Claude AI for Finance and Accounts Professionals From first login to firm-wide workflows: setup, prompting, review discipline and a working prompt library.

Claude is a general purpose AI assistant built by Anthropic. For a finance professional it behaves less like a search engine and more like a very fast, very well-read junior who has read everything you give it, never gets tired, and occasionally states something wrong with complete confidence. Everything in this manual follows from that last point.

What it is genuinely good at

• Reading long documents. A 180-page annual report, a loan agreement, a Bangladesh Bank circular, three years of board minutes. It will pull out what you ask for in seconds.
• Drafting in a fixed structure. Management letter points, audit memoranda, disclosure notes, credit appraisal narratives, board papers, client emails.
• Turning mess into structure. A pasted trial balance, an unformatted bank statement, a list of fixed asset purchases, all converted into a clean schedule.
• Explaining and challenging. Talking through IFRS 9 staging logic, or arguing the other side of a position you have taken so you can test it before the client does.
• Building small tools. Excel formulas, Power Query steps, Google Apps Script, Python for a data set that is too big for Excel.

Where it will hurt you if you are careless

• Local statute and case law. Its recall of the Income Tax Act 2023, the VAT and SD Act 2012, SROs and Bangladesh Bank circulars is incomplete and can be out of date. It may produce a plausible section number that does not exist.
• Arithmetic it does in its head. Language models are not calculators by default. The fix is in Part 08: make it compute, do not let it estimate.
• Confidential client data. Covered in Part 09. This is a professional obligation, not an IT preference.
• Sounding right. Fluent, well-formatted output creates a feeling of reliability that the content has not earned. Your professional scepticism applies to AI output exactly as it applies to a client schedule.

The model family, in plain terms

Anthropic ships several models. You choose one from a dropdown in the chat box. For finance work you rarely need to think harder than this:

Practical habit: start on Sonnet 5. If the answer feels shallow on a technical question, switch to Opus 5 and ask again in the same conversation.

Fifteen minutes of setup saves you from the two most common beginner complaints: "it keeps giving me generic answers" and "I have to explain my firm every single time".

Step 1: Create the account

1. Go to claude.ai and sign up with an email address. Use your firm email if the account will be used for client work.
2. Verify the email, then set your name.
3. Check availability for your country first if you are unsure, at anthropic.com/supported-countries. Paid plans need an international card or an app store subscription.
4. Install the apps you will actually use: desktop (Windows or macOS) and mobile (iOS or Android) from claude.com/download. The desktop app matters more than people expect, because it can connect to files and other applications on your machine.

Step 2: Pick the right plan

You can learn everything in Parts 03 to 05 on the free plan. You will hit its limits within a week of real use.

Step 3: The interface, labelled

Below are annotated diagrams of the current layout. They are drawings, not captured screenshots, so that each control can be numbered and explained. Menu wording shifts between releases; the positions and the ideas are stable.

Step 4: Set your profile once

In Settings, find your profile and preferences. Write who you are and how you want answers. This travels with every new conversation and is the single 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.

How I want you to respond:
- Professional but conversational. No marketing language, no filler.
- Lead with the answer, then the reasoning. Short paragraphs.
- Use British spelling and present amounts in BDT with thousands separators.
- Cite the standard or section you are relying on (IFRS, IAS, ISA, ITA 2023,
  VAT and SD Act 2012, Bangladesh Bank circular). If you are not certain the
  provision is current, say so explicitly instead of guessing.
- Whenever a calculation is involved, compute it with the analysis tool and
  show the workings. Never estimate arithmetic.
- Challenge me when my premise looks wrong. Do not agree by default.

Step 5: Set your privacy position before you upload anything

In Settings, review the data controls, including whether your conversations may be used to improve the models. On Team and Enterprise plans that is off by default. On individual plans, check the setting yourself and record the decision. Part 09 covers what this means for client confidentiality.

