GST Adjudication in the Age of AI - Assistance, Not Substitution

AI Enters Tax Adjudication - Where Should the Line Be Drawn?

Artificial Intelligence( (AI) is rapidly entering professional decision-making. Legal research, drafting, summarisation and identification of authorities can now be undertaken with remarkable speed. Tax administration cannot remain untouched by this technological change. Properly used, AI can help an adjudicating officer locate statutory provisions, identify judicial precedents and organise complex factual material.

But adjudication is fundamentally different from mechanical information processing. A show cause notice represents the formation of a statutory view by the proper officer. An adjudication order determines civil consequences after considering the taxpayer's defence. The responsibility to understand the facts, verify the law and apply an independent judicial mind therefore remains with the statutory authority.

The Gujarat High Court has now addressed this concern directly in Faiz Enterprise Through Proprietor Mahetar Mahir Farukbhai Versus State Tax Officer, Unit-67, Surat & Anr. - 2026 (8) TMI 1389 - GUJARAT HIGH COURT . The ruling is significant because the impugned GST proceedings relied on AI-generated authorities that were either non-existent or irrelevant. The High Court characterised such adjudication as fundamentally defective and procedurally unsustainable.

The judgment is not merely about incorrect citations. It raises a much larger question for contemporary tax administration: How far can Artificial Intelligence participate in a statutory decision that the law requires a human authority to make independently?

When Technology Enters the Adjudication Room

The controversy arose out of proceedings culminating in the cancellation of the petitioner's GST registration. The State Tax Officer passed the cancellation order dated 15.12.2025, relying on certain judicial decisions.

When the matter reached the Gujarat High Court, the petitioner highlighted a disturbing feature of the adjudication. One of the authorities cited in the order apparently did not exist, while the citations or ratios attributed to other judgments did not align with the propositions for which they were relied upon.

The petitioner specifically pointed out that State of Gujarat v. Aarbee Structures Pvt. Ltd., cited, did not exist. It was further submitted that the Madras High Court decision in M/s Sri Vinayaga Agencies v. Assistant Commissioner had been cited with a different citation, and its ratio was inapplicable. A similar objection was raised regarding Jyoti Tar Products Pvt. Ltd. v. Deputy Commissioner, whose citation and ratio did not match the controversy before the authority.

On 13.08.2026, the High Court recorded its prima facie impression that the impugned order appeared to have been passed by placing exclusive reliance on AI-generated case law. That observation transformed what might otherwise have been an ordinary challenge to the cancellation of registration into an important examination of the permissible role of AI in quasi-judicial administration.

An Apology That Revealed a Larger Institutional Problem

On the subsequent date of hearing, the officer who had passed the impugned order filed an affidavit and tendered an unconditional apology. He acknowledged that Artificial Intelligence had been used while drafting the order and that it referred to judgments that turned out to be non-existent or irrelevant.

The explanation was that the officer was probationary and the mistake had occurred due to lack of experience. The affidavit also informed the High Court that, after its earlier order, the officer had attended a training programme on the use of Artificial Intelligence for drafting special civil applications and for adjudication/appellate orders.

The case therefore illustrates a problem far wider than the individual mistake of one officer. Generative AI can produce legal language of considerable fluency. It can provide citations, summaries and apparently reasoned propositions in a format that closely resembles conventional legal research. But linguistic confidence is not proof of legal correctness. A citation may be fabricated, a genuine judgment may be associated with an incorrect proposition, or a valid precedent may have subsequently been reversed or distinguished.

That distinction is particularly important in tax adjudication because a seemingly polished order can still be legally defective if the authority relied upon either never existed or does not apply to the facts.

AI Can Find an Authority - It Cannot Verify Itself

The administrative response to the controversy was immediate. Following the High Court's intervention, the Additional Commissioner of State Tax issued detailed instructions dated 18.08.2026 governing the use of Artificial Intelligence by adjudicating and quasi-judicial authorities.

The first principle emerging from those instructions is elementary yet critical: anything identified through AI must be independently verified. This applies not only to case law but also to statutory provisions, rules, circulars and notifications.

When superior court judgments are relied upon, authorities have been instructed to ensure that the citation is accurate, the decision is relevant to the dispute, the source is authentic, and the judgment remains good law. The legal status of the judgment must therefore also be checked so that an authority does not unknowingly rely on a precedent subsequently overruled by a superior forum.

