Gemini for Accountants: Gmail, Sheets and AI() for Practice Tasks

An accountant can use Gemini in Google Workspace to summarise client email threads, draft routine replies, tidy and categorise lists in Sheets, and turn notes into documents. Check =AI() results by tying them back to the source: compare counts and totals, test a random sample against the originals, and flag anything outside your allowed categories.

The check matters more than the formula. =AI() returns text that looks tidy whether it is right or wrong, and it cannot undo its own output. A practice that adds a tie-out step, the same way it would for a junior's working paper, gets the speed without importing the risk.

Follow me on Instagram@sagnikteaches

What your plan includes before you plan the work

Gemini arrives in Workspace through the plan you are on. As of November 2026, Google's plan comparison gives Business Starter Gemini in Gmail, the Gemini app and Vids, while Business Standard and above add Gemini in Docs, Sheets, Slides, Meet, Drive and Chat. Google's USD list prices are $7, $14 and $22 per user per month for Starter, Standard and Plus on a one-year commitment; month-to-month billing costs more, and new customers are often shown an introductory discount, so confirm the figure on your own billing page. Google's admin help says changes to Gemini settings can take up to 24 hours to apply, so switch things on before the day you need them.

Connect on LinkedInSagnik Bhattacharya

The AI() function in Sheets needs an eligible plan too, and because Gemini in Sheets is not on Starter, a practice on Starter will need Standard for the people doing this work. If a user sees "AI function not available", Google lists three causes: the plan, the administrator's settings or the language setting. Sort that out for one pilot user first. The full plan-by-plan breakdown is in what Gemini in Google Workspace includes on each business plan.

Subscribe on YouTube@sagnikteaches

One more point on data. Google's Workspace privacy hub says customer data, including prompts, is not used to train models without the customer's permission. That settles training, not access: who in the practice can open a shared sheet, how long chats are kept and whether staff use personal accounts are still settings and habits you check in your own admin console rather than assume.

Gmail: four jobs that save real minutes

Summarising a long client thread

A client has sent twenty-two messages over three weeks about a late filing, a director's loan and a change of year end. Before a call, open the thread and ask Gemini for a summary with decisions, open questions and dates. The output saves you rereading, but it is a map, not the record. Click through to any message the summary relies on for a figure or a promise.

Summarise this thread for a partner picking it up cold.
Give me: (1) what the client wants, (2) what we have already told them,
(3) figures or dates mentioned, with the message they came from,
(4) anything we have promised and not yet delivered.
If something is unclear, say "unclear" instead of guessing.

An illustrative result might list "Client believes the director's loan was repaid in March; our last email said we had no evidence of this." That is the line you click into. In practice, the summary's most valuable sentence is usually the one that exposes a disagreement.

Drafting the routine reply

Records request, deadline reminder, "do we need your signed engagement letter" chaser: all are formulaic. Write one good version, then ask Gemini to adapt it to the client's wording. Keep the facts you supply (dates, document names) in the prompt, so the model has nothing to invent.

A before and after: the draft chaser you typed reads "Hi, still waiting on your bank statements." After asking for a polite version that lists the three missing statements and states the date they are needed by, you get a message that names the account, gives a deadline and offers a secure upload route. Read it once, fix the tone, send it. Never let an unreviewed draft out when it quotes a figure or a legal deadline.

Triage of an inbox after leave

Ask Gemini to list messages needing action by client and urgency. Use it to prioritise, then open the urgent ones yourself. For more on this, how to use Gemini in Gmail to clear your inbox faster shows the prompts and the limits.

Turning a query into a checklist

A client emails: "I started selling on a marketplace and bought a van, what do I need to tell you?" Ask for a checklist of the information a practice would normally need for a new income stream and a new vehicle, framed as questions for the client. You then edit it to match your own onboarding checklist. The value is the first draft in thirty seconds. The accuracy of tax treatment is yours, so the checklist should ask for facts and documents, never state the answer.