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

The five-part structure

Compare the two

Six habits that separate confident users from frustrated ones

1. One task, one conversation. Start a new chat when the subject changes. Threads that drift lose accuracy.
2. Give it the source. Uploading the actual circular, contract or trial balance beats relying on its memory every single time.
3. Ask it to ask you. End with "ask me any questions before you start". It will surface the assumptions you forgot to state.
4. Iterate rather than restart. "Too long, cut to 250 words and drop the introduction" is faster than rewriting the prompt.
5. Make it show workings. An answer you cannot audit is not usable in 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 rather than took from 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 an AI assistant stops being a novelty and starts saving real hours. You attach the document. It reads all of it.

What you can attach

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

What it can produce back

Claude can create and edit real files: Excel workbooks with formulas, Word documents, PowerPoint decks, PDFs, and interactive HTML tools. Ask for the format you want. "Give me this as an Excel file with the working columns preserved" is a valid instruction.

A worked routine: the trial balance review

1. Export the current year and prior year trial balance to one Excel file, with account code, description, and both balances.
2. Attach it and use the prompt below.
3. Take the flagged items into your own review, do not accept the conclusions.

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

Using the analysis tool so the arithmetic is computed rather than estimated:
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, state the two most likely legitimate business
   explanations and the two most likely misstatement risks, referencing the
   relevant assertion.
4. List accounts that moved suspiciously little, which can indicate a rolled
   forward balance nobody reconciled.
5. Identify anything that looks like a classification error 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.
Ask me for any missing context before you begin.

Reading a long document properly

For a 200-page document, do not ask "summarise this". Ask targeted questions, one at a time, and ask for the page or clause reference so you can go and read it yourself.

Attached is a [lease / facility / service] agreement. Extract into a table,
with the clause number beside each item:
- Parties, commencement date, term, and any renewal or 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 accounting treatment under
IFRS 16 and IFRS 9, and what specific figures would I need to extract to
build the schedule. Quote nothing longer than a phrase. Give me clause
references so I can verify each line myself.

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

Financial reporting

• Disclosure checklists. Upload your draft financial statements and ask for a gap analysis against a named standard. Treat the output as a first pass, then run your firm's own checklist.
• Technical position papers. Give it the facts, ask for the treatment under the specific standard, then ask it to argue the opposite position. The disagreement between the two answers is where your real work is.
• Note drafting. Feed it the schedule and the prior year note, ask it to update the wording and keep the house style.
• Consolidation logic. Explaining and testing elimination entries, non-controlling interest calculations, 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, with the actual 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 logic. Ask it to write the filters: entries posted after period end, round-sum entries, entries by unusual users, entries with rare account pairings.

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

Raw finding: [describe what you actually found, including sample size,
exceptions found, 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 say "not quantified"
  where they do not. Do not invent a number.
- The recommendation must be specific enough that a client can implement it
  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 the board narrative. Always check that it has not invented a cause.
• Ratio analysis with meaning. Not just the ratios, but which three matter for this business and why the others are noise.
• Model review. Upload a projection model and ask what assumptions drive the answer and where the model breaks.
• Due diligence question lists. Tailored to the target's industry and the specific deal structure.

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

The safe pattern: give it the law

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: [set out the transaction, amounts, dates and parties]

Questions:
1. Which provision in the attached text applies, quoted 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 under the third schedule logic you supply, minimum tax comparison, advance tax instalment scheduling.
• VAT reconciliation. Reconciling sales per accounts to Mushak 9.1 returns to the general ledger, identifying the gaps and listing the likely causes 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. Upload the order and ask it to extract every ground of assessment into a table, with the amount in dispute and the paragraph reference, so you can plan the appeal.
• Appeal and ADR drafting. Structure, sequencing of grounds, and consistency between the statement of facts and the 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 your staging logic, explaining a probability weighted outcome to a credit committee, reviewing the internal consistency of a provisioning matrix, and drafting the model documentation. Give it your own PD and LGD assumptions, never ask it to supply them.
• Regulatory circular digestion. Upload a Bangladesh Bank circular and ask for a one-page impact note: what changed, effective date, which departments are affected, what a compliance checklist would contain.
• Credit appraisal. Structuring the financial analysis section, stress testing the borrower's projections, listing the covenants that would actually protect the bank given this borrower's risk profile.
• Basel and prudential reporting. Explaining the mechanics, reconciling reported figures to the underlying 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, against what applied before, 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 we should raise with the regulator or our lawyers.