This is an important practical distinction. AI-assisted research may inform an officer that a judgment exists. That does not relieve the officer of the obligation to open the judgment, read the material portion, understand its factual context, and satisfy himself that the proposition attributed to it actually emerges from the decision.

The technology can accelerate the search, but verification remains a human responsibility.

A Paraphrased Ratio Is Not a Substitute for Reading the Judgment

Another significant feature of the Gujarat instructions is AI-generated summaries of judicial decisions.

Artificial Intelligence often provides a condensed statement of what a court supposedly decided. Such summarisation may be useful at the preliminary research stage, but a statutory order cannot safely proceed on that summary alone. Accordingly, the instructions require that, where an AI tool paraphrases the ratio of a judgment, the officer must locate the original relevant paragraph before relying on it.

The importance of this safeguard should not be underestimated. Judicial observations derive meaning from their factual and statutory context. A sentence extracted from one dispute cannot be transplanted into another merely because the same expression appears in both.

In GST litigation, small factual distinctions can materially alter the outcome. The tax period may differ. The governing provision may have been amended subsequently. A judgment may address registration, whereas the pending adjudication concerns input tax credit or valuation. Even within the same statutory provision, the factual foundation required to apply a precedent may be absent.

AI can summarise a judgment. It cannot relieve an adjudicating authority of the need to understand why that judgment applies.

The Taxpayer's Case Law Cannot Be Ignored

The administrative instructions go beyond verification of authorities discovered through AI. They specifically require adjudicating authorities to consider and discuss judgments cited by the taxpayer. If an authority considers such a decision inapplicable, reasons must be recorded for reaching that conclusion.

This requirement reinforces an established principle of reasoned adjudication. Application of mind is not demonstrated merely by reproducing authorities supporting the Department's view. The taxpayer's legal defence must also be addressed.

An order that cites several judgments but fails to examine the authorities relied upon by the noticee may appear elaborate on the surface while remaining substantively unreasoned. AI can make such an order longer and more polished, but it cannot cure the absence of engagement with the defence.

The quality of adjudication is therefore not measured by the number of citations. What matters is whether the authority has understood the competing propositions and explained why it accepts one and rejects the other.

The Most Important Requirement - Human Oversight

The Gujarat instructions expressly recognise that AI tools may be used to improve the efficiency and effectiveness of adjudication. The approach is therefore not one of technological rejection.

The safeguard lies in human oversight. The instructions acknowledge concerns about the accuracy, transparency, and the underlying information used by AI tools, and therefore require that their use remain subject to appropriate human supervision. More importantly, every order must reflect independent legal reasoning and the officer's own application of mind to the facts and law. AI output may assist the reasoning process; it cannot substitute for it.

This perhaps represents the most balanced principle emerging from Faiz Enterprise. The issue is not whether a government officer has used technology. Modern administration inevitably uses technological tools. The decisive question is whether the statutory authority continues to perform the function entrusted to it by Parliament.

If AI researches while the officer decides, technology remains an aid. If AI effectively determines the reasoning and the officer merely adopts its output, assistance has crossed into abdication.

The Signature on the Order Carries the Responsibility

The instructions also make another proposition unambiguous: ultimate responsibility for an order's correctness lies with the authority issuing it, regardless of the technological tools used. Failure to comply with the instructions, including passing orders based on unverified AI-generated material, may amount to misconduct and attract disciplinary action.

This principle has consequences far beyond AI. An adjudicating officer cannot justify an erroneous statutory order by claiming that a junior officer prepared the draft, a database supplied an incorrect judgment or a software application generated the reasoning. The legal order remains the act of the statutory authority whose name and designation appear on it.

Technology may change how a draft is prepared, but it does not transfer statutory accountability from the decision-maker to the tool.

From Departmental Instruction to Judicial Mandate

Faiz Enterprise is particularly significant because the High Court did not merely take note of the administrative instructions dated 18.08.2026. It directed that those instructions must be followed scrupulously and further stated that violating them would amount to contempt of the Court.

The instructions therefore acquired immediate practical importance for GST adjudication within the State.

The prescribed safeguards are neither complicated nor anti-technology. They require verification of primary sources, correct citation, examination of relevance, verification of the present legal status of judgments, consideration of taxpayer authorities, human oversight, and independent reasoning.

