ChatGPT for Finance and Accounts Professionals From first login to firm-wide workflows: setup, model selection, data analysis, custom GPTs and review discipline.

ChatGPT is OpenAI's assistant, and the most widely used AI tool in the world. For a finance professional the important thing to understand on day one is that "ChatGPT" is not one model. Several sit behind the same box, they differ enormously in quality, and the cheap ones are the default.

The model picker is the most important control on the screen

A fast model answers instantly and is fine for tidying a paragraph. A reasoning model thinks before it replies and is the only sensible choice for a technical accounting question. Picking the wrong one is the single most common reason people conclude that AI is not useful for professional work.

What it is genuinely good at

• Data analysis on real files. Upload an Excel or CSV and it writes and runs Python to analyse it. This is the strongest feature in the product for accountants, and the most underused.
• Long document review. Annual reports, agreements, assessment orders, circulars.
• Drafting to a fixed structure. Memoranda, management letter points, disclosure notes, board narrative, client correspondence.
• Building small reusable tools. Custom GPTs that hold your firm's format and rules so a whole team gets consistent output.
• Deep research. A longer multi-source investigation that comes back with a sourced report.

Where it will hurt you if you are careless

• Bangladeshi statute. Its recall of the Income Tax Act 2023, the VAT and SD Act 2012, SROs and Bangladesh Bank circulars is incomplete and may be out of date. It can produce a section number that does not exist, stated with total confidence.
• Arithmetic done in prose. If it does not run the analysis tool, the numbers are predicted rather than calculated. Part 04 covers how to force real computation.
• The default model. A fast model will confidently give a shallow answer to a hard question and you will not be told that a better model was available.
• Training on your conversations. On individual plans your content may be used to improve the models unless you opt out. Part 09 explains why this is a professional issue and not an IT preference.

Twenty minutes of setup, most of it in two settings screens that almost nobody opens.

Step 1: Create the account

1. Go to chatgpt.com and sign up with an email address or a Google account. Use your firm email if the account will touch client work.
2. Install the desktop app and the mobile app from chatgpt.com/download. The desktop app matters because that is where the Work features and the Office extensions live.
3. Paid plans need an international credit card. Business and Enterprise can be invoiced.

Step 2: Pick the plan, and read the context column

Most comparisons focus on price and message limits. For our work the deciding number is how much text the model can actually take in at once, because that is what determines whether it can read a full annual report or only a third of it.

Step 3: The interface, labelled

These are annotated drawings rather than captured screenshots, so each control can be numbered and explained. Menu wording shifts between releases; the positions and the ideas are stable.

Step 4: Write your custom instructions once

In Settings, open personalisation and custom instructions. There are two boxes: what ChatGPT should know about you, and how it should respond. This travels with every new chat and is the highest-return five minutes in this manual.

I am a practising Chartered Accountant in Dhaka, Bangladesh, with 12 years of
banking experience. My work covers statutory audit, IFRS reporting, income tax
and VAT under Bangladeshi law, and bank regulatory reporting. I work in BDT.
My outputs go into audit files, tax returns, client reports and board papers,
so accuracy matters more than speed and everything I use gets reviewed.

- Professional but conversational. No marketing language, no filler, no
  restating my question back to me.
- Lead with the answer, then the reasoning. Short paragraphs.
- British spelling. Amounts in BDT with thousands separators.
- Cite the standard or section you rely on (IFRS, IAS, ISA, ITA 2023, VAT and
  SD Act 2012, Bangladesh Bank circular). If you are not certain a provision
  is current, say so explicitly rather than guessing.
- Whenever a calculation is involved, run it with the analysis tool and show
  the code. Never do arithmetic in prose.
- Tell me when my premise looks wrong. Do not agree by default.
- If a question needs the reasoning model to answer properly, say so before
  you answer.

Step 5: Set the data controls before you upload anything

In Settings, under data controls, find the option covering whether your content is used to improve the models. On Free, Go, Plus and Pro this is on by default with an opt-out available. On Business and Enterprise, OpenAI states that your content is not used for training. Make the decision deliberately, record it, and re-check it after major product updates.