Work only from the attached circular. Do not rely on your own knowledge of
prior circulars. Where you need a cross reference I have not given you, list
it under "documents to obtain".

Everything so far has been one conversation at a time. This part is about making Claude carry context so you stop re-explaining yourself.

Projects: a workspace per client or per subject

A Project is a container that holds reference files and standing instructions. Every chat you start inside it inherits both. Available on paid plans, unlimited on Pro and above.

Projects worth setting up on day one

• One per significant client or engagement, holding prior year accounts and the current audit programme.
• One called "Firm templates", holding your engagement letter, management letter format, and report templates.
• One called "Bangladesh tax reference", holding the current Act text, relevant SROs and the Finance Act you rely on. This is how you get statutory answers you can trust: the source lives in the Project.
• One called "Bank regulatory", holding the circulars your reporting depends on.

Memory

Claude can retain useful context across conversations, so it remembers your role, your recurring projects and your preferences. Review what it has retained in Settings, and clear anything client-specific that should not persist. Memory is a convenience feature, not a document management system.

Connectors

Connectors let Claude read from applications you already use, including Google Drive, Gmail, Google Calendar and Microsoft 365, and any other tool exposed through the connector directory. Claude for Chrome and Claude for Microsoft 365 extend the same idea into the browser and Office.

Research

On paid plans, Research runs a longer multi-source investigation and returns a sourced report. Useful for market data, industry benchmarking and regulatory landscape reviews. Not a substitute for reading the primary source on a technical accounting or tax question.

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

Make it compute, do not let it estimate

Claude can execute real code in a sandbox to do analysis. When it does, the arithmetic is performed by a computer, not predicted by a language model. Two ways to trigger it:

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

Asking to see the code is not about reading Python. It is about being able to check the logic: which rate was applied, how the days were counted, whether the closing balance ties.

Build a [lease amortisation / loan repayment / depreciation / ECL] schedule
from the attached data.

Requirements:
- Use the analysis tool. Show me the code you ran.
- State every assumption in a list before the schedule, and stop to confirm
  them with me if any assumption is material to the result.
- Produce the schedule as an Excel file with the working columns intact, not
  just the final figures.
- Add a proof row: the schedule must tie back to [total cost / opening
  liability / gross exposure]. Show the tie.
- If any input is missing or internally inconsistent, stop and tell me
  rather than assuming a value.

Claude in Excel

Claude works inside a spreadsheet as a side panel: it can read your workbook, explain a formula chain, trace where a number comes from, build new schedules and fix broken references. For anyone who lives in Excel, this is the highest impact single feature in the product.

Cowork, Claude Code and Skills

• Claude Cowork is an agentic desktop application for knowledge work: give it a folder of files and a multi-step job, and it works through the whole thing rather than answering one message. It is included on every plan; installing plugins requires a paid plan.
• Claude Code is the terminal and IDE tool. For a finance team this is how you build genuinely useful internal tools: Google Apps Script automations, a VAT reconciliation web app, a fund accounting tool, a Python script that parses 400 bank statements.
• Skills are reusable instruction packs. A firm can build a Skill for "our management letter format" or "our IFRS 9 model documentation standard" and deploy it across the organisation on Team and Enterprise plans.

The API, and what a firm should expect to pay

A subscription covers you using Claude. The API is for building Claude into something else: a client portal, an internal review tool, a bulk document processor. It is billed per token, separately from any subscription. A token is roughly three quarters of an English word.

Three costed illustrations

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

A Chartered Accountant's duty of confidentiality does not have an exception for convenient software. Settle this before the tool spreads through your firm informally, because it will.

