CBIC's analytics arm is already risk-scoring your returns. The "AI drafts your notice" part isn't confirmed.

Tax-tech blogs are circulating a specific claim — that GSTN uses generative AI to draft your GST notice from an anomaly cluster. Some of the underlying capability is real and on the record. The GenAI-drafting part is not confirmed by GSTN or CBIC anywhere we could find. Here's exactly where that line sits, and why it barely changes what you should do about your GSTR-2B/3B gap either way.

The claim doing the rounds

A tax-tech blog — Binary Semantics — describes a GSTN/CBIC "AI roadmap" in which generative models read anomaly clusters (recurring gaps between GSTR-1, GSTR-2B and GSTR-3B) and auto-draft structured notice text under provisions like DRC-01 or ASMT-10, with an officer reviewing before issue. It's a plausible-sounding story, and parts of the surrounding picture are real. But we could not find a single dated GSTN advisory, CBIC circular or Ministry of Finance statement — a Parliament reply, a press release, anything on gst.gov.in or cbic.gov.in — that confirms GenAI is drafting notice text. Until one surfaces, treat that specific claim as third-party characterisation, not GSTN/CBIC fact. (Position as of August 2026 — if GSTN publishes something more specific, this page should be updated.)

What is actually on the record

Three things sit behind the claim, and each is independently verifiable — none of them says "GenAI writes the notice":

• DGARM exists and does exactly the analytics job you'd expect. The Directorate General of Analytics and Risk Management was set up as "an apex body of CBEC for data analytics and risk management" across Customs, Central Excise and GST, tasked to "utilize internal and external data sources for detailed data mining and analysis to generate outputs for focused and targeted action by field formations and investigation wings" (CBEC — renamed CBIC in 2018 — Office Memorandum F.No. A-11013/19/2017-Ad.IV, 11 July 2017). Separately, CBIC's own scrutiny SOPs confirm DGARM selects GSTINs for return scrutiny on risk parameters (Instruction No. 02/2022-GST, 22 March 2022, for FY 2017-18 and 2018-19), and under the later instruction those GSTINs are pushed to the proper officer's scrutiny dashboard on ACES-GST (Instruction No. 02/2023-GST, 26 May 2023, for FY 2019-20 onwards). This is CBIC's own description of the office, not a third-party gloss.
• BIFA (Business Intelligence and Fraud Analytics) is a real, named tool used by field officers. An establishment order of the Chief Commissioner, Customs, Central Excise & CGST, NER Guwahati Zone (25 September 2020) records a Zonal BIFA Cell being created there in compliance with CBIC instructions, and a Press Information Bureau release (7 March 2024) on a Meerut CGST investigation (₹1,048 crore of inadmissible ITC across 232 fake firms) records officers using "various analytical tools such as E-way Comprehensive Portal, Advait and Business Intelligence and Fraud Analytics (BIFA)" to build the case. That is BIFA used for post-hoc fraud investigation by a Commissionerate — not evidence that BIFA (or anything else) is auto-drafting the notice a routine, non-fraud taxpayer receives.
• GSTN has publicly invested in predictive analytics — as a hackathon, not a live drafting tool. GSTN ran a "GST Analytics Hackathon" (PIB, 23 August 2024), inviting students, researchers and companies to build predictive models for a GST analytics framework. That is GSTN's own account of its intent to build better predictive models for tax compliance. It says nothing about generative drafting of notice text, and a hackathon prototype is not a production system.

The intimation that already is automated — and it isn't GenAI

If you want a real example of a GST communication issued with no officer in the loop at the point of trigger, you already have one, and it long predates any GenAI story: the DRC-01C intimation under Rule 88D. It fires automatically from the portal's rules engine when your claimed GSTR-3B ITC exceeds your GSTR-2B ITC beyond a threshold — Rule 88D(3) makes the unexplained excess liable to be demanded under Sections 73, 74 or 74A; the intimation itself is not that demand. FORM GST DRC-01C Part A is headed "(System Generated)" in the Rules themselves.

The distinction matters for how you read the "AI notice" story: DRC-01C is a deterministic, rules-engine comparison of two numbers — not a language model interpreting an anomaly cluster and writing prose. It is real, dated (inserted by Notification 38/2023-Central Tax, 4 August 2023) and already produces an intimation with zero human review at the point of issue. If a truly generative step is ever added upstream of a demand notice, the department hasn't said so yet — but the underlying comparison logic doesn't need to be generative to already be automated and consequential.

What does still run through a human step: Section 61 scrutiny

The other real, citable mechanism worth knowing — and the one ASMT-10 actually belongs to — is return scrutiny under Section 61 of the CGST Act, read with Rule 99. This is genuinely settled, unglamorous law, not an AI story:

The demand that can follow scrutiny sits under Section 73/74 for periods up to FY 2023-24, or the unified Section 74A for FY 2024-25 onwards — three different provisions depending on the period, not one. The instructions above set the selection mechanism and the ASMT-10/11/12 clock; we don't assert the specific risk parameters DGARM scores against, since neither instruction publishes that list.

Why the "who wrote it" question matters less than it sounds

Here's the part that should actually change how you operate: whether the officer's discrepancy notice, or a future demand, is drafted by a person, a template, or eventually a generative model doesn't change what triggers it. Every mechanism above — DGARM's risk-based selection, BIFA's fraud flags, Section 61 scrutiny, Rule 88D's automated comparison — runs on the same underlying returns data: your ITC claim in GSTR-3B, your GSTR-2B, and your GSTR-1 — even though the specific parameters DGARM or BIFA weigh are not published. The department has a standing analytics directorate, a named fraud-analytics tool and a published appetite for predictive models. That's the confirmed trend, GenAI or not.

Which is exactly the point where reconciling before you file — rather than after a discrepancy notice lands — is the only lever you actually control. Waiting to see whether this month's gap gets risk-scored into a drafted notice is a bet against a system the department is visibly investing in; closing the 2B-versus-3B gap before filing, the way tools like Recoup do continuously, is the alternative that doesn't depend on guessing what the department's tooling looks like this quarter.

What to actually do about it this month

• Reconcile your GSTR-3B ITC to your GSTR-2B before you file, invoice by invoice — not after. This is the same discipline that avoids a DRC-01C, and it's the same gap any scrutiny mechanism, human or automated, is reading.
• Action your IMS queue properly. Reject duplicates and wrong-value records before they're deemed accepted into your 2B — an unclean 2B is the raw material every one of these mechanisms scores against.
• Don't manage to a rumoured threshold. Whether it's DRC-01C's unpublished limit or a scrutiny-selection parameter nobody has published, the only defence that works at any setting is a genuinely reconciled position, not a guess at where the line is drawn.
• Keep your invoice-level workpapers current, not just your return. If a discrepancy communication does arrive — an ASMT-10 notice, a DRC-01C intimation, or eventually a Section 74A show-cause — the fastest response is one where the invoice-level explanation already exists, rather than being reconstructed against a 30-day or 7-day clock.

Recoup reconciles your Zoho Books ITC against your GSTR-2B continuously, names the exact invoice and vendor behind any gap, and keeps that workpaper ready — so whatever mechanism the department is running underneath its notices, your position doesn't depend on knowing which one it used.

Related