A prompt is an instruction to a capable assistant who cannot see your screen, does not know your client, and will fill any gap you leave with an assumption.

The five-part structure

Six habits that separate confident users from frustrated ones

1. Choose the model first. Reasoning model for judgement, fast model for tidying.
2. Give it the source. Uploading the actual circular or trial balance beats relying on its memory every time.
3. Ask it to ask you. End with "ask me any questions before you start" and it will surface what you forgot to state.
4. Iterate rather than restart. "Too long, cut to 250 words and drop the introduction" beats rewriting the prompt.
5. Make it show workings. An answer you cannot audit is not usable on a professional file.
6. Ask for the counter-argument. "Now argue why this treatment is wrong" catches more errors than any other single instruction.

Before I use this, review your own output as a sceptical engagement partner
would. List, in order of severity:
1. Any figure you calculated in prose rather than computed with the analysis
   tool, or took from something other than my source data.
2. Any statement of law or standard you are less than fully confident is
   current, with what I should verify and where.
3. Any assumption you made that I did not give you.
4. Anything a reviewer would send back.
Do not rewrite the draft yet. Just give me the list.

This is where ChatGPT stops being a chat toy for an accountant. Attach a spreadsheet, tell it to analyse, and it writes and runs real Python on your data.

Make it compute, do not let it estimate

A language model predicting the next word is not a calculator. When the analysis tool runs, actual code executes and the arithmetic is real. Two ways to trigger it:

• Put the standing instruction in your custom instructions, as in Part 02.
• Say it in the prompt: "compute this with the analysis tool and show me the code".

Asking to see the code is not about reading Python. It is about checking which rate was applied, how the days were counted, and whether the closing balance ties.

What you can attach, and what comes back

• Excel and CSV: trial balances, ledgers, ageing schedules, fixed asset registers, loan portfolios.
• PDF: annual reports, circulars, agreements, assessment orders, scanned vouchers.
• Word: draft financial statements, prior year memoranda, engagement letters.
• Images: a photo of a handwritten cash book, a screenshot of an error in the NBR portal.

It returns Excel files, Word documents, PowerPoint decks, charts, and interactive tables you can sort in the chat. Ask for the format you want.

Attached is a two-year trial balance for [client], [industry], year ended
[date]. Materiality is BDT [amount].

Use the analysis tool so the arithmetic is computed rather than estimated,
and show me the code.
1. Build a variance table: account, prior year, current year, movement in BDT
   and percent. Sort by absolute movement.
2. Flag every account where the movement exceeds materiality OR exceeds 25
   percent, whichever captures more.
3. For each flagged account, give the two most likely legitimate business
   explanations and the two most likely misstatement risks, with the relevant
   assertion.
4. List accounts that moved suspiciously little, which can indicate a rolled
   forward balance nobody reconciled.
5. Identify apparent classification errors between current and non-current,
   or between cost of sales and operating expenses.

Do not conclude on any item. Give me a review agenda, not an opinion.
Return the variance table as an Excel file.

Canvas

Canvas opens a document beside the chat that you and ChatGPT edit together, rather than regenerating the whole reply each time. Use it for anything you will iterate on: a memorandum, a report section, a policy note. You can edit directly, highlight a paragraph and ask for a change to that part only, and keep the rest untouched.

Reading a long document properly

Attached is a [lease / facility / service] agreement. Use the reasoning
model. Extract into a table, with the clause number beside each item:
- Parties, commencement, term, renewal and termination options
- All payment obligations, amounts, escalation clauses and payment dates
- Any variable consideration or contingent payment
- Security, guarantees and covenants, including financial covenants with
  their defined ratios
- Change of control, penalty and default provisions

Then, separately: which clauses drive the treatment under IFRS 16 and
IFRS 9, and what figures I would need to extract to build the schedule.
Quote nothing longer than a phrase. Give clause references so I can verify
every line myself.

Twelve places where this pays for itself in the first month. Each is a prompt pattern to adapt, not a magic button.

