AI for GST compliance: from 12 hours to 2 hours a month

GST is one of those problems that looked AI-shaped from day one and stayed stubbornly resistant for three years. The data is structured. The rules are written down. The reconciliation is mechanical. Yet the actual work, until very recently, still ate one accountant for a few days every month at every Indian SMB we've worked with.

That changed in 2026. Not because LLMs got smarter, but because the surrounding infrastructure finally caught up. This is what works now, what still doesn't, and why a small textile manufacturer can plausibly cut GST workload from 8-12 hours a month to 2-3.

Why it didn't work in 2022

For context: when GST went fully digital, the moving parts were already in place. GSTN APIs existed. The Invoice Registration Portal (IRP) issued IRNs. GSTR-1 and GSTR-3B filings were structured XML. In theory, a script could pull all of it.

In practice, three things kept it manual.

API access was uneven. GSTN APIs were gated by GSPs (GST Suvidha Providers). Direct access for SMBs was painful. The official endpoints had quirks that varied by state.

Accounting software lagged. Tally was the dominant tool but its API surface was limited. Zoho Books and Marg had better APIs but smaller installed base. Most reconciliation was still happening in Excel.

The "soft" judgement calls were everywhere. Is this expense eligible for ITC? Is this reverse charge? Does this exemption apply? The rules are written but ambiguous at the edges, and SMBs don't have the in-house tax brain to resolve them quickly.

LLMs in 2022-2023 couldn't help with any of this reliably. The prompt would happily hallucinate a Section number that didn't exist.

What changed by 2026

Three things, in order of impact.

Tally and Zoho Books shipped real AI features. Not chat windows pasted on. Actual automated reconciliation, anomaly flagging, ITC suggestions, and IRP push from inside the product. Tally AI in particular, as it matured through 2025, made the per-invoice classification step nearly free.

E-invoice rules tightened. The threshold dropped progressively. By 2026 a much larger slice of B2B invoices is e-invoiced via the IRP, which means the data is structured at the source, with an IRN and QR code. Less PDF parsing, more direct API consumption.

Frontier models got accurate enough on Indian tax content to be useful as a first-pass classifier. Claude Sonnet 4.6 and GPT-5 will give you a defensible answer on "is this expense eligible for ITC under Section 17(5)" with citations to the Act. They're not infallible, but they're better than most non-CA staff.

The combination is what made the workflow cleanly automatable.

What we actually wire up

A typical engagement for an SMB textile manufacturer:

• Source data: Tally for sales and purchases, IRP for inbound e-invoices, GSTN for GSTR-2A/2B downloads, bank statements for reconciliation.
• Pipeline: A self-hosted n8n workflow pulls all sources daily into a Postgres staging table.
• AI layer: An agent classifies each line. Is the GST rate right? Is the HSN right? Is ITC eligible? Is this a reverse-charge case? It flags anomalies and writes them to a queue.
• Human layer: The owner or accountant reviews flagged items in a simple web UI. Approve, reject, ask for more info.
• Filing layer: Approved data flows back to Tally and forward to GSTR filings. The CA reviews the consolidated return.

The agent is doing the labour-intensive comparison work. The human is doing the judgement work. The CA is doing the strategic review. Each does what they're good at.

-- A typical anomaly query the agent investigates
SELECT
  i.invoice_number,
  i.vendor_gstin,
  i.taxable_value,
  i.gst_rate,
  v.gst_rate AS gstr2a_rate,
  v.taxable_value AS gstr2a_value
FROM purchase_invoices i
LEFT JOIN gstr2a_lines v
  ON v.invoice_number = i.invoice_number
  AND v.vendor_gstin = i.vendor_gstin
WHERE
  v.invoice_number IS NULL
  OR ABS(i.taxable_value - v.taxable_value) > 1
  OR i.gst_rate <> v.gstr2a_rate
ORDER BY i.invoice_date DESC;

That single query, run nightly, surfaces 80% of what an accountant used to find manually. The agent looks at each row, classifies the likely cause (vendor hasn't filed yet, rate mismatch, duplicate, fraud), and proposes an action.

What still needs human eyes

We're explicit with clients about the limits.

ITC ambiguity. Section 17(5) has a long list of blocked credits with edge cases. Membership fees, food and beverages, employee-related expenses, motor vehicles. The model is good at the obvious cases and confident on the wrong cases. We always send borderline ITC to a human.

Reverse charge edges. Some services are reverse charge, some aren't. The thresholds and sector rules change. We give the model the current rule set in context, but we still keep a human in the loop for the unusual cases.

Sector exemptions. Textile, leather, agricultural products, exports. The exemption rules are sector-specific and change at the budget. We keep a versioned rule set per sector and update it on schedule. We do not let the model "remember" the rules from training data.

Notices and replies. When a GST notice arrives, the model can summarise it and draft a response. We do not let it file the response. That's the CA, always.

The mistake we keep seeing in DIY setups: clients let the agent file the return directly because the dashboard makes it easy. Don't. The CA's job changed. It didn't disappear.

What the numbers actually look like

A real client, 35-person textile manufacturer in Surat. Before:

• Senior accountant: 8-12 hours/month on GST reconciliation.
• Junior staff: another 4-6 hours/month on data entry.
• CA: ₹20-25,000/month retainer, mostly compliance.
• Annual losses from missed ITC and late filings: estimated ₹40-60,000.

After about three months of the workflow being live:

• Senior accountant: 2-3 hours/month, almost all of it judgement calls and approvals.
• Junior staff: data entry mostly gone, now doing customer follow-ups.
• CA: ₹8-10,000/month, strategic review only.
• Missed ITC: tracked to near zero through the optimisation suggestions.

The CA's fee dropped, but the CA's hour rate went up because they now do interesting work. The CA, we think, will be fine.

The honest summary

GST compliance is one of the cleanest examples of where the new AI stack actually delivers in India. The infrastructure is mature, the data is structured, the models are accurate enough, and the workflow is well-bounded.

It is also one of the places where the worst kind of vendor pitch is most common. "AI GST" SaaS is having a moment. Most of it is a thin wrapper over GSTN APIs and a chatbot. Before paying ₹3,000/month for one, ask whether your existing accounting software's AI features cover 80% of what you need. They usually do.

If they don't, a custom workflow on top of Tally and the IRP is well within reach for a small studio engagement. We have built several. They are not exotic.

The accountant's job got more interesting. The CA's job got more interesting. That's what good automation does.