Playbook: Finance and Accounting

Finance work is mostly two things: getting numbers right, and explaining what they mean to people who don't read spreadsheets. AI is excellent at the second, and — used the right way — a fast assistant for the first.

Used the wrong way, it's dangerous: a language model will add up a column in prose, get it wrong, and say so with total confidence.

---

1. The two non-negotiable rules

Rule 1 — never let a model do arithmetic in text. Language models predict words; they don't calculate. For any sum, ratio, variance or reconciliation, make the AI run code (ChatGPT's data analysis, Claude's analysis tool, Gemini in Sheets, or Copilot in Excel) or produce a formula you run yourself. Then spot-check totals against the source.

Rule 2 — protect the data. Bank statements, salary data, client ledgers and PAN/GST details don't go into consumer AI accounts. Use:

• Company-approved business/enterprise tools with a data agreement and training disabled, or
• Anonymised extracts (client codes instead of names, rounded or indexed figures), or
• A local model for sensitive analysis where policy requires it.

Check your firm's policy and client engagement terms before using AI on client books.

2. The financial narrative

Leaders don't want the P&L; they want to know what changed, why, and what to do. That translation is where AI saves the most time.

You are a finance business partner writing for the CEO, who is not
an accountant.

Attached: monthly P&L for the last 6 months (CSV).
First, use code to calculate month-on-month and year-to-date
changes for each line. Show me the calculations table.

Then write a one-page memo:
1. The 3 most important trends, with the numbers
2. For each, the likely drivers — label clearly which are
   supported by the data and which are hypotheses to confirm
3. One risk and one opportunity
4. Decisions or questions for the CEO

Plain language. No figure that isn't in the calculations table.

Your job: confirm the drivers. AI can see that marketing spend jumped in March; only you know it was the trade show.

3. Reconciliation with AI

Matching two messy lists — bank vs ledger, invoices vs payments, vendor statement vs payables — is tedious and error-prone.

I have two files: bank_statement.csv and ledger_export.csv.
Using code:
1. Standardise dates, amounts (to 2 decimals) and remove extra
   spaces in references
2. Match on transaction ID where present; otherwise on exact amount
   and date within ±3 days
3. Output three tables: matched, in bank only, in ledger only
4. Flag near-matches (same amount, different date beyond 3 days,
   or amounts differing by under ₹10)
5. Give counts and totals for each table

Show the code you ran.

Review every unmatched and near-matched item yourself. The AI finds candidates; an accountant decides what they are — timing difference, bank charge, error or something that needs escalation.

4. Chat with policies and circulars

Tax rules, GST notifications and internal finance policies are long and change often. A project with the documents loaded answers questions faster than searching PDFs.

1. Create a project (Claude Project, Custom GPT, NotebookLM) and upload the current official documents — the relevant Finance Act, CBIC/CBDT circulars, your internal policy.
2. Instruction: "Answer only from the uploaded documents. Quote the section and page for every answer. If the documents don't cover it, say so."
3. Ask real questions: "Under the uploaded circular, what documentation does a startup need to claim [deduction]?"

Always open the cited section and read it. Treat the answer as research, not advice. Tax positions for clients still need professional judgement and sign-off, and rules change with every Budget — replace the documents when they do.

5. Useful everyday prompts

• Variance commentary: "Using the calculations table, draft two-sentence commentary for each line with variance above 10%."
• Excel help: "Write an Excel formula that returns the latest invoice date per customer from this layout: [describe columns]. Explain it."
• Clean an export: "Using code, split this 'Party Name / GSTIN' column into two columns and flag invalid GSTIN formats."
• Board summary: "Turn this management report into 5 bullet points for a board pre-read, keeping every number exactly as written."

6. Controls that make AI safe in finance

• Maker-checker stays: AI output is the maker; a person is always the checker.
• Keep the working: save the code and calculation tables with the output.
• Tie out: totals in any AI output must match the source system.
• Log AI use on client deliverables where your firm or clients require it.

---

⚠️ Common mistakes

• Arithmetic in chat text. Always make it run code.
• Uploading raw bank statements or client ledgers to personal AI accounts.
• Trusting a tax answer without reading the cited section.
• Outdated documents in your policy project after a Budget or notification.
• Presenting AI-guessed drivers as facts in a CEO memo.

---

What's next: from numbers to people — AI in HR and recruiting, where the fairness stakes are highest.