It's the 4th of the month. Six clients are waiting on books, one has sent a bank statement as a phone photo, another has an expense line called "misc 2" for 48,000 rupees, and you still have the month-end checklist in your head, not on paper. AI prompts for accountants and bookkeepers in India will not close the books for you. What they can do is shorten the boring parts: a checklist you actually follow, queries you'd otherwise type out ten times, and a first read of why a reconciliation won't tie.
This pack is built around one split. You and your software do the arithmetic and apply the rules. The model drafts, asks questions and flags oddities. These are untested templates; I haven't run them on a client's books. The worked example uses invented figures, and I checked its arithmetic with a short script.
One scope note. This post gives no tax, GST or compliance rules, on purpose. Those change by notification, and a confident wrong answer costs your client money. Where a prompt touches a rule, the prompt tells the model to ask you for the source. For the rules side, take them from the official government pages or check with a CA. The separate post on prompts for chartered accountants covers GST and income tax workflows, and you should read it with the same caution.
How do I make a month-end close checklist with AI?
Start from your own process, not the model's idea of one. Generic checklists miss what makes a client different: a cash-heavy shop, a services firm with milestone billing, a trading business with stock.
You are helping a bookkeeper in India prepare a month-end close checklist
for one client.
CLIENT PROFILE:
- Business type: [TRADING / SERVICES / MANUFACTURING / OTHER]
- Software: [TALLY / ZOHO BOOKS / EXCEL / OTHER]
- Bank accounts: [N], credit cards: [N], cash handled: [YES/NO]
- Stock tracked: [YES/NO]; fixed assets: [YES/NO]; payroll: [YES/NO]
- Month being closed: [MONTH YEAR]
- Things that went wrong in earlier closes: [LIST]
OUTPUT:
1. A checklist grouped as: bank and cash, receivables, payables, stock,
payroll, fixed assets, accruals and prepayments, final review.
2. For each item: what to check, what document or report to use, and who
supplies it (me or the client).
3. A "chase the client" list, separated, with what is needed by when.
RULES:
- Do not state any tax, GST or statutory rule or due date. Where an item
depends on one, write: "Rule to confirm from official source or CA".
- Do not invent figures. Use [AMOUNT] placeholders.
- Keep it to one page.
Run it once per client type and save the result as a template. Then edit by hand. The "things that went wrong before" line is where the value is. A checklist that includes "client forgets to send card statements" gets you a real chase item.
Which prompts help with bank reconciliation?
Matching is a job for your software or a spreadsheet. The model's use is the leftovers: it can suggest the usual reasons a small set of items stays unmatched, and it can write the test for each.
Bank reconciliation for [CLIENT CODE], [MONTH YEAR].
Book balance (from ledger): [AMOUNT]. Bank statement balance: [AMOUNT].
Difference I cannot explain: [AMOUNT].
Unmatched items (parties masked as V1, C1 etc.; 10 to 25 rows):
Side | Date | Amount | Narration | Cheque or ref no.
[PASTE ROWS]
Tasks:
1. Group the unmatched items by likely cause: deposit in transit,
cheque not yet presented, bank charge not booked, duplicate entry,
wrong date, wrong sign, amount keyed with a transposed digit.
2. For each group, say what evidence in the statement or ledger would
confirm it.
3. List queries to send the client, one line each.
4. Do NOT total anything or state that the reconciliation ties. I will
compute the adjusted balances myself.
Rule 4 matters. A model that says "this ties" without a computed check is the worst kind of helper. Here is the shape of a correct check, with invented numbers I ran in a script:
| Line | Amount |
|---|---|
| Bank statement balance | 402,546.00 |
| Add: deposits in transit | 18,200.00 |
| Less: cheques issued, not presented (6,500 and 2,250) | 8,750.00 |
| Adjusted bank balance | 411,996.00 |
| Book balance | 412,350.00 |
| Less: bank charges not yet booked | 354.00 |
| Adjusted book balance | 411,996.00 |
Both sides come to 411,996, so the difference is zero. The script printed adjusted bank 411996.0, adjusted book 411996.0 and difference 0.0. The model's contribution was to suggest "bank charge not booked" as the reason for a 354 gap and to propose checking the statement narration. It did not do the sums, and shouldn't.
What about ledger scrutiny and odd entries?
Accountants spend hours looking for the entry that doesn't belong. A model can help you design the questions, particularly for a ledger export you have already cleaned.
Below is a masked ledger extract for [ACCOUNT NAME], [PERIOD], from a
[BUSINESS TYPE]. Columns: Date | Voucher type | Narration | Debit | Credit.
[PASTE UP TO 50 ROWS]
List entries that deserve a question, grouped as:
- Round-number amounts
- Vague narrations ("misc", "adjustment", "as discussed")
- Same amount appearing more than once within [N] days
- Entries dated at period end that look like cut-off issues
- Entries that look posted to the wrong account for the narration
For each, give the row, why it stands out, and the question I should ask
the client. These are leads, not findings. Do not say any entry is wrong.
