GST Reconciliation Using AI: A Practical Playbook for 2B, Books and GSTR-3B

GST reconciliation is not one matching exercise. It is a set of timing, identity, eligibility and reporting tests across the purchase register, GSTR-2B, GSTR-3B and the annual return. AI helps most when it keeps those reasons separate.

Build the four-way population

Start with one row per invoice or credit note and retain the source file and period. At minimum, carry supplier GSTIN, document number, document date, taxable value, IGST/CGST/SGST/Cess, 2B period, books period, 3B claim period and an eligibility flag.

Then reconcile in four directions:

1. Books to 2B: recorded purchase that is not available in the statement.
2. 2B to books: supplier-reported document missing from the ledger.
3. 2B to 3B: available credit that was claimed, deferred or omitted.
4. 3B to books: return disclosure that does not tie to the ITC ledger.

An AI classifier can propose a reason code, but it should not invent an invoice match where the GSTIN or document number is materially different.

The exception taxonomy

Use stable labels so a partner can read the dashboard quickly:

• timing difference;
• supplier has not filed or amended;
• GSTIN or document-number mismatch;
• duplicate in books;
• credit note not linked;
• blocked or ineligible ITC;
• reverse-charge or import-of-services treatment;
• claim deferred for a later period;
• manual review required.

The output should show gross value, tax value, owner, next action and ageing. “Mismatch” alone is not actionable.

A safe AI workflow

Ask the model to normalise dates and invoice-number punctuation, cluster probable matches and produce a proposed reason code with confidence. Keep exact matching deterministic first: supplier GSTIN, document number, tax amounts and period. Use fuzzy matching only for the residual population, and send low-confidence matches to a person.

For an exception note, give the model the verified row, the source reference and the firm’s approved wording. Ask it to draft—not decide—whether the item is to be claimed, deferred, reversed or followed up.

Do’s and don’ts

Do keep 2B snapshots by period, retain the pre-clean and post-clean counts, and reconcile the final claimed amount to GSTR-3B. Do record who accepted each material exception.

Do not upload a client’s raw GST data to an unapproved public model. Do not treat a high similarity score as proof of a match. Do not mix 2A, 2B and GSTR-3B as if they were interchangeable evidence.

What to measure

Track the percentage of rows auto-matched, the percentage sent to review, unresolved tax value, average exception age and the number of supplier follow-ups closed. The goal is not a 100% automatic match; it is a smaller, better-explained review queue.

Start with the GST reconciliation working paper, the GSTR-9 ITC mapper and CORAA’s GST reconciliation workflow.