How to Use ChatGPT and Claude to Draft a Utilization Certificate: A Step-by-Step Guide for CAs
Yes, a general AI assistant such as ChatGPT, Claude or Gemini can speed up a utilization certificate (UC), but only if you supply the sanction terms and the figures and treat its output as a draft to be checked. The method below keeps the rules and numbers in your hands and uses the AI for structure, arithmetic re-performance and wording.
This is the hands-on companion to our reference post, Utilization Certificate: Format, Who Signs and a CA Checklist. Read that first for the rules. Here you learn the workflow.
What can ChatGPT or Claude do here, and what should you never delegate?
| Good uses (AI drafts, you check) | Never delegate |
|---|---|
| Laying out the statement in the grantor's format once you paste it | Deciding who is entitled to sign |
| Re-adding the receipts and spending table | Confirming that spending is within the sanctioned heads |
| Turning your notes into scope and basis wording | Vouching, bank tracing and reading the sanction letter |
| Listing checks from the sanction text you provide | Quoting rules or time limits from memory |
| Spotting inconsistencies between two tables you paste | Signing, or issuing the UDIN |
The rule is simple: provide the source, ask it to apply it. Do not ask "what does GFR 12-A require?" and trust the answer. Paste the form text or sanction clause and ask the model to work only from that.
Step 1: Set up a safe workspace
Client data is confidential. Before typing anything:
- Never paste client names, PAN, GSTIN, CIN, bank account numbers or sanction numbers that identify the client.
- Replace them with labels such as "Grantee A", "Scheme X" and "Sanction 1".
- Round or shift figures if the numbers themselves could identify the client, or use a synthetic case that mirrors the structure.
- Use a paid or business workspace where your organisation has reviewed the data-use settings, and switch off training on your chats where the tool allows.
- Keep the real figures in your Excel working paper, not in the chat.
For ready-made anonymisation wording, see our DPDP-safe prompt template library for CA firms.
Step 2: Give the AI a role and a boundary
Start a fresh chat and paste this.
You are an assistant to a chartered accountant preparing a utilization
certificate. Rules for this chat:
1. Use ONLY the sanction terms and form text I paste. If something is not
in them, say "not stated" - do not fill the gap from memory.
2. Do not name any law, rule, form number or time limit unless it appears
in the text I paste.
3. Show every calculation line by line.
4. List every assumption you make, and ask me before making it.
5. Flag anything that looks inconsistent. You do not sign or certify;
you only draft and check.
Reply "ready" if you understand.
Step 3: Paste the governing text yourself
Paste the format the grantor prescribed (for government grants, the GFR 12-A form text) and the relevant sanction clauses, with identifiers removed.
Here is the certificate format required by the grantor:
[PASTE FORM TEXT]
Here are the sanction conditions (identifiers removed):
[PASTE CLAUSES: purpose, heads and proportions, period, interest
treatment, treatment of unspent balance, who must sign]
Task 1: List, as a numbered checklist, every check and every figure the
format and conditions require me to support. Quote the clause for each.
Do not add any requirement that is not in the text above.
Check the list against the sanction letter. If it contains a requirement you cannot find in the text, it was invented.
Step 4: Feed the worked case
Use anonymised or synthetic figures. Here is a practice case you can paste as is.
Case (synthetic, all amounts in rupees):
- Unspent balance of earlier grant brought forward: 1,50,000
- Grant received this year under Sanction 1: 20,00,000
- Interest earned on grant money: 30,000
- Expenditure per ledger: 19,50,000, made up of programme costs
12,00,000, salaries 4,50,000, capital assets 3,00,000
- Note: the capital assets line includes 1,00,000 paid as an advance
to a vendor; the asset has not been delivered.
- Sanction terms: interest to be reported; unspent balance carried
forward; capital assets capped at 15% of the grant.
Task 2: Prepare the receipts-and-spending table in the format above.
Task 3: Test each sanction term against the numbers.
Step 5: Ask for working and assumptions
Show the table line by line with the arithmetic for every total.
