobody opens. It reads across systems that were never built to talk to each other, and it shows its work on the way through.

It is not your modeling stack and it is not your reporting layer. For where the rest of Google’s tools land in a finance function, I went through the wider set in Google Gemini for finance.

The prompts

Every prompt in this article is in my AI Library, along with the correction prompt for partial periods and the three-file tie-out sequence as a reusable set. They’re written to work on any ledger, not just my coffee data.

Grab them at f9finance.com/ai-library, and I’ll send you Finance AI Insider each week with the next thing I’ve tested.

If you try the tie-out on your own general ledger, I want to know what it found. That’s the prompt where this stops being a demo.