In substance, these are the same disciplines that good adjudication required even before generative AI entered legal practice. Artificial Intelligence has simply made their observance more urgent because it can produce convincing but incorrect legal material at extraordinary speed.

The Entire Chain of Proceedings Fell

The defect in the present case did not merely result in the deletion of an incorrect citation from the cancellation order. The High Court set aside the entire chain of proceedings.

The show cause notice dated 30.10.2025, the cancellation order dated 15.12.2025, the order rejecting the revocation application dated 10.03.2026 and issued on 23.03.2026, the appellate order dated 25.05.2026, and GST APL-04 dated 26.05.2026 were all quashed.

The authority was directed to issue a fresh notice. The taxpayer was required to cooperate with the fresh proceedings, and its reply and defence were to be duly considered before passing a reasoned order in accordance with law. All rights and contentions were left open.

The remedy adopted by the Court is revealing. The taxpayer was not granted immunity from lawful proceedings merely because AI had been improperly used. Instead, the Court erased the defective decision-making process and permitted the authority to begin afresh in the manner contemplated by law.

The Punjab & Haryana High Court Had Already Sounded the Warning

The principle emerging from Faiz Enterprise may usefully be read alongside the earlier decision of the Punjab & Haryana High Court in M/s. SRO India Versus State Of Punjab And Another - 2026 (7) TMI 1113 - PUNJAB AND HARYANA HIGH COURT.

There, Form GST DRC-01A and the accompanying material uploaded on the portal showed traces of AI-assisted drafting. The uploaded material included suggestions to improve the proposed reasoning and to prepare the order. The taxpayer therefore challenged the proceedings on the ground that the statutory authority had failed to apply its mind independently.

The Punjab & Haryana High Court emphasised that the statute expects the competent authority to examine the facts and apply its own mind before issuing a notice. Finding that the notice had primarily relied on an AI tool, the Court quashed the DRC-01A and consequential proceedings, while preserving the competent authority's liberty to act afresh after independently examining the facts.

The factual settings of the two cases are different, but the underlying concern is remarkably similar. SRO India dealt essentially with AI-assisted preparation of the notice and the absence of independent application of mind. Faiz Enterprise went a stage further and exposed the danger of AI-generated legal authorities entering a final adjudication order without verification.

Together, the judgments are beginning to define the judicial boundary between legitimate technological assistance and impermissible substitution of statutory judgment.

Independent Application of Mind Is Technology-Neutral

The enduring significance of these rulings is that they should not be read as creating an exceptional legal principle applicable only to Artificial Intelligence.

Independent application of mind has always been indispensable to quasi-judicial proceedings. AI merely provides a new context in which that principle must operate.

The same defect would arise if an adjudicating officer mechanically copied another order, reproduced an investigation report without examining the taxpayer's defence, or accepted a draft prepared by somebody else without independently considering the material. What matters is not who or what prepared the first draft. What matters is whether the statutory authority ultimately made the decision.

The test is therefore technology-neutral: Does the notice or order reflect the judgment of the officer whom the statute authorised to act?

The Real Risk Is Not AI - It Is Unverified AI

It would be equally incorrect to treat the two High Court rulings as establishing that Artificial Intelligence must never enter tax administration.

AI can significantly improve research efficiency. It can help identify relevant provisions, locate possible authorities, structure large records, and detect issues requiring closer examination. Used responsibly, it may even improve the quality and consistency of adjudication.

The difficulty arises when efficiency is mistaken for reliability. An AI-generated citation must be checked. An AI-generated ratio must be compared with the original judgment. An AI-generated statutory proposition must be tested against the bare provision and relevant amendments. An AI-generated conclusion must be reconsidered after examining the taxpayer's response. AI should therefore be treated as a research assistant, not an adjudicating authority.

Author's Perspective - Technology May Assist the Mind, Not Replace It

Faiz Enterprise and SRO India draw a clear boundary for AI in tax adjudication. AI may assist with research and drafting, but its output must be independently verified, and the statutory authority must itself examine the facts, the law and the taxpayer's defence.

The principle is straightforward: technology may assist human judgment, but it cannot replace statutory satisfaction or quasi-judicial responsibility. The decisive question must remain not what AI suggested, but what the proper officer independently decided-and why.

***