Practitioner training manual · Version 1.0 · 29 July 2026
Gemini Notebook, which almost everyone still calls NotebookLM, is Google's source-grounded research assistant. You give it documents, and it answers only from those documents, with a citation on every claim that you can click to see the exact passage it came from. For an auditor or a tax practitioner that constraint is not a limitation. It is the whole point.
| Gemini Notebook | A general assistant such as ChatGPT or Claude | |
|---|---|---|
| Where answers come from | Only the sources you loaded | Its training, plus anything you attach, plus the web |
| Citations | Inline on every claim, clickable to the passage | Only when you demand them, and often not verifiable |
| Risk of invented facts | Much lower, because it is grounded | Real, and the main thing you manage |
| Computation | Now runs code in a secure cloud computer, on higher tiers first. See the caution below | Runs real code on your data when you ask it to |
| Best at | Finding, comparing and explaining what is in a body of documents | Drafting, computing, building things, open-ended reasoning |
Read that table twice. It tells you the division of labour: Gemini Notebook for what the documents say, a general assistant for what to do about it. Part 08 covers using them together.
The professional problem with AI has always been the same: fluent output you cannot trace to a source. Gemini Notebook addresses that directly. When it tells you what the group accounting manual says about capitalisation, the citation takes you to the paragraph. You are no longer taking the machine's word for it, which means the output can support work you sign.
Until this update the answer was simple: never take a number from a notebook. That has now changed, but only partly, and the distinction is one you have to hold in your head.
| Situation | Where the number comes from | What you do |
|---|---|---|
| The secure cloud computer runs and you can see it executed code | Actual computation on your sources | Treat it like any other computed working: read the logic, re-perform the material ones, put your own working on the file |
| It answers with a figure in prose, with no code run | Predicted text, not arithmetic | Do not use it. Take the number from the source document or your own schedule |
| You are on a tier where code execution has not reached you yet | Prose, always | Same as above. Use a tool that executes code |
Setup is genuinely quick, because there is nothing to install and the free tier is not crippled. What takes thought is which account you use.
You cannot buy Gemini Notebook on its own. It comes with a Google AI plan, a qualifying Google Workspace plan, or a Google Cloud licence. Prices differ by country and are billed through Google One, so check one.google.com for the rate in Bangladesh rather than trusting a US figure.
| Limit | Standard (free) | In Plus | In Pro | In Ultra |
|---|---|---|---|---|
| Notebooks per user | 100 | 200 | 500 | 500 |
| Sources per notebook | 50 | 100 | 300 | 500 or 600 |
| Chats per day | 50 | 200 | 500 | 2,500 or 5,000 |
| Audio Overviews per day | 3 | 6 | 20 | 100 or 200 |
| Video Overviews per day | 3 | 6 | 20 | 100 or 200 |
| Reports per day | 10 | 20 | 100 | 500 or 1,000 |
| Mind Maps, Flashcards, Quizzes per day | 10 each | 20 each | 100 each | 500 or 1,000 each |
| Deep Research | 10 per month | 3 per day | 20 per day | 75 or 200 per day |
| Watermark removal on slide decks and infographics | No | No | No | Yes |
The per-source ceiling does not change with your plan. A single very large file has to be split before it will load, no matter how much you spend. Plan for that when you load a multi-volume manual or a full year of board papers.
Three panels, and understanding the split is most of the learning curve. This is a drawing rather than a captured screenshot so that each area can be numbered.
The single biggest mistake is treating a notebook like a folder. It is closer to a witness: what it can tell you is bounded entirely by what it has seen.
A notebook that contains the Income Tax Act, three clients' financial statements and a Bangladesh Bank circular will give you muddled answers, because retrieval has to compete across unrelated material. A notebook scoped to one subject gives sharp ones.
| Good notebook | What goes in it | Question it answers well |
|---|---|---|
| Bangladesh income tax reference | Current Act text, operative Finance Act, relevant SROs and general orders | "What does the loaded text say about withholding on this payment type?" |
| Client permanent file | Memorandum and articles, prior year accounts, board minutes, major contracts, accounting manual | "What has the board approved about related party lending?" |
| Bank regulatory library | The circulars your reporting depends on, plus your internal policy | "Which circulars touch provisioning on restructured loans?" |
| Engagement working file | This year's trial balance commentary, audit programme, client responses, PY management letter | "What did we raise last year that is still open?" |
| Standards notebook | The IFRS and ISA texts your firm relies on | "What disclosure does the loaded standard actually require?" |
Inside a notebook you can tick and untick individual sources before asking. Use it deliberately. Asking a question with only the 2026 Finance Act selected gives you a different and more reliable answer than asking with three years of Finance Acts loaded together, because the model cannot then blend them.
