AI for finance professionals: 8 month-end tasks for ChatGPT, Claude and Claude Code

In short

AI for finance professionals is already useful in the least glamorous part of the job: the month-end close. A model can code a bank statement to your chart of accounts, match it against the general ledger, explain budget-vs-actual variances, build a 13-week cash view, write Excel formulas, and draft a plain-language memo for the owner. Approvals, journal entries, and the signature on the numbers stay with a person.

Start in a chat tool such as ChatGPT or Claude. When the same work repeats every month, move it into Claude Code or Codex. These are agents that work on a folder of files on your own computer: they write the calculation, run it, and save the result. You can open that calculation, check it, and run it again next month.

Below are eight tasks walked through one close. The company is hypothetical: a mid-size wholesale distributor with two operating accounts, about 600 bank transactions a month, QuickBooks Online or Xero as the ledger, a 40-line budget-vs-actual in Excel, and an owner who wants a short memo by business day five. The before/after figures in each task are our illustrative estimates for this example, not measured client results.

What AI can and cannot do in finance today

Current models read spreadsheets, PDFs, and scans well, make sense of messy payment descriptions, and handle Excel confidently. Give them a file and they calculate rather than recall: both ChatGPT and Claude run code in the chat to work through a table. They also write well, so turning ten lines of variances into a paragraph an owner will actually read is easy for them.

Where they help:

• suggest an account for every bank line
• pair bank transactions with ledger entries and list what is left unmatched
• compute budget-vs-actual variances and group them by cause
• draft a cash forecast from contracts and expected receipts
• pull payment terms, late fees, and deposit requirements out of contracts
• write a formula, pivot, or macro from a plain description
• explain the month in plain language

Where they are still weak:

• Accuracy without checks. A model can get one row in five hundred wrong and not tell you. Every total needs a check against the source, and a calculation you can rerun is better than an answer in a chat window.
• Accounting and tax judgment. The model will happily discuss whether a cost can be capitalized. The answer still belongs to your accountant or advisor, with the standard in hand.
• Knowing your company. It does not know that "Arden Supply" and "Arden Logistics LLC" are the same vendor, that warehouse rent is due on the tenth, or that marketing moved to a different cost center this year. You have to tell it.
• System access. A chat tool cannot log into your ERP or bank portal. Data reaches it as exports. A dependable connection to the ledger is a system project, covered at the end.

A useful rule: AI prepares, a person approves. Anything that moves money or changes the books goes through you.

Which tools to use

For most finance people the choice comes down to five options. The first three are chat tools. The last two are agents that work with files on your machine.

If you are just starting, pick ChatGPT or Claude on a work plan. Move to Claude Code or Codex when the same steps repeat every month.

8 tasks, one month-end close

The example is the September close. The tasks follow the order of the work: bank, reconciliation, variances, next month's cash, contracts, Excel, memo, forecast.

Before the first prompt, strip what the model does not need: employee IDs, card numbers, salaries by name. More on that in the security section.

1. Coding the bank statement

The September export has about 600 lines with descriptions like "PMT INV 118" or "RFND EXP". Before, the analyst went through it by hand and typed an account in the next column.

Here is the September bank export (Excel) and our chart of accounts with examples. Suggest an account for every line with a one-line reason. If the account is not obvious, write "unclear" instead of guessing. Return a table: date, counterparty, amount, account, confidence (high, medium, low).

Result: 80–90% of lines coded with high confidence, the rest flagged for review. You check the flagged lines and spot-check the confident ones.

Before/after (estimate): about 4 hours of manual coding versus 1 hour of review.

Tip: attach last month's coded statement. The model learns your habits from it faster than from any description.

2. Bank-to-ledger reconciliation

Next, the bank needs to match the books: is every payment recorded, are there duplicates, do amounts and counterparties agree. Export the period's transactions from the ledger to Excel.

Compare two files: the bank export and the ledger transactions for September. Match lines by date (within 2 days), amount, and counterparty. Show three lists: in the bank but not the ledger; in the ledger but not the bank; matched but with a different amount or counterparty. For each item give the row numbers in both files.