Sheets and AI(): what the function does and doesn't do

The syntax is =AI("prompt", [optional range]); Google also accepts =Gemini() as the same function. The prompt describes the task, and the range supplies the data. Nothing generates until you select the cells and click Generate and Insert. Google's help page lists what matters for practice use:

• Responses are text only. Numbers come back as text unless you convert them.
• The function does not see the rest of the sheet or your Drive files, only the range you pass.
• Only the first 350 selected AI cells generate per run.
• AI functions cannot be nested inside other functions such as IF.
• Results cannot be undone or redone, only regenerated with Refresh and Insert.
• Short-term and long-term generation limits apply, and a long-term limit can mean a 24-hour wait.

Three further facts to build into your routine. First, output does not regenerate by itself when an input changes: you click Refresh and Insert, and if you build the prompt by joining cells with &, Google says Sheets will not even prompt you that the data changed. Second, the function can pull real-time information from Google Search, so a prompt such as "what is the current mileage rate" returns a web answer with no source you can file; rates and thresholds belong in a reference tab you maintain. Third, Google says plainly that Gemini output may be inaccurate and should not be relied on as financial or professional advice. That last point is the reason for the tie-out step. One helpful detail for your file: each Generate and Insert is attributed to the person who clicked it in the sheet's version history, so you can see who produced a set of results and when.

Five Sheets tasks that suit AI() in a practice

Notice the pattern in the prompts. Each limits the answer to a fixed set or a fixed format, and each offers an exit such as Other or NOT FOUND. Open-ended prompts produce open-ended outputs you cannot test mechanically. For a wider set of ideas see Gemini in Google Sheets for small businesses: 10 practical uses.

A worked example: categorising 200 bank lines, then tying out

Say a sole-trader client sends a year's bank export of 200 lines, with a total of money out of 48,615.40 in the period. You have a tab called Source with columns A (date), B (narrative) and C (amount, as a number). On a tab called Work, you copy the three columns, add column D for AI category, and keep Source untouched.

In D2 you enter:

=AI("Assign exactly one category from this list: Travel, Software, Rent, Meals, Fees, Stock, Other. Reply with the category word only.", B2)

Because only the first 350 AI cells generate per run, 200 rows is fine in one go. Fill down, then check the results before you trust any of them. Here is the tie-out, in the order I would run it.

1. Completeness. Count rows holding a text result with =COUNTIF(D2:D201,"?*"), which ignores empty strings and errors; it should equal 200. Avoid COUNTA here, because it counts a formula cell even when nothing has generated. Any shortfall means some cells did not generate, so select them and run again.
2. Allowed values only. On a Lists tab put the seven allowed categories in A2:A8. Then =SUMPRODUCT(--(COUNTIF(Lists!A2:A8, D2:D201)=0)) counts rows whose result is not on the list. It should be zero. A non-zero count means the model added a word, a full stop or a new category.
3. Totals reconcile. In a summary tab use SUMIF per category on column C and add the seven results. The total must equal 48,615.40 to the cent. This catches lost or duplicated rows, though not wrong categories.
4. Sample test. Copy the Work tab and paste it as values, so sorting cannot disturb the AI cells. Add =RAND() in a helper column, paste that as values too, sort by it and read the first 25 rows against their narratives. Record how many are wrong. If three or more are wrong, your prompt needs examples or the task is too ambiguous for AI().
5. Review the high-value lines. Filter amounts over 500 and check every one by hand, whatever the sample showed. Mistakes on big items matter most.
6. Freeze it. Copy the results and paste as values into a final tab, then note the date, the prompt text and who checked. Pasting values stops a later refresh silently changing a signed-off result.

An illustrative result: completeness 200, off-list 3, totals tie, sample 2 wrong of 25, high-value check found one fuel purchase coded Travel that was actually Stock for a market stall. The three off-list rows, on inspection, were "Meals." with a full stop. A small fix, but without the SUMPRODUCT check they would have dropped out of the category totals and made the summary short by the sum of those lines. The total tie-out would have caught that too, which is why you run both.

