India’s tax authorities detected fake Input Tax Credit worth over Rs 1.14 lakh crore between 2020 and 2025, most of it routed through shell companies that exist only on paper. The bottleneck isn’t finding fake firms anymore, GSTN’s own risk engines already flag thousands of suspicious registrations every year.
The bottleneck is what happens next: connecting a flagged entity to every other shell company the same operator controls, verifying whether goods actually moved, and building an evidence chain fast enough to act before the money exits the network.
That is the specific problem agentic AI workflows are built to close, and it is worth being precise about what “agentic” means here versus what is still a manual, officer-driven grind.
Key Takeaways
- Shell company networks follow a repeatable pattern: one operator, multiple PANs, no genuine business activity, circular trading to inflate turnover
- The current detection model works in silos, GSTN, MCA records, e-way bills, and FASTag data sit in separate systems that investigators cross-reference manually
- Real shell company cases show 6 to 70+ linked firms per network, often built by misusing the Aadhaar and PAN details of unrelated individuals
- Agentic AI means a system that chains investigative steps on its own, an address anomaly triggers a cluster search, which triggers a movement check, which triggers an evidence compilation, without an officer manually initiating each query
- No fully autonomous, end-to-end shell company investigation system is deployed anywhere in India today, current tools accelerate correlation and evidence gathering, they don’t replace the officer’s judgment or the legal process
- Prophecy Eagle I is built for exactly this data fusion layer within the GST and indirect tax ecosystem, and nothing beyond it
The Shell Company Problem, By the Numbers
The scale here is not abstract. In the first Special All-India Drive against fake GST registrations (May to July 2023), tax authorities found 21,791 entities that did not actually exist. The second drive, run from August to October 2024, went further: of 67,970 GSTINs flagged for verification, 27 percent were confirmed non-existent, accounting for Rs 10,179 crore in detected evasion and Rs 2,994 crore in blocked ITC. Source: Times of India.
Parliament has been given similarly large numbers directly. In a written Lok Sabha reply, the Minister of State for Finance disclosed that GST officers detected Rs 35,132 crore in ITC evasion linked to 17,818 fake firms in a single reporting window, with 69 arrests. Source: CFO Economic Times.
These aren’t edge cases buried in footnotes. This is the baseline rate at which shell companies are being manufactured to defraud the exchequer, and every one of these numbers represents cases that were eventually caught, not the ones still active.
How Shell Company Networks Actually Operate
Publicly reported enforcement cases show the same modus operandi recurring with minor variations.
In Bengaluru, DGGI unearthed a network of six shell companies built around a chartered accountant who also served as statutory auditor for some of the entities. The companies had no real business activity, generated fake invoices worth Rs 266 crore, and used circular trading to inflate turnover before claiming Rs 48 crore in fraudulent ITC. One of the six was listed on a stock exchange, which is why DGGI shared its findings with SEBI as well. Source: PIB
In Thane, a single individual registered 22 non-genuine firms using misappropriated Aadhaar and PAN details, availing Rs 48 crore and passing on Rs 44 crore in fraudulent ITC. Source: PIB A related case in Palghar involved 70 shell firms built around Rs 320 crore in fake invoices. Source: The Print. In Lucknow, CGST officers traced 131 fake non-existent firms that had shown Rs 1,355.74 crore worth of supplies to roughly 1,100 recipient firms across Uttar Pradesh, Haryana, and Delhi, passing on Rs 197.20 crore in ineligible credit, without a single unit of actual goods movement. Source: PIB.
Strip away the specific numbers and the pattern is consistent:
- One operator, many PANs: The same individual registers multiple companies using different PAN cards, or misuses the identity documents of people who have no idea their KYC details are being used.
- Directorship stacking: One person shows up as director across an unusual number of entities, often with fabricated or forged address proofs like fake rent agreements and electricity bills.
- No genuine activity, real paper trail: These firms generate invoices, GST filings, and e-way bills, everything looks compliant on paper, but no goods or services actually move.
- Circular trading: Entities in the network trade with each other repeatedly to inflate turnover and legitimize the appearance of business volume.
The mechanics aren’t sophisticated. What makes them hard to catch quickly is that each of these signals lives in a different data source.