For "same amount twice", do it yourself in a spreadsheet too (a COUNTIFS or pivot is faster and complete). The model reads a pasted extract, so it can miss rows and it can't see the rest of the ledger. If you want spreadsheet formulas for this, ChatGPT Excel and Google Sheets formula prompts covers the pattern.
How do I draft accrual and prepayment queries?
Accruals depend on facts only the client knows: work done but not billed, a contract that spans months, an annual subscription paid upfront. Use the model to turn your gaps into questions, not to decide amounts.
For [CLIENT CODE], month [MONTH YEAR], I need to confirm accruals and
prepayments.
Known recurring items: [LIST: rent, internet, insurance, software, audit fee]
Items paid this month that may cover more than one month: [LIST]
Work received but not yet invoiced by suppliers: [LIST OR UNKNOWN]
Write:
1. A question list for the client, plain English, no accounting jargon,
each with the document I need (contract, invoice, email).
2. A table with columns: Item | Period covered | Amount paid | Months in
this period | Months in future | Amount to defer (formula in words only).
Leave the amounts blank for me to compute.
3. A reminder line at the bottom: "Treatment depends on the client's
accounting policy and applicable standards: confirm with the
responsible CA."
Leaving the amount blank is deliberate. The table gives you the structure, you do the division, and the "confirm with the CA" line stops a template from becoming a policy.
Can AI write client reminders that actually get a reply?
Yes, and this is probably the safest use. Chasing documents is the part of month-end that eats time and goodwill.
Write a WhatsApp-length message (under 90 words) to [CLIENT CONTACT],
who runs [BUSINESS TYPE]. I need these by [DATE] to close [MONTH]:
[LIST OF ITEMS, max 6]
Tone: polite, direct, no pressure language. Say what happens to the
timeline if the items come late: [CONSEQUENCE, e.g. "I'll close without
the card statement and flag it as unreconciled"].
End with one clear question they can answer with "yes" or a date.
Write it in [ENGLISH / HINGLISH / HINDI].
Keep the consequence true. If you won't really close without the statement, don't threaten it. For another example of structuring client messages when documents arrive as images and PDFs, see AI invoice data extraction into Google Sheets.
A worked example (illustrative)
An invented client: a small trading business. Your masked inputs say the bank statement and book differ by 9,804 before adjustments (the same figures as the table above), there's one deposit of 18,200 that only shows on the next month's statement, two cheques issued but not presented, and one debit on the statement with the narration "chgs".
What good output looks like (shape):
Group A: Timing. Deposit of 18,200 appears on the next statement; likely a
deposit in transit. Confirm the deposit slip date.
Group B: Cheques. Two issued cheques not yet in the statement; confirm
issue dates and cheque numbers in the cheque register.
Group C: Bank charge. Debit narrated "chgs" has no matching ledger entry;
confirm against the bank's charge advice and book it.
Queries for client: (1) Deposit slip for the 18,200 item. (2) Confirm the
two cheques were handed over.
I have not totalled or confirmed that the reconciliation ties.
Then you compute the adjusted balances as in the table above. The model gave you structure and three leads. The tie-out is yours.
Where this goes wrong
| Failure | What it looks like | Fix |
|---|---|---|
| Invented rule | A GST or tax statement with a confident tone | Forbid rules in the prompt; use official pages or a CA |
| Mental arithmetic | A total that is off by a few rupees | Compute outside, paste results, "do not calculate" |
| "It ties" | A tie-out claimed from pasted rows only | Forbid verdicts; compute adjusted balances yourself |
| Unmasked client data | GSTIN, PAN or bank numbers in a chat | Mask with codes; keep the key; check policy |
| Missed rows | Anomaly list covers only part of an extract | Do completeness checks in the spreadsheet |
| Generic checklist | Items irrelevant to this client | Fill the profile and "past problems" lines |
| Wrong language tone | Stiff Hindi or Hinglish | Read aloud before sending; edit |
What not to automate
- Tax, GST and statutory treatment. The model can't be your source. Take rules from the official pages and your CA, and write the source on the working paper.
- The reconciliation sign-off. You compute the adjusted balances. You say it ties.
- Posting entries. Draft suggestions are fine; posting to the ledger is your professional act.
- Judgements on estimates and policy. Provisions, depreciation choices and write-offs need the responsible professional.
- Client data in unapproved tools. Mask it, minimise it, or leave it out. The post on what not to paste into ChatGPT is a good checklist.
- Anything sent unread. Reminders and queries go out under your name.
For the logic behind "use only these inputs", the lesson on avoiding hallucinations explains why. If you work on analysis and commentary rather than books, prompts for finance analysts pairs well with this pack, and the prompt library has more templates you can copy and adapt.
