CORAA
Blog/AI in Audit

How to Use ChatGPT and Claude for NPA Classification of Agricultural Loans: A Step-by-Step Guide for CAs

A step-by-step method for using ChatGPT, Claude or Gemini to test agricultural and allied loans for NPA classification and provisioning: safe setup, copy-paste prompts, a worked answer, and the checks that catch AI errors.

CCORAA Team26 September 20269 min read

How to Use ChatGPT and Claude for NPA Classification of Agricultural Loans: A Step-by-Step Guide for CAs

Yes, a general AI assistant can help you test agricultural and allied loans for NPA status, provided you paste the governing paragraphs yourself, give it anonymised data, and treat every date and rupee figure as a draft you recompute. This guide shows the exact prompts, what a correct answer looks like, and the checks that catch the usual errors.

It pairs with our reference post on NPA classification for agricultural and allied loans under the RBI Master Circular 2025. Read that first for the rules; this one teaches the workflow.

Last reviewed: 26 September 2026. Educational guide, not legal or professional advice.

What can ChatGPT or Claude do here, and what should you never delegate?

Good use of AI Never delegate
Applying a rule you pasted to a sample account and showing the steps Recalling RBI thresholds or paragraph numbers from memory
Recomputing overdue days and provision percentages Deciding whether a product falls within Annex-2 without the sanction terms
Drafting a test programme or a query list for the branch Signing off a classification or a provisioning conclusion
Explaining a paragraph in plain language Handling real borrower names, account numbers or PAN
Formatting findings into a working-paper table Getting the SLBC crop calendar: that comes from the state SLBC, not from the AI

The AI is a fast junior who reads carefully when you hand over the text and improvises when you do not. Your job is to hand over the text.

Step 1: Set up a safe workspace

Before you type anything:

  • Never paste borrower names, PAN, GSTIN, CIN, account numbers or branch identifiers. Replace them with "Borrower A", "Loan 1" and so on.
  • Use synthetic or anonymised figures. Round amounts if the exact value adds nothing.
  • Use a paid or enterprise plan with training on your data switched off, where your firm's policy allows AI at all.
  • Save your role prompt once. In ChatGPT you can store it in Custom Instructions or a Project; in Claude you can store it in a Project's instructions; in Gemini a saved Gem does the same job. Generic fallback: keep it in a text file and paste it at the start of every chat.

For ready-made wording on redaction, use our DPDP-safe prompt template library for CA firms.

Step 2: Give the AI a role and rules of engagement

This is the system-style prompt. It stops the AI from filling gaps with guesses.

You are an audit assistant helping a chartered accountant test bank
advances for NPA classification under RBI prudential norms.

Rules:
1. Use ONLY the rule text I paste in this chat. If a rule you need is
   not in the pasted text, say "not provided" and stop. Do not recall
   thresholds or paragraph numbers from memory.
2. Quote the exact sentence you rely on for every conclusion.
3. Show every calculation step by step, including day counts.
4. List every assumption you make. If a fact is missing (crop
   duration, security value, due date), ask me instead of assuming.
5. Finish with a table: conclusion, rule relied on, open questions.

Step 3: Paste the governing rule yourself

Open the RBI Master Circular dated 1 April 2025 (or the current Directions) from rbi.org.in and copy the actual text. Provide the source, ask it to apply it. A good paste covers the crop-season paragraph, the Annex-2 scope paragraph and the provisioning paragraphs. Or attach the PDF where your tool supports file upload, and still quote the paragraphs you need in the chat so the AI works from visible text.

RULE TEXT (pasted from the RBI Master Circular, 1 April 2025):

[Paste para 2.1.2 (iv)-(v), para 4.2.13.1, para 4.2.13.2, Annex-2
heading and Parts A-C, paras 5.4.1, 5.3.1, 5.3.2, 4.1.1 and 4.1.2 here.]

Confirm you have read this text and list, in one line each, the tests
it creates. Do not analyse any loan yet.

If the summary of tests it returns does not match the text, fix that before moving on.

Step 4: Run the worked case

Use two anonymised loans, one inside Annex-2 and one outside it. Every fact the AI needs is stated; nothing is left for it to assume.

Apply the rule text above to these two loans. Today's date for
staging purposes is 30 September 2026.

Loan 1: Short-duration crop loan, individual farmer (Annex-2 Part A
activity). Outstanding Rs 10,00,000. Realisable security Rs 6,00,000.
The instalment fell due before the kharif harvest and stayed unpaid
through the rabi season too, so it is overdue for two crop seasons on
the SLBC calendar.
Task: classify it. Then compute the provision (a) as substandard and
(b) after 12 months, when it turns doubtful.

Loan 2: Dairy loan, Rs 5,00,000 outstanding. One instalment fell due
on 15 February 2026 and is unpaid. No Annex-2 mapping is documented.
Task: state which test applies and why, compute the date on which it
becomes an NPA (show the day count), and compute the provision as
substandard.

Show working step by step. State all assumptions. Quote the rule
sentence you rely on for each conclusion.

Step 4b: Challenge it

Follow up with: "Which of your conclusions depends on a fact I did not give you? What would change if the dairy loan were bundled in a Kisan Credit Card limit?" A good answer flags the product-mapping question rather than picking a side.

What a good answer looks like

These are the results you should compute yourself from the verified rules in the companion post, so you know what to expect.

Loan 1. Two crop seasons overdue on a short-duration crop loan: NPA, substandard.