When the prompt needs examples

If the sample test fails, add two or three worked examples to the prompt: "STRIPE PAYOUT = Other; ADOBE *CREATIVE = Software". Better still, put your mapping in a small table and apply it with a normal lookup first, leaving AI() only for the narratives the table does not recognise. That hybrid is faster to audit. A lookup is deterministic and AI() is not, so every row the lookup handles is a row you never need to question.

A before and after to show it: raw narratives "AMZNMKTP*2K4", "ADOBE *CREATIVE CLD" and "RAILCO TICKETS" arrive. The lookup table handles ADOBE and RAILCO. The AI() column handles the Amazon marketplace code, with the prompt told to answer Other when the item is unclear. You end up with a smaller AI footprint and a clearer audit trail.

A second example: pulling fields from forty pasted invoices

Another practice task. A client forwards forty supplier invoices as pasted text in a single column, and you need date, supplier, net amount and total for each. On the Work tab, column A holds the pasted text, and B to F hold one AI() formula each: date, supplier, net, tax and total. For the net amount (column D) the prompt is:

=AI("From this invoice text return only the net amount before tax as a plain number with a decimal point, or NOT FOUND. No currency symbol, no words.", A2)

Because results are text, wrap the numeric column in VALUE() in a separate column when you add them up, and count rows where VALUE errors. Then the tie-out has its own twist: an invoice's net plus tax should equal the total. In a check column G write =ROUND(VALUE(D2)+VALUE(E2)-VALUE(F2),2) and filter for anything other than zero. Illustrative result: thirty-six rows tie exactly, three rows are NOT FOUND because the pasted text was cut off, and one row differs by 10.00 because the model took a delivery charge as the net line. Four rows to open against the PDF instead of forty. That is the benefit: the tie-out turns a vague worry about accuracy into a short, specific list.

Note that the check column uses ordinary formulas. Nesting AI() inside other functions is not allowed, so keep AI() cells as plain outputs and do every comparison in separate cells beside them.

A one-page procedure to put in the practice manual

Write the routine down so any team member can follow it and a partner can inspect it. A filled-in version of the core of such a page:

• Scope: Gemini in Gmail and Sheets may be used for drafting, summarising, categorising and extraction. It may not be used to state a tax or legal position.
• Data allowed: client names only where the task needs them; no bank account numbers, national identifiers or passwords in prompts or ranges.
• Workbook rule: run AI() in a copy; keep Source untouched; paste results as values once checked.
• Tie-out minimum: completeness count, off-list count, totals reconcile, 25-row sample, every line over the practice's review threshold.
• Sign-off: the preparer records prompt text, date and error count; a reviewer other than the preparer signs off anything going to a client.
• Review date: reread after any Gemini change notice, and every quarter regardless.

The sample size is a judgement. Twenty-five rows from two hundred catches a prompt that is badly off, but it will not find a one per cent error rate. If a task feeds a figure into a filing, check more, or check all of the rows that move the totals. If you cannot afford that, the task is not yet ready for AI().

Docs, Meet and the other places Gemini turns up

On Standard and above you also get Gemini in Docs and Meet. For an accountant the practical uses are drafting a client letter from bullet points, summarising a meeting for the file, and restructuring a long advice note. Two cautions. First, a Meet summary is notes taken by a machine, not a record of advice, so a human confirms any action involving money or a deadline. Second, ask clients before you record or transcribe, and follow your own firm's policy. Our tutorial on the best AI note taker for accountants compares options if Meet's isn't enough.

For drafting with house style, a Gemini Gem can hold your standard instructions, such as tone, sign-off and "never state a tax conclusion". Google is replacing Gems with skills: its transition page gives March 2027 for Workspace business accounts, existing Gems move across automatically, and Google's admin guidance says migrated ones arrive as inactive drafts the creator must review and switch on. Until then, anyone you share a Gem with can view its instructions and uploaded files, so keep client material out of it, and check the sharing rules again once your Gems become skills. See building Gemini Gems for repeat business tasks.