Result: a short exception list instead of two long tables. Typically it holds a payment posted on the wrong day, a bank fee nobody recorded, and a couple of vendors named differently in the ledger.

Before/after (estimate): 3 hours versus 1 hour, including working the exceptions.

Ask for row numbers every time. Without them the model can "find" a difference that does not exist, and you may not notice.

3. Budget vs actual and variance explanations

Computing variances on 40 lines is easy. Explaining them so the owner sees the cause, not one more column of percentages, is the hard part.

Here is the September budget vs actual, the coded bank statement, and department heads' comments. Find lines where the variance is above 10% and above $20,000. For each, suggest a cause using only data from these files and cite the rows you used. If the data does not show a cause, say so.

Result: six to eight lines with explanations such as "freight 18% over budget: two unplanned shipments to Dallas, invoices dated Sep 12 and Sep 19". Where the data runs out, the model says so. That tells you who to ask.

Before/after (estimate): about 3 hours versus 1.5.

You set the thresholds. Do not let the model decide what counts as material.

4. Cash forecast and the gap week

For October you need to know whether cash covers commitments. Inputs: opening balances, the vendor payment schedule, expected customer receipts, payroll, tax, and rent.

Build a weekly cash forecast for October: opening balance, receipts, payments, closing balance. Sources: balances on Oct 1, the AP schedule, AR aging with expected pay dates, recurring payments. Flag weeks where closing cash drops below $250,000. Suggest which payments could move, but do not move anything.

Result: a forecast with one tight week mid-month and a short list of payments worth discussing with vendors. The decision stays with you and the owner.

Before/after (estimate): 2 hours versus 40 minutes.

5. Checking contracts and invoices for payment terms

September brought invoices from new vendors and amended contracts from old ones. Do the invoice terms match the contracts: due date, deposit, currency, late fees, remittance details?

Here are 20 vendor contracts (PDF) and the September invoices. Match each invoice to its contract and compare amount, currency, due date, deposit, and bank details. Return a table and a separate list of every mismatch with the exact contract clause quoted.

Result: a match table and three or four mismatches. An invoice demands full prepayment where the contract says 30%, or the bank details on the invoice differ from the contract. That last one matters most: changed remittance details are a classic sign of vendor fraud, so confirm by phone before paying.

Before/after (estimate): about 2 hours versus 40 minutes.

If documents arrive every day, this is a flow rather than a monthly task. We describe how an agent handles that flow in how an AI agent checks documents.

6. Excel formulas and macros

The most underrated task. Finance people lose hours hunting for a formula or repeating the same manual steps each month: paste an export, rename vendors, roll up locations.

Sheet "Bank" has descriptions in column C and amounts in D. Sheet "Map" has keywords in column A and accounts in B. Write a formula that returns the account for the first matching keyword, or "unclear" if none match. Excel 365.

Result: a working formula with an explanation. For repeated steps, ask for a macro and where to paste it. State your Excel version, since available functions differ.

Before/after (estimate): anywhere from half an hour to half a day of searching, versus 10–15 minutes.

7. The owner memo

The owner does not want 40 lines of budget vs actual. They want three answers: what happened, why, and what to do.

Here are the September results: budget vs actual, variance explanations, and the October cash forecast. Write a half-page memo for the owner: the three main events of the month, their causes, the cash risk in October, and two decisions we need from them. Plain language, no accounting jargon. Use only numbers from these files.

Result: a draft you edit for 10–15 minutes. The edit is not optional. The model does not know the owner already heard about the Dallas shipments, or that marketing is a sensitive topic this quarter.

Before/after (estimate): 1.5 hours versus 30 minutes.

8. Scenario forecast

The last step of the close looks forward. What if the largest customer pays a month late? What if import costs rise 10%?

Using January–September actuals and the October cash forecast, project cash to year end in three scenarios: base, largest customer pays 30 days late, 10% increase in imported purchase costs. Put the assumptions in a separate table so I can change them.

Result: three scenarios and an assumptions table. The value is speed rather than precision: change an assumption and see the ending cash move.

Before/after (estimate): about 3 hours versus 1.