The five decisions to make at partner level

1. What data may leave the firm. Decide the categories: anonymised extracts only, or identified client data under an appropriate plan and terms. Write the policy in one page.
2. Which plan. Team and Enterprise carry no model training on your content by default, plus administrative control. Individual accounts spread across staff give you neither.
3. Client consent and engagement letters. Consider whether your engagement terms should disclose the use of AI tools. Some clients, particularly banks and listed companies, will ask.
4. Which connectors are allowed. A connected mailbox is a much bigger exposure than a single uploaded PDF.
5. Retention. Know where conversations live, how long they are retained, and who can export them.

Practical anonymisation

For most analytical work you do not need identifying data at all. Replace the client name with "the Company", strip the TIN, BIN and account numbers, and 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 are unaffected by tooling. 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]. Using the
analysis tool:
1. Match transactions by amount and date, allowing a tolerance of [n] days.
2. List unmatched bank items and unmatched book items separately, with dates
   and amounts.
3. Build the reconciliation statement from book balance to bank balance.
4. Flag anything unusual: round sums, transactions on non-working days,
   payments to the same party split just below any approval threshold I give
   you, and reversals.
Show the code. The reconciliation must tie exactly. If it does not tie,
tell me the difference rather than forcing it.

Attached: our loan portfolio extract and our staging policy document.

Using only 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 where two criteria
   conflict, and list them for manual judgement.
3. Recompute ECL using the PD, LGD and EAD I have 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 specific document I should
request to resolve it. Do not tell me the VAT treatment of anything unless
I have given you the provision.

Attached: an assessment order for [taxpayer], assessment year [year].

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

Then, separately:
- Which additions appear to 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, in 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" instead
  of speculating.
- Close with the decisions the board is being asked to make.
Plain language. No adjectives that are not doing work. Under 600 words.

Facts: [set out the transaction fully, with dates, amounts and contract terms]

Part 1: Set out the accounting treatment under [standard], with the specific
paragraph references you are relying on. Flag any paragraph 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: Set out 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.

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. Put them in a
  short footnote instead.
- 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, and end with the one question I need them to answer.

Create a practical exercise for [article students / junior bank finance
staff] on [topic].

Structure:
- A realistic Bangladeshi scenario with figures, one paragraph.
- Four questions rising in difficulty from recall to judgement.
- A worked answer for each, with the arithmetic computed using the analysis
  tool so I can verify it.
- Three common mistakes and why each one happens.
- A one-page summary they can keep.
Use BDT throughout. Flag any tax or VAT rate you use as illustrative and
tell me to confirm it against the current Finance Act.

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

The answer was weak because: [too generic / wrong format / invented figures
/ missed the point].

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. Tick items as you go. 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 your profile and preferences, including the compute instruction
• Check the model training and data controls in Settings and record your decision
• Run five prompts using the five-part structure and compare them with lazy versions
• 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
• Draft one management letter point and compare it with your own version
• Build one schedule with the analysis tool and re-perform it in Excel to confirm it ties
• 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
• Set up a Bangladesh tax reference Project holding the source provisions
• Run a source-bound statutory analysis and verify every reference yourself
• Try Claude in Excel on a real workbook
• Decide your firm position on connectors and write it down

Week 4: leverage

• Build one reusable tool: a checklist, a calculator, or a template generator
• Write your firm's one-page AI usage policy covering the five partner decisions
• Train one colleague using this document and the prompt library
• Identify the one recurring task in your practice worth automating properly
• Re-read Part 09 and check your own last month of use against it

Primary sources used in this manual

• Plans, features and API pricing: claude.com/pricing, accessed 29 July 2026
• Product help and how-to articles: support.claude.com
• Developer documentation: platform.claude.com/docs
• Prompt engineering guidance: docs.claude.com prompt engineering overview
• Country availability: anthropic.com/supported-countries
• Product news and release notes: anthropic.com/news
• Apps: claude.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