Then list every assumption you made, and every place where you needed
a fact I did not give you. Do not treat an unadjusted advance as
expenditure unless the sanction terms say so. Mark any figure you are
unsure about with [CHECK].
Reply to each assumption. Then ask for the scope wording:
Draft two sentences stating what I checked (sanction letter, bank
statement, books, vouchers on a sample basis) and what I relied on
management for. Leave the sample size and dates blank for me.
What a good answer looks like
The figures below are our own computation from the synthetic case, not a transcript. Any AI output is illustrative: wording varies run to run, and you should expect different layouts. What must not vary is the arithmetic.
| Item | Rs |
|---|---|
| Opening unspent balance | 1,50,000 |
| Grant received | 20,00,000 |
| Interest earned | 30,000 |
| Total funds available | 21,80,000 |
| Expenditure (ledger 19,50,000 less unadjusted advance 1,00,000) | 18,50,000 |
| Closing position | 3,30,000 |
Of the closing position, Rs 2,30,000 is cash or bank balance and Rs 1,00,000 is an unadjusted advance, which follows the companion post's rule that an advance not yet adjusted is not reported as spent. Capital assets actually incurred come to Rs 2,00,000, which is 10% of the Rs 20,00,000 grant, so the illustrative 15% cap is met. A good draft states the unspent balance is carried forward, as the sanction terms say.
A weak draft would report expenditure at Rs 19,50,000 and closing at Rs 2,30,000, quietly treating the advance as spent.
Where does AI go wrong on a UC?
- Invented rules. It cites a rule number or time limit you never pasted. Check: any citation must appear in text you supplied.
- Arithmetic slips. Totals may be off by a line, especially with Indian digit grouping. Check: re-add in Excel, and confirm opening plus receipts less spending equals closing.
- Silent assumptions. It may treat advances as expenditure or assume interest is returned. Check: the assumptions list must be complete, and you must agree with each.
- Wrong heads. It may push spending into the wrong component. Check: compare each head with your voucher analysis, not with its table.
- Signatory guesses. It may state that a CA must sign. Check: only the grantor's format and sanction decide, as the companion post explains.
Verify in 60 seconds
Do not accept the table until you have re-run it independently. Open CORAA's fund utilization certificate template, enter the same figures and compare totals line by line. Then re-add opening plus receipts plus interest less expenditure in your own sheet. If the three agree, the AI's structure is safe to use. If not, the spreadsheet wins.
How do ChatGPT, Claude and Gemini differ for this task?
Only differences we are sure of:
- All three let you attach documents such as a PDF or a sheet. Attach only anonymised copies.
- All three let you save a role prompt so you do not retype it: custom instructions or projects in ChatGPT and Claude, and saved instructions in Gemini. Menu names and plan availability change, so check your account.
- Model behaviour, file limits and data settings vary by plan and change often. Read the current data-use terms before use.
Whichever tool you use, the checks above are the same.
Frequently asked questions
Can ChatGPT or Claude issue a utilization certificate?
No. It can draft the layout and re-add figures. The certificate is signed by the person the grantor's format and sanction name, and a CA who signs is professionally responsible for it, including a UDIN.
Is it safe to paste a sanction letter into an AI tool?
Not with identifiers. Remove the client, scheme, sanction number and bank details, or paste only the clauses that matter. Follow your firm's confidentiality policy and the tool's data settings.
Which prompt works best?
One that gives the source and asks the model to apply it, requires line-by-line working, and demands a list of assumptions. Prompts that ask the model to recall rules from memory are the weakest.
How do I know the numbers are right?
Recompute them yourself and compare with the fund utilization certificate template. Never rely on an AI-generated total.
Can AI decide whether spending was allowed under the sanction?
It can compare figures with terms you paste, such as a percentage cap. It cannot judge whether a payment was genuinely for the sanctioned purpose. That needs vouchers and your judgement.
Does using AI change my responsibility as a CA?
No. Responsibility for the certificate stays with you. Document what you checked and keep your own working papers.
Last reviewed: 26 September 2026. This post is educational and not a substitute for the sanction letter or advice from your professional body.
CORAA's internal audit product page describes how engagement working papers and evidence can be kept together. To try it, start a trial.