Across the eight tasks, our example goes from roughly 20 hours of manual work to 6–7 hours. The saved time goes into checking and into conversations with department heads, which is the work a finance professional is hired for.

When chat is not enough: Claude Code and Codex

Chat has a ceiling. Every month you upload the files again, explain the rules again, and check again that the model did not change its logic. If the close repeats, move it into an agent.

Claude Code from Anthropic and Codex from OpenAI are programs that work on a folder on your computer. You can run them in a terminal, a code editor, or a desktop app. You describe the task in words, as in chat. The agent reads the files, writes a small program for the calculation, runs it, checks the output, and saves an Excel file. You do not need to program, but it helps to read what the agent plans to do and approve each step.

A close folder and a rules file

A simple setup: a "Close" folder with subfolders for bank exports, ledger exports, the budget, contracts, and outputs. At the root sits a rules file. Claude Code reads it as CLAUDE.md, Codex as AGENTS.md. Five lines are enough to start:

# Month-end close
- Bank exports in bank/, ledger exports in ledger/, budget in budget.xlsx, chart of accounts in coa.xlsx
- Do every calculation in a script in scripts/, never in your head
- For every number in a report, cite the file and row it came from
- Never edit or delete source files
- If an account or match is unclear, write "unclear"; do not guess

From then on, the monthly instruction is one sentence: "close September by the rules: code the bank, reconcile to the ledger, run budget vs actual, draft the memo." The agent runs the same eight steps you did in chat and keeps the rules between months.

Anthropic publishes a starting point: an open finance plugin with commands for reconciliation, variance analysis, journal entries, and a month-end close checklist. It is written for US GAAP and SOX work, so adapt the rules to your own policies.

Why an agent is easier to trust

In chat, the model gives an answer and you cannot always see how it got there. An agent leaves the calculation behind as a script. That changes three things.

You can rerun it. If a missing bank line turns up, add it and run the script again. Nothing is recalculated from memory.

You can check it. A colleague, an auditor, or the agent in a fresh session can read the script and see which rule matched the payments and why row 312 landed in "unclear".

Next month runs the same logic. The rule you fixed in September already applies in October. In chat you would have to remind the model.

This is the principle we treat as the core of AI in finance: the model shows where every number came from. If it cannot, do not use the number.

Two practical tips. Start the agent with read-only access to source files and write access only to the outputs folder; do not connect it to the bank portal or give it write access to the ledger. For each new task, run the first two or three months in parallel with your manual process and compare totals.

If you want to learn this on your own tasks, join the course waitlist. While the course is paused, 1-on-1 sessions are available on request.

Security: what not to give the model

In finance this is the first question. The short answer: anonymize what you can, and do not put work data into a personal free account.

What to keep out of an ordinary chat:

• personal data: national IDs, salaries by name, card numbers
• full bank details next to amounts and counterparty names, when you can do without them
• data covered by confidentiality clauses with clients or lenders
• logins, passwords, and API keys for the bank portal or ERP

What helps:

• Anonymization. Replace counterparties with codes ("Vendor 07"), drop IDs and account numbers. For bank coding and variance work that is almost always enough.
• A work plan. On ChatGPT and Claude business plans, your data is not used to train models by default. On personal accounts, turn training off in settings.
• Your company's policy. Many firms already have an approved tool list, and regulated data such as personal data under GDPR has rules about where it can go. Check before the first upload, not after.

If the company wants AI for everyone, not just the finance team, it needs a shared setup with access rights and logs. We cover that in an internal ChatGPT for your company.

Will AI replace finance professionals

No, but the job changes. Manual coding, formula hunting, and eyeballing reconciliations leave the close. Three things stay and matter more.

Judgment. Is the variance material, should the payment move, is the model's suggested cause right. The model proposes, you decide.

Accountability. You sign the numbers, not the model. A mistake in the owner memo is your mistake, even if the model wrote the draft.

Conversations. Finding out from operations why there were two extra shipments, agreeing a new payment date with a vendor, explaining a risk to the owner. AI can prepare you for these, but you still have to have them.

Data entry and first-pass coding change fastest. A finance professional who can give AI a clear task and check the result becomes more valuable, because they fit more analysis into the same week.