Financial reporting

• Disclosure checklists. Upload the draft financial statements, ask for a gap analysis against a named standard, then run your firm's own checklist over the result.
• Technical position papers. Give the facts, ask for the treatment, then ask it to argue the opposite. The gap between the two answers is where your real work sits.
• Note drafting. Feed it the schedule and the prior year note, ask it to update the wording and keep house style.
• Consolidation logic. Testing elimination entries, non-controlling interest computations, and translation of a foreign operation.

Audit

• Risk assessment. Industry risks, entity-specific risks, and the assertions each one touches.
• Working paper drafting. Purpose, work performed, results, conclusion, in your firm's structure, using the numbers you supply.
• Management letter points. Condition, criteria, cause, effect, recommendation. It is very good at this format and it is tedious for humans.
• Journal entry testing. Upload the journal listing and have it run the filters: entries after period end, round sums, unusual users, rare account pairings, entries just under an approval threshold.

Turn this raw finding into a management letter point in condition, criteria,
cause, effect, recommendation format.

Raw finding: [what you found, including sample size, exceptions, amounts and
the period]

Rules:
- Criteria must cite the specific control objective or standard, not a vague
  reference to "best practice".
- Quantify the effect in BDT where my facts allow it, and write "not
  quantified" where they do not. Do not invent a number.
- The recommendation must be specific enough for the client to implement
  without asking us what we meant.
- Rate it high, medium or low and justify the rating in one sentence.
- Neutral professional tone, no blame language, under 220 words.

Corporate finance and management reporting

• MIS commentary. Give it the variance table, ask for board narrative, then check it has not invented a cause.
• Ratio analysis with meaning. Not just the ratios, but which three matter for this business and why the rest are noise.
• Model review. Upload a projection model and ask which assumptions drive the answer and where the model breaks.
• Due diligence question lists. Tailored to the target's industry and the deal structure.

The highest value and highest risk area in this manual. The value is in structure, computation and drafting. The risk is in statutory recall.

The safe pattern: give it the law

Use the reasoning model. I have attached [the relevant sections / the SRO /
the circular]. Work only from the attached text. If something needed for the
answer is not in the attachment, say "not covered in the provided text" and
tell me exactly what else to give you. Do not fill gaps from your general
knowledge of tax law.

Facts: [the transaction, amounts, dates and parties]

Questions:
1. Which provision in the attached text applies, by section and sub-section
   reference only?
2. What is the resulting treatment, computed step by step with the analysis
   tool?
3. What is the strongest argument the tax authority could make against this
   position?
4. What documentation should be on file to support it?

Where it works well without statutory recall

• Computation schedules. Taxable income build-up, depreciation on the basis you supply, minimum tax comparison, advance tax instalment scheduling.
• VAT reconciliation. Reconciling sales per accounts to filed returns to the general ledger, and listing the likely cause of each gap.
• TDS and VDS working papers. You supply the current rate matrix, it applies it consistently across a payment listing and produces the deduction schedule.
• Notice and assessment order review. Extracting every ground of assessment into a table with amounts and paragraph references so you can plan the appeal.
• Appeal and ADR drafting. Structure, sequencing of grounds, and consistency between statement of facts and grounds of appeal.
• Client explanation letters. Turning a technical position into something a managing director will actually read.

Banking and financial institutions

• IFRS 9 ECL. Testing staging logic, explaining a probability weighted outcome to a credit committee, reviewing the internal consistency of a provisioning matrix, and drafting model documentation. Supply your own PD and LGD assumptions, never ask it to produce them.
• Regulatory circular digestion. Upload a Bangladesh Bank circular and ask for a one-page impact note.
• Credit appraisal. Structuring the financial analysis, stress testing borrower projections, and listing covenants that would actually protect the bank given this borrower.
• Prudential reporting. Explaining the mechanics, reconciling reported figures to the ledger, and drafting the narrative that accompanies a return.

Attached is a circular issued by [regulator] dated [date]. Produce a
one-page impact note for our [finance / credit risk / compliance] committee:

1. What changed, in plain language, but only where the circular itself states
   the previous position. Where it does not, say so.
2. Effective date and any transitional relief, quoted by paragraph.
3. Which functions are affected and what each must do differently.
4. What our systems and reports must capture that they may not capture today.
5. A compliance checklist of no more than eight items, each one testable.
6. Open questions to raise with the regulator or our lawyers.