  • Substandard provision: 15 per cent of Rs 10,00,000 = Rs 1,50,000 (no allowance for security).
  • Once doubtful (after 12 months): unsecured portion Rs 10,00,000 less Rs 6,00,000 = Rs 4,00,000 at 100 per cent = Rs 4,00,000. Secured portion Rs 6,00,000 at 25 per cent (first year of doubtful) = Rs 1,50,000. Total Rs 5,50,000.

Loan 2. Not in Annex-2, so the ordinary 90-day norm applies. From 15 February 2026: 13 days to 28 February, 31 in March (44), 30 in April (74), 16 in May (90). Day 90 is 16 May 2026, so the account is overdue for more than 90 days on 17 May 2026, which is the NPA date. Provision as substandard: 15 per cent of Rs 5,00,000 = Rs 75,000.

Item Expected result
Loan 1 classification NPA, substandard
Loan 1 provision, substandard Rs 1,50,000
Loan 1 provision, doubtful year 1 Rs 5,50,000
Loan 2 test 90-day norm
Loan 2 NPA date 17 May 2026
Loan 2 provision Rs 75,000

Illustrative, wording varies run to run: a good AI response will reach these numbers with its working shown, quote the paragraph text you pasted, and note that the dairy conclusion depends on the absence of a documented Annex-2 mapping. If your run gives different numbers, do not average the runs. Find the step that differs.

Where AI goes wrong

Failure What it looks like Check
Outdated or remembered thresholds Quotes 10 per cent, or another old provision rate, that is not in your paste Search its answer for any percentage not in the pasted text
Invented paragraph numbers Cites "para 4.2.13.5" or a clause that does not exist Open the circular and confirm each cited paragraph
Arithmetic slips Off-by-one on day counts; secured/unsecured split wrong Recompute the day count on a calendar and the provision on a calculator
Silent assumptions Assumes crop is short duration or that the loan is in Annex-2 Look at the assumptions list; if it is empty, the prompt was ignored, so re-run
Over-extending the relief Applies the two-season test to the dairy loan Ask "Is this activity listed in Annex-2?" and make it quote the entry

Five checks to run on every output:

  1. Every percentage and paragraph number appears in your pasted text.
  2. The day count matches your own calendar count.
  3. Secured plus unsecured portions add back to the outstanding.
  4. The assumptions list is not empty and contains nothing you did not provide.
  5. The crop-season conclusion is backed by the SLBC calendar you hold, not by the AI's claim about "typical" seasons.

Step 5: Verify in 60 seconds

Take the same loan and run it through the free NPA classification calculator. It applies the 90-day test, the out-of-order test for CC/OD and the crop-season test for agricultural advances and shows staging and provision. If the calculator and the AI disagree, the AI is wrong until you prove otherwise. Then check the rule against the circular text. The calculator is a sanity check on a sample item; it does not replace the SLBC calendar or the bank's own records.

Do ChatGPT, Claude and Gemini differ for this task?

The method is the same. Only the mechanics differ, and only where it is safe to say:

  • All three accept a pasted paragraph, and all three let you attach a PDF in current paid plans. Pasting the specific paragraphs is still safer than trusting the model to find them in a long PDF.
  • ChatGPT (Custom Instructions or Projects), Claude (Projects) and Gemini (Gems) can each store your Step 2 role prompt so you do not retype it. Menu names change, so check the current interface.
  • Any of them can return a wrong number with confidence. None replaces the calculator check.

Frequently asked questions

Can ChatGPT or Claude decide whether an agricultural loan is an NPA?

It can apply a rule you paste to facts you supply and show its working. The classification is your professional judgement, and it rests on the bank's records and the SLBC crop calendar, which the AI does not have.

Is it safe to paste bank loan data into ChatGPT or Claude?

Not with real borrower details. Anonymise names, account numbers and identifiers, and check your firm's policy and the tool's data settings first. See the DPDP-safe prompt library linked above.

Why should I paste the RBI text instead of asking the AI what the rule is?

Models can recall thresholds from older versions or invent paragraph numbers. Pasting the current text turns the task into applying a rule you supplied, which is far easier to verify.

How do I know if the AI has invented a paragraph number?

Open the Master Circular and look it up. If you cannot find the cited paragraph or the words differ from what the AI quoted, discard that conclusion.

Does the crop-season relief apply to dairy or poultry loans?

On the circular text, the two-season and one-season test applies only to Annex-2 farm credit; other loans follow the 90-day norm. Ask the bank for any documented product mapping before concluding. The reference post covers this in detail.

Which is better for this work: ChatGPT, Claude or Gemini?

For a rule-application task with pasted text, the discipline of the prompt matters more than the model. Test your own prompt on a case where you already know the answer, and use the model that follows the "not provided" instruction most reliably for you.

Where CORAA fits

A chat window gives you a draft. CORAA is built for CA firms that need the recomputation, evidence and working papers kept together on a real engagement; see AI agents for audit. When you are ready to try it on your own file, start a free trial.

Last reviewed: 26 September 2026. Sources: RBI Master Circular RBI/2025-26/13 dated 1 April 2025, as summarised in our companion post.

Topics
how to use chatgpt for npa classificationclaude for bank audit npa testingAI agricultural loan NPA crop seasonChatGPT prompts for bank statutory auditAI provisioning calculation agricultural advances
Share
← Back to all articles
Keep reading

More in ai in audit.

Built for India · DPDPA compliant

Ready to automate your audit work.

See how Coraa reduces audit engagement time by 60%, from ledger scrutiny to working papers, all from one Tally import.

Run one complete audit free