Everything downstream depends on this step. Curation is the professional skill here, and it is one accountants already have.
The chat looks like every other AI chat box. The discipline around it is completely different, because here you can and must verify.
Weak: "Tell me about withholding tax."
You get a summary of everything, cited everywhere, useful nowhere.
Strong: "Working only from the sources currently selected, what do they say about withholding on payments to a non-resident for technical services? Quote the operative words, give the section reference, and tell me if any selected source contradicts another."
You get something you can take to a file and verify in ninety seconds.
Working only from the sources I have selected: Question: [state it precisely, with the transaction, parties and period] Answer in this order: 1. The direct answer, in two sentences. 2. The operative words from the source, quoted briefly, with the section, paragraph or clause reference. 3. Any condition, exception or proviso in the sources that qualifies it. 4. Anything in the selected sources that points the other way. 5. What the selected sources do not cover that I would need to resolve this properly. If the sources do not answer the question, say so plainly instead of reasoning around it.
Click every citation that supports a conclusion you will rely on. You are checking three things, and it takes seconds:
The Studio panel turns the same sources into other formats. Some of it is genuinely useful in practice. Some of it is a novelty. Here is the honest split.
| Output | Practice use | Verdict |
|---|---|---|
| Reports | A briefing on what a body of documents says. Good starting point for a technical note or a client update. | Highest value. Use weekly. |
| Data Tables | Pulls structured facts out of unstructured documents, for example every covenant across twelve facility agreements. | Very high value, and underrated. |
| Mind Maps | Seeing the structure of a statute or a manual before you read it. Also good for planning an audit approach. | Useful for orientation. |
| Slide Decks | A first draft of a training session or client presentation, exportable and editable. | Saves an hour, needs your edit. |
| Audio Overviews | Two hosts discussing your sources. Genuinely good for consuming a long circular in Dhaka traffic. Interactive mode lets you interrupt and ask. | Better than expected. Not evidence. |
| Video Overviews | Visual explainer from your sources. | For internal training, not client work. |
| Flashcards and Quizzes | Testing article students on the standards or the Act after you have loaded the material. | Excellent for staff development. |
| Infographics | Visual summaries. Watermarked below the Ultra tier. | Occasional use. |
Generate a report from the selected sources for [audience: engagement team / audit committee / credit committee]. Structure: 1. What the sources collectively establish, in five bullets. 2. The specific obligations or requirements they impose, each with its source and reference. 3. Where the sources conflict or leave a question open. 4. What is not addressed anywhere in the loaded set. 5. Five questions this raises that a reviewer should ask. Keep every statement traceable to a citation. Do not add context from outside the sources. Where you are summarising rather than quoting, say so.
Build a data table from the selected sources, one row per [agreement / circular / assessment order], with these columns: [counterparty] | [date] | [term or effective date] | [key obligation] | [financial covenant and its defined ratio] | [penalty or default clause] | [source reference] Rules: - Every cell must be traceable to a citation. - Where a document is silent on a column, write "not stated" rather than inferring a market norm. - Do not normalise different wordings into one phrase if the difference could matter legally. Quote the distinguishing words.
Five concrete builds, in the order I would do them. Each takes under an hour to set up and pays back on the first use.
Current Act texts, the operative Finance Act, the SROs and general orders you actually use, plus a dated note of what is loaded. This is the answer to the biggest weakness in every general AI assistant: statutory recall. Here the statute is in the room.
Every circular your reporting and provisioning depends on. Then ask questions like "which of these deal with classification of restructured facilities, and do any of them conflict?" That question is hard for a human across forty documents and easy here.