Work only from the attached circular. Do not rely on your own knowledge of
prior circulars. List anything you need under "documents to obtain".

Everything so far has been one conversation at a time. This part is about making ChatGPT carry context so you stop re-explaining yourself, and about making a whole team produce consistent output.

Projects: a workspace per client or subject

A Project holds files and standing instructions, and every chat started inside it inherits both. Available across plans, with shared projects on Business and Enterprise.

Custom GPTs: your firm's methodology, saved

A custom GPT is a saved assistant with fixed instructions and attached reference files. Build it once, use it forever, share it with the team. This is the feature that turns individual productivity into firm capability, and it is the one most firms never get to.

Four worth building in your first month:

• Management letter drafter, holding your format, tone rules and rating scale.
• Working paper reviewer, holding your firm's review checklist and ISA references.
• Bangladesh tax reference, holding the current Act text, the operative Finance Act and the SROs you rely on, with instructions to answer only from those files.
• Client communication translator, turning technical positions into plain language for a non-accountant.

You are a Bangladesh tax reference assistant for a chartered accountancy
practice. You answer only from the files attached to this GPT.

Rules you never break:
1. Every substantive answer cites the section, sub-section, SRO number or
   paragraph in the attached files. No citation, no answer.
2. If the attached files do not cover the question, say "not covered in the
   loaded sources" and list what should be uploaded. Never answer from
   general knowledge of tax law, including law from other countries.
3. Never state a rate, threshold or due date that is not in the files.
4. Compute with the analysis tool and show workings. Never do arithmetic in
   prose.
5. Where the files contain provisions from different years, say which year
   you are applying and warn me to confirm it against the operative
   Finance Act.
6. End every answer with: "Verify against the current text before use."

Memory, scheduled tasks and apps

• Memory carries context across chats. Review what has been retained in Settings and clear anything client-specific that should not persist.
• Scheduled tasks run a prompt on a schedule, for example a Monday morning summary of a regulator's published updates.
• Apps and connectors let it read from Drive, SharePoint, Slack and dozens of other tools. Business plans carry 60 plus.
• Company knowledge on Business and Enterprise searches across your connected workspace rather than one attached file.

The gap between an average user and a strong one is almost entirely this: strong users make the machine compute, and they automate anything they do twice.

Deep research

Deep research runs a longer multi-source investigation and returns a report with citations. It genuinely helps with market and industry work, regulatory landscape reviews, and comparator analysis. It does not replace reading the primary source on a technical accounting or tax question, and every citation still needs opening.

Research question: [state it in one sentence]

Scope:
- Jurisdiction: [Bangladesh / regional / global], period: [dates]
- Include: [regulator publications, listed company financial statements,
  industry association data]
- Exclude: vendor marketing, undated blog posts, aggregator sites

Deliverable:
1. Findings, each one attributed to a source I can open
2. Where sources disagree, show the disagreement rather than resolving it
3. A table of the quantitative data points with source and date beside each
4. What you could not find, stated explicitly rather than filled in
5. The three questions this research raises that I should investigate next

Do not estimate any figure. If a number is not in a source, mark it as
"not found".

The Office extensions

On Plus and above, ChatGPT works inside Excel, PowerPoint and Google Sheets. For someone who lives in a spreadsheet, this is the highest-impact single feature in the product: it reads the live workbook, traces where a number comes from, explains a formula chain, builds schedules and fixes broken references. Review every change before accepting it, exactly as you would review a junior's edit.

Agentic work: ChatGPT Work and Codex

• ChatGPT Work handles multi-step jobs on the desktop rather than answering one message at a time.
• Codex is the coding agent. For a finance team this is how you build genuinely useful internal tools: a script that parses 400 bank statements, a VAT reconciliation utility, a Google Apps Script automation.
• Record mode captures a meeting and turns it into notes and actions. Get consent from everyone in the room first, and check your firm's position on recording client meetings.