Why the Current Investigation Model Struggles
Here’s the honest structural problem. A GST officer working a suspicious registration has to check it against GSTN filing history, cross-reference the director’s PAN against MCA’s company registration data, pull e-way bill records to see if the claimed goods movement is corroborated, and check FASTag data to confirm whether a vehicle physically made the trip an invoice claims it did.
Each of these is a separate system. None of them talk to each other by default. An investigator doing this manually is essentially running the same query, five or six times, across five or six different portals, then manually cross-referencing the outputs in a spreadsheet or on paper. By the time a pattern surfaces, the network has often already moved on, either by winding down the flagged entities or by spinning up new ones under fresh identities.
This is precisely why the two Special All-India Drives needed months of dedicated, coordinated effort involving both Central and State tax administrations just to work through a batch of flagged registrations. The detection signal existed. The correlation work is what consumed the time.
What “Agentic” Actually Means, Versus Just Automation
There’s a meaningful difference between an alert system and an agentic one, and it’s worth being precise about it rather than using “agentic AI” as a buzzword.
A rules-based alert system flags an anomaly and stops. Someone registered ten companies under different PANs this week, alert generated, ticket raised, an officer picks it up whenever they get to it.
An agentic workflow keeps going. It takes that same flag and asks the next investigative question on its own: does this individual also appear as a director elsewhere? Do any of those other entities share a registered address? If they do, does the transaction pattern between them look like circular trading, and can that be checked against e-way bill records for the same period? If the e-way bills claim goods moved, does FASTag data for the relevant route and timeframe actually corroborate a vehicle making that trip?
Each answer determines what gets checked next, without an officer having to manually initiate each individual query. That’s the actual distinction. It’s not that the AI is smarter, it’s that the investigative sequence that used to require a human to manually chain five separate lookups now runs as one continuous workflow, and what lands on the officer’s desk is a compiled, cross-verified picture instead of five disconnected data points they have to stitch together themselves.
Where This Fits Today
This is exactly the layer Prophecy Eagle I, Innefu’s Financial Fusion Centre, is built for, and it’s worth being specific about the boundary rather than vague about it. Eagle I fuses GSTN filing data, PAN and directorship records, e-way bill data, and FASTag records to flag the specific patterns described above: one individual behind multiple company registrations, unusual directorship counts, and GST transactions where the paper trail doesn’t match physical movement of goods. When a network like the ones described above starts forming, that’s the signal set Eagle I is designed to surface, and to keep surfacing across the network as new linked entities appear.
What it does not do, and this matters for anyone evaluating it, is function as an anti-money-laundering or SAR-filing tool, and it does not extend into direct tax or PMLA territory. Its data ecosystem is GSTN, e-way bill, FASTag, and company registration data. If a shell company investigation moves into layered money laundering through the banking system or benami property holdings, that’s a different data ecosystem entirely, and claiming otherwise would be overselling what the tool actually does.
Request a demo of Prophecy Eagle I.
Frequently Asked Questions
1. What is a shell company in the GST context?
A shell company in the GST context is a registered entity with no genuine business activity that exists solely to generate invoices, claim or pass on fraudulent Input Tax Credit, or inflate turnover through circular trading with other linked entities.
2. How do Indian tax authorities currently detect shell companies?
Detection primarily happens through GSTN’s data analytics and risk profiling (via the Directorate General of Analytics and Risk Management), coordinated Special All-India Drives involving physical verification, and DGGI intelligence operations that trace director and PAN patterns across multiple company registrations.
3. What data sources are needed to trace a shell company network?
A complete picture typically requires GSTN filing and registration data, MCA company and directorship records, e-way bill data to check claimed goods movement, and FASTag records to verify whether vehicles actually made the trips an invoice claims.
4. Is agentic AI already deployed in Indian shell company investigations?
Not in a fully autonomous, end-to-end form. Financial fusion tools that cross-reference GST, directorship, and logistics data to flag networks are in use, but the investigation, verification, and prosecution steps still require human officers.
5. What’s the practical difference between an automated alert and an agentic AI workflow?
An automated alert flags one anomaly and stops there. An agentic workflow uses that flag to trigger the next relevant check on its own, tracing linked entities, verifying movement data, and compiling a correlated evidence picture, without an officer manually initiating each step.
[Internal link: Fake Invoicing and GST Fraud blog] [Internal link: How AI Tools Are Helping the GST Department Detect Evasion] [Internal link: E-Way Bill Fraud Analytics blog, once live] [Internal link: Prophecy Eagle I product page]