Memorandum and articles, the last three years of financial statements, board minutes, major contracts, and the group accounting manual. Now "what has the board actually approved about intercompany lending?" is a thirty second question instead of an afternoon.
The IFRS and ISA texts your firm relies on. Use it to check what a standard requires before you write the note, and to generate disclosure checklists you then run against your own.
Your audit manual, methodology notes, house templates and technical updates. Then generate flashcards and quizzes for article students, and audio overviews for staff who are travelling. This turns dormant firm knowledge into something people consume.
Working only from the selected sources for this client: 1. List every matter raised in prior year management letters or audit memoranda, with the year and the source reference. 2. For each, state whether anything in the loaded sources indicates it was resolved, and cite what indicates it. 3. Flag every matter where the sources are silent on resolution. These are the ones to carry into this year's planning. 4. List commitments management made in writing, with dates and references. 5. Identify anything in the board minutes that suggests a transaction or arrangement not reflected in the financial statements I loaded. Do not conclude. Give me a planning agenda with references I can verify.
Two things separate an occasional user from someone who has built a system: sharing notebooks properly, and knowing which tool to reach for.
Notebooks can be shared with colleagues, and advanced sharing is available on paid plans. Two points a partner should hold in mind: sharing a notebook does not raise a collaborator's own source limit, and sharing a notebook shares everything in it. Check the source list before you share a client notebook with anyone outside the engagement team.
Deep Research investigates a question, then imports both its report and the sources it relied on directly into your notebook, where they become part of the grounded set for every future question. For industry background, regulatory landscape and comparator work this is a strong workflow. Free accounts get a monthly allowance rather than a daily one, so spend those runs on questions that matter.
This is the other half of the July 2026 rename. You can now access and create notebooks directly inside the Gemini app, with syncing between the app and the standalone Gemini Notebook experience, and Google says notebooks are coming to AI Mode in Google Search. The old frustration that every notebook was an island is largely gone.
For a practice this cuts both ways. Grounded firm knowledge becomes reachable wherever you are working, which is genuinely useful. It also means a notebook holding client material is now reachable from more surfaces than the one you deliberately opened. Fold that into the sharing decision in Part 09 rather than treating it as a convenience feature.
Every notebook is being given a sandbox that lets it write and execute code against your loaded sources, rather than only summarising them. For an accountant this turns a reading tool into something that can also tabulate, reconcile and chart the material it has read, without leaving the grounded set.
Three practical notes before you build a workflow on it:
| Task | Reach for | Why |
|---|---|---|
| "What does the circular actually say?" | Gemini Notebook | Grounded, cited, verifiable |
| "Build the lease amortisation schedule" | ChatGPT or Claude | Mature code execution, returns a workbook you can file. Gemini Notebook can now compute too, but only on tiers that have the cloud computer |
| "Draft the management letter point" | ChatGPT or Claude | Better at structured drafting |
| "Compare twelve facility agreements" | Gemini Notebook | Data tables across a document set |
| "Which prior year issues are still open?" | Gemini Notebook | The permanent file is the source |
| "Explain this to the managing director" | ChatGPT or Claude | Better at register and tone |
| "Train the article students on IFRS 16" | Gemini Notebook | Flashcards, quizzes and audio from the standard itself |
The strongest workflow uses both. Find and verify the provision in Gemini Notebook, take the verified reference into a general assistant for the computation and the drafting, then bring the finished output back to your own review. Neither tool replaces the other, and neither replaces you.
The data position here is better than most AI tools, and the account you use decides how much better.
Copy, paste, replace the bracketed parts. Type in the box to filter.
Across the selected sources, list every place that deals with [topic]. For each: the source title, the section or paragraph reference, and one line on what it says there. Order them by how directly they address the topic. Do not summarise the topic itself. I only want a map of where to read.
Read the selected sources against each other and find every point where they conflict, overlap inconsistently, or use the same term differently. For each conflict: - Quote the competing words from each source, briefly, with references. - State plainly whether one supersedes the other on the face of the documents, or whether that is not determinable from what I have loaded. - Do not resolve the conflict yourself. Show it to me.