The API, and what a firm should expect to pay

A subscription covers you using ChatGPT. The API is for building the model into something else, and is billed per token, entirely separately. A token is roughly three quarters of an English word.

Four costed illustrations

Computed, not estimated. Assumptions stated so you can re-run them with your own volumes.

A Chartered Accountant's duty of confidentiality has no exception for convenient software. With ChatGPT this section carries extra weight, because on individual plans the training default runs against you.

The five decisions to make at partner level

1. What data may leave the firm. Anonymised extracts only, or identified client data under an appropriate plan and terms. Write the policy in one page.
2. Which plan. Business at two seats costs about the same per head as Plus and removes the training default. That comparison usually ends the discussion.
3. Client consent and engagement letters. Consider whether your terms should disclose AI use. Banks and listed clients will ask.
4. Which apps are connected. A connected mailbox is a far bigger exposure than one uploaded PDF.
5. Retention. Know where conversations live, how long they persist, and who can export them.

Practical anonymisation

For most analytical work you do not need identifying data. Replace the client name with "the Company", strip the TIN, BIN and account numbers, keep the figures. The analysis is identical and the exposure is far lower.

Professional obligations that do not change

• Confidentiality under the IESBA Code as adopted by ICAB applies to anything you paste, upload or connect.
• Professional competence and due care means you remain responsible for the output. "The AI produced it" is not a defence and never will be.
• Audit documentation under ISA 230 requires the file to show the work performed. If AI assisted a procedure, the file should show what was done, what evidence was obtained, and who reviewed it.
• Independence and objectivity. A tool that agrees with you enthusiastically is a risk to objectivity, not a support for it.

Copy, paste, replace the bracketed parts. Type in the box to filter.

Attached: the bank statement and the cash book for [period]. Use the
analysis tool and show the code.
1. Match transactions by amount and date, allowing a tolerance of [n] days.
2. List unmatched bank items and unmatched book items separately.
3. Build the reconciliation statement from book balance to bank balance.
4. Flag anything unusual: round sums, transactions on non-working days,
   payments to one party split just below [approval threshold], reversals.
The reconciliation must tie exactly. If it does not tie, tell me the
difference rather than forcing it. Return the working as an Excel file.

Use the reasoning model. Attached: our loan portfolio extract and our
staging policy document.

Working only from my attached policy, and computing with the analysis tool:
1. Apply the staging criteria to each exposure and produce a stage summary
   by count and by exposure amount.
2. Identify exposures where the policy is ambiguous or two criteria conflict,
   and list them for manual judgement.
3. Recompute ECL using the PD, LGD and EAD I supplied. Do not supply your own
   risk parameters under any circumstances.
4. Reconcile your total to our reported provision and explain each difference.
5. Draft the model documentation section describing the staging methodology,
   suitable for our audit file and regulatory review.

Attached: monthly sales per the general ledger and the VAT returns filed for
the same period.

Using the analysis tool, build a reconciliation from accounting turnover to
declared taxable turnover, month by month, showing each reconciling item
separately: exports and zero rated supplies, exempt supplies, non-taxable
income, timing differences on advances, credit notes, related party
transfers, and anything else the data supports.

For each unexplained difference, list the three most likely causes ranked by
how often they occur in practice, and the document I should request to
resolve it. Do not tell me the VAT treatment of anything unless I have given
you the provision.

Use the reasoning model. Attached: an assessment order for [taxpayer],
assessment year [year].

Produce a table with one row per addition or disallowance, columns:
paragraph reference, description, amount, the reason the officer gave, and
the evidence the order says was considered.

Then, separately:
- Which additions rest on a factual finding we could rebut with documents,
  and what documents would be needed.
- Which rest on an interpretation of law, quoted from the order itself.
- The order in which grounds of appeal should be argued, strongest first,
  with one sentence on why.

Work only from the attached order. Do not assess the merits under any
provision I have not given you.

Attached: management accounts for [period] with budget and prior year
comparatives.

Write the finance commentary for the board pack:
- Open with the three things the board must know, under 80 words total.
- Then revenue, margin, cost and cash, one short paragraph each, each
  anchored to a specific figure from the attachment.
- Explain variances only where the data supports an explanation. Where it
  does not, write "driver not identifiable from the data provided".
- Close with the decisions the board is being asked to make.
Plain language, no adjectives that are not doing work, under 600 words.