A two-week starting plan

Learn on a real close with real files.

Days 1–2. Choose the tool and plan. Turn off training on your data or use a work account. Put together your chart of accounts with a few examples per account.

Days 3–5. Take last month's close, where you already know the answer. Redo tasks 1–3 in chat and compare with your manual result. That shows you where the model goes wrong on your data.

Days 6–8. Add tasks 6 and 7: formulas and the owner memo. Fastest win, lowest risk.

Days 9–10. Save the prompts that worked into one document. It becomes your future agent instructions.

Days 11–14. If the tasks repeat, try Claude Code or Codex on a copy of the close folder. Start with one task, such as the reconciliation, and run it alongside your manual process.

By the end of week two you will know which tasks save time and which are still easier by hand.

Training the finance team

When one person learns AI, the gain stays with that person. When the finance team works by shared rules, the gain becomes a process: the same prompts, one chart-of-accounts map, an agreement on what data can be uploaded and who checks the output.

A team needs to settle four things: which close tasks go to AI and which do not, how to anonymize data, how to check model output, and where prompts and rules live so a new hire works the same way. We run this training on the company's own files and tasks. Details are on the corporate AI training page.

When you need a system

A personal tool stops being enough when the work outgrows one person and one folder. Signs:

• statements and documents arrive daily from several channels: bank, email, shared drives, chat
• you need direct read access to the ERP and the bank rather than exports
• several people use the output, and it matters who saw and approved what
• an auditor or the owner asks where a number came from, and the answer is needed in a minute, not a day

At that point you need a system: agents on top of the ERP and the bank with read access, roles, and an audit trail. Every number in a report links to its source, every human decision is logged, and permissions mirror the finance team's. The agent gathers and matches, a person approves, and posting and payments stay in the ERP and the bank.

These projects usually stall on gray areas in the process rather than on the model. Urgent payments get approved in a chat thread, the latest contract version sits on someone's laptop, the vendor master was never cleaned. A system will not fix those, but it shows them quickly. What AI looks like at the department level, what to automate first, and what to leave alone at the start is covered in AI for finance departments. Such a system is tested on a set of real past cases; see why AI projects need evals.

If you need that kind of system on top of your ERP, bank, and email, see our GPT integration service.

FAQ

What is the best AI tool for accountants?

For spreadsheets, bank exports, and Excel formulas, ChatGPT and Claude both work well: they calculate with code and handle files. Microsoft 365 Copilot fits teams that live in Excel and Outlook. For a close that repeats every month, the agents Claude Code and Codex are a better fit.

Can I upload a bank statement to ChatGPT?

Yes, if you strip what is not needed and use a work account. Replace counterparty names with codes, remove IDs and account numbers, and either use a business plan or turn off training on your data. Personal data about employees or customers needs a legal basis and your company's approval before it goes to any outside service.

Can AI work directly with my accounting system?

A chat tool does not log into your ledger, but it works well with Excel exports. Connecting AI to the ledger with read access is possible, but that is an integration project with permissions and logging.

Do I need to code to use Claude Code or Codex?

No. You describe the task in words and the agent writes the calculation. It helps to read what it plans to do and to keep it away from editing source files. The first setup of the folder and rules file is easier with a colleague or in a training session.

Will AI replace finance professionals?

No. AI takes over manual coding, line-by-line matching, and first drafts. Judgment on materiality, accountability for the numbers, and conversations with the business stay with people. A finance professional who works well with AI simply fits more analysis into the same time.

How far can I trust numbers from a model?

As far as you can check them. Ask for row numbers and file references, tie totals back to the source, and run the first months in parallel by hand. A calculation the agent saved as a script is much easier to check than an answer in a chat window.

Bottom line

AI for finance professionals does not close the month on its own. It takes on coding, matching, calculations, and drafts. Start in chat on one real close, check the output against a month you already finished, then move the repeating steps into Claude Code or Codex with a rules file. To learn it yourself, join the course waitlist. To bring the whole finance team along, look at corporate AI training. And if you need a system on top of the ERP and the bank with roles and an audit trail, that is our GPT integration work.