Working only from the selected circulars: 1. Which of them apply to [transaction or exposure type]? List with dates and references. 2. In date order, what did each one change, quoting the operative words. 3. Which is currently operative on the face of these documents, and which have been expressly superseded? Where supersession is not stated, say so. 4. What obligations does the current position impose on us, each one testable. 5. What I would need to load to be confident this list is complete.
From the selected agreements, build a table with one row per agreement: counterparty, date, term, financial covenants with their defined ratios and test dates, security given, events of default, change of control provisions, and the clause reference for each. Where an agreement is silent on a column, write "not stated". Do not infer a market standard. Where two agreements define the same ratio differently, add a note quoting both definitions.
Compare the accounting policy notes across the financial statements in the selected sources. 1. Every policy that changed wording between years, with both versions quoted and both references. 2. Every policy that appears in one year and not another. 3. Whether the statements explain the change, and where. 4. Which changes could affect comparability, and which look like drafting tidy-ups rather than substantive change. Do not conclude on whether any change was appropriate.
From the selected board and committee minutes for [period]: 1. Every approval of a transaction, borrowing, guarantee, related party arrangement or capital commitment, with date and reference. 2. Every matter recorded as deferred, unresolved or to be reviewed later. 3. Every mention of litigation, regulatory contact, investigation or material dispute. 4. Every reference to a going concern, liquidity or covenant discussion. 5. Anything approved that I should expect to see reflected in the accounts. Give me references for all of it. Do not assess whether anything was correctly accounted for.
Using only the standard loaded in the selected sources, produce a disclosure checklist for [standard]. One row per requirement: the requirement in plain language, the paragraph reference, and whether the standard makes it mandatory or conditional on circumstances. Separate the presentation requirements from the note disclosure requirements. Do not add anything from your general knowledge, and do not include requirements from other standards even where they usually travel together.
From the selected sources, build training material for [article students / new bank finance staff] on [topic]. 1. A one-page briefing on what the sources say, in plain language, cited. 2. Twelve flashcards covering the points people most often get wrong. 3. A ten-question quiz with answers and the source reference for each. 4. Three realistic Bangladeshi scenarios that test judgement rather than recall, with a model answer for each drawn from the sources. Keep every technical statement traceable. Where the sources do not settle a point of judgement, say so in the model answer.
I am about to advise on [matter], relying on the sources currently selected. Tell me what these sources do not address that a reviewer would expect me to have considered. Be specific: name the question, and say what kind of document would answer it. Do not attempt to answer those gaps from general knowledge. I want the list, not the fill.
Generate an audio overview of the selected sources, but focus it on: - What changed and when, in date order - The practical obligations it creates for a finance team - The three points a practitioner most often misunderstands Audience: qualified accountants in Bangladesh. Assume technical vocabulary is understood. Skip the general introduction to the topic and get straight to the substance.
No prompt matches that word. Clear the filter to see all of them.
Twenty minutes a day, on documents you already have to read. Progress is held in the page only, so it resets if you reload.
Bangladeshi statutory references in this manual are illustrative. Always work from the current text of the Income Tax Act 2023, the Value Added Tax and Supplementary Duty Act 2012, the operative Finance Act, applicable SROs, and current Bangladesh Bank circulars.
| Term | What it means in practice |
|---|---|
| Notebook | A workspace holding a bounded set of sources. Scope it to one subject. |
| Source | One document, page, video or audio file loaded into a notebook. |
| Grounding | Answering only from the loaded sources. The reason this tool suits professional work. |
| Citation chip | The clickable marker on a claim that opens the exact passage it came from. |
| Studio | The panel that turns sources into reports, tables, mind maps, audio, video and decks. |
| Audio Overview | A generated two-host discussion of your sources. Interactive mode lets you interrupt with a question. |
| Data Table | Structured extraction across many documents into rows and columns. |
| Deep Research | An investigation that imports both its report and the sources it cited into your notebook. |
| Discover sources | Describe a topic and it proposes sources to add to the notebook. |
| Secure cloud computer | The sandbox introduced with the July 2026 rename that lets a notebook write and run code against your sources. |
| Per-source ceiling | The size cap on a single file, which does not increase with your plan. |