Use the reasoning model.
Facts: [the transaction in full, with dates, amounts and contract terms]

Part 1: The accounting treatment under [standard], with the paragraph
references you rely on. Flag any reference you are not fully confident about
and tell me to verify it.

Part 2: Now act as the reviewing partner who disagrees. Write the strongest
argument for a different treatment.

Part 3: What additional facts or documents would settle the question, and
what the disclosure would need to say under each treatment.

Do not reconcile the two positions. I want the disagreement visible.

Attached workbook. Before changing anything:
1. Map the calculation chain: which sheet feeds which, and where the inputs
   actually live.
2. List every hard-coded number sitting inside a formula, with its cell
   reference. These are the errors waiting to happen.
3. List broken references, inconsistent formulas within a row or column, and
   any circular logic.
4. Tell me which three fixes carry the highest risk of changing a reported
   figure, so I can check those myself first.
Then wait for my instruction before you edit anything.

Attached: the full year journal listing. Use the analysis tool and show the
code for every test.

Run and report separately:
1. Entries posted after the period end date but dated within the period.
2. Round-sum entries above BDT [threshold].
3. Entries posted by users who post fewer than [n] entries a year.
4. Entries posted outside working hours or on public holidays.
5. Account pairings that occur fewer than [n] times in the year.
6. Entries just below BDT [approval threshold].

For each population give the count, the total value, and the ten largest
items. Do not conclude on any entry. This is a selection basis, not a
finding.

Rewrite the attached technical note for [the managing director / the audit
committee chair / a family shareholder], who is intelligent but not an
accountant.

Rules:
- Lead with what it means for them: the cash effect, the reported profit
  effect, and the decision they need to make.
- No standard numbers or section references in the body. Footnote them.
- No jargon without a plain definition on first use.
- Keep every figure exactly as it appears in the source. Change no number.
- Under 350 words, ending with the one question I need them to answer.

This prompt gave me a weak answer:
"[paste your prompt]"

The answer was weak because: [too generic / wrong format / invented figures
/ missed the point / used the fast model when it needed reasoning].

Rewrite the prompt so it works. Show me the improved version, then explain in
three bullets what you changed and why, so I can apply the same thinking
myself next time.

No prompt matches that word. Clear the filter to see all of them.

Thirty minutes a day on real work, not practice work. Progress is held in the page only, so it resets if you reload.

Week 1: foundations

• Create the account, install desktop and mobile apps
• Write both custom instruction boxes, including the compute rule
• Open data controls, decide your training position, and record the decision
• Ask the same hard question on the fast model and the reasoning model, and compare
• Upload one real PDF and extract something you would otherwise have hunted for

Week 2: your actual work

• Run the trial balance analytical review on a live engagement
• Read the code it wrote for one analysis and check the logic yourself
• Draft one management letter point and compare it with your own version
• Use the self-review prompt on every draft this week
• Deliberately catch it being wrong once, and note how you caught it

Week 3: context and structure

• Set up a Project for one client with instructions and reference files
• Build the source-bound Bangladesh tax GPT and load the current provisions
• Run a source-bound statutory analysis and verify every reference yourself
• Try the Excel extension on a real workbook
• Decide your firm position on connected apps and write it down

Week 4: leverage

• Build one custom GPT your whole team can use
• Run one deep research brief and open every citation
• Write your firm's one-page AI usage policy covering the five partner decisions
• Train one colleague using this document and the prompt library
• Re-read Part 09 and check your own last month of use against it

Primary sources used in this manual

• Individual plans, model availability and context windows: chatgpt.com/pricing, accessed 29 July 2026
• Business and Enterprise plans, security and privacy commitments: openai.com/business/chatgpt-pricing, accessed 29 July 2026
• Help centre and usage limits: help.openai.com
• Developer documentation and API pricing: developers.openai.com
• Enterprise privacy: openai.com/enterprise-privacy
• Apps: chatgpt.com/download

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.

Glossary