Where a mistake would hurt and what to do instead

• Anything that states a tax or legal conclusion. Use AI to draft questions, then decide yourself.
• Calculations. Use native formulas for arithmetic. Ask Gemini to explain a formula, not to produce the number.
• Personal identifiers. Strip names, account numbers and reference codes from ranges where the task does not need them.
• Filed or signed documents. No AI output goes into a filing without a named person reading it.

A realistic failure: an assistant asks AI() to "extract the net-of-tax total" from pasted invoice text and gets a plausible number from the wrong line. That kind of slip passes a glance but not a tie-out to the invoice. The lesson is to extract fields, then recompute the arithmetic in a normal cell and compare.

Setting this up for a team of three

Roll it out in one week for a pilot user. Day one: confirm plan, admin settings and a test of =AI() on dummy data. Day two: write the five prompts in the table and save them in a shared sheet. Day three: run a closed client job through the process and score the tie-out. Day four: fix prompts. Day five: write a one-page procedure covering allowed data, the tie-out steps, who signs off and how outputs are frozen. Then widen to the rest of the team, one task at a time.

Keep a running log of tie-out results. If the off-list count or sample error rate drifts upward after a Gemini update, you will see it in the log. For a general library of Sheets prompts beyond accountancy, there is also Google Sheets AI prompts. If you want help deciding which of your tasks suit this route, that is what my AI implementation consultation covers, at the contact page.

Gemini in Workspace: questions from practices

Which Workspace plan do I need for the AI() function?

Google says the function needs an eligible Workspace or Google AI plan, and its plan comparison lists Gemini in Sheets from Business Standard upwards, not on Starter. A notice such as 'AI function not available' can also mean admin settings or the language setting block it, so test one account before promising it to staff.

Does Gemini see my whole spreadsheet?

No. Google's help page states that the AI function cannot see your whole spreadsheet or other Drive files, so the data it should use has to be in the range you pass. That is a useful limit, because it makes you choose exactly what is exposed.

Can I undo an AI() result?

Google says output cannot be undone or redone, only regenerated. Keep an untouched copy of the source tab, work in a copy of the workbook, and treat each run as something you verify rather than accept.

Is it safe to put client data in Gemini?

Google's Workspace privacy hub says customer data, prompts included, is not used to train models without the customer's permission. That covers training, not everything: still review your own admin settings, sharing and retention, and remove names and identifiers where the task does not need them.

Should clients be told I use it?

Many practices mention AI-assisted drafting in the engagement letter or a short client notice. Describe the tasks, the data involved and the review that follows, and have a partner approve the wording.

Further reads

• Best AI Tools for Chartered Accountants: Seven for a Small Practice — Where Gemini sits among seven tools for a small practice.
• AI Engagement Letters for Accountants: Faster Drafts, Clear Scope — Drafting scope wording, including a line on AI use.
• How Accountants Use AI to Spot Errors in Client Books — Using AI to flag oddities in client ledgers.
• Gemini or ChatGPT for a Google Workspace Business? — Choosing between Gemini and ChatGPT inside Workspace.
• How to Use Gemini in Google Docs to Draft and Edit Faster — Drafting and editing client letters in Docs.
• ChatGPT for Accounting With Excel: Seven Tasks, Each With Its Check — The same checks-first approach, done in Excel.
• Is Gemini Safe for Confidential Business Data in Workspace? — Where to check Gemini's data settings in your admin console.
• 7 AI Mistakes That Put an Accounting Firm's Client Trust at Risk — Seven AI slips that cost practices client trust.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Google Docs Editors Help, 'Use the AI function in Google Sheets'; Google Workspace pricing and plan comparison page (USD list prices); Google Workspace generative AI privacy hub; Google Workspace Admin Help on turning the Gemini app on or off; Gemini Apps Help, 'Share a Gem' and 'About the transition from Gems to skills'. All checked November 2026.