Issuu

AI-Based Risk Detection in Indian Legal Contracts Using NLP and Large Language Models

Shiva Vishwakarma¹, Utkarsh Srivastava², Etisham³, Devesh Katiyar´, Gaurav Goelµ

¹²³ Students, ´µ Assistant Professors Dr. Shakuntala Misra National Rehabilitation University, Lucknow, Uttar Pradesh, India.

ABSTRACT - Indian legal contracts present unique difficulties for automated risk detection as they blend very old legalese with jurisdiction specific statutory references and switch between English and regional languages at different points of the document and have been evolving very rapidly in recent years with the Digital Personal Data Protection Act 2023, the Arbitration and Conciliation Act 1996 and the Real Estate (Regulation and Development) Act 2016. This paper discusses the use of Natural Language Processing (NLP)andLargeLanguage Models(LLMs)to detecthigh risk clauses in Indian commercial contracts by using approaches such as Named Entity Recognition and clause classification to prompting GPT-4 inference and Retrieval augmented generation (RAG) based on Indian statutory law. The main finding is that LLMs markedly outperform classical NLP when it comes to clause level contextual risk identification, yet the lack of an annotated Indian commercial contract corpus remains the single most critical constraint facing the field. The paper proposes a four stage hybrid pipeline that brings together In LegalBERT based clause classification, LLM driven risk reasoning, and Indian legal knowledge retrieval. These results are immediately applicable to Indian law firms, corporate counsel and legal tech startups seeking to scale contract review past the ability ofmanualreviewtosupport.

Key Words - IndianlegalNLP,contractriskdetection, LLMs,InLegalBERT,DPDPA2023,IndianContractAct, clauseclassification,RAG,legaltech.

1. INTRODUCTION

1.1. The Problem in Practice

Take a technology vendor agreement signed by a mid sized Indian IT firm in 2021.Enshrined in Clause 18.4 of a 220 page document was an indemnity clause that,readinconjunctionwiththelimitationofliability clause in Clause 9.2, rendered the vendor financially liable for any data breach in an unlimited amount of time without any clarification of what constitutes a breach, what law is applicable or what notice obligationsthevendorisunder.Theproblemwasonly discovered when a dispute emerged 18 months after

signingthedealandatthistimethelegalfeestodefend the claim had already eaten in to the entire annual value of the contract. This is hardly an isolated incident. Indian companies process thousands of commercial agreements each year spanning vendor arrangements, real estate deals, technology licensing, employment relationships, and mergers and acquisitions. Junior lawyers who devote an estimated 70–80% of their billable hours to routine contract reading [1] remain both the primary review mechanismanditsmostvulnerablepointoffailure.

1.2. Why Indian Contracts Are Uniquely Difficult

Indian commercial contracts are unusually difficulttoworkwith,forfourdistinctreasons.Initially they are embedded in a multi-tiered statutory structure: the Indian Contract Act 1872 (ICA) the Companies Act 2013 the Arbitration and Conciliation Act 1996 the Information Technology Act 2000 RERA 2016 and more recently the Digital Personal Data ProtectionAct2023DPDPA.

Secondly, Indian legal drafting is steeped in colonial era law and idioms, and so you see ‘notwithstanding the aforesaid’, ‘save and except’, and ‘mutatismutandis’allthetimeandthosephrasesneed to be interpreted by the law and not by the syntax engine.

Third, contracts in property and construction often change between the languages, with the clause text being in English and the property description or the schedule references being in Hindi or other regional languages. Fourth, a few of the most critical risk indicators, whether a seat in an arbitration clause isdifferentfromtheplaceofarbitration,whenisasum specified in a contract under ICA Section 74 a penalty and when is it liquidated damages, and whether a non compete clause can be enforced under ICA Section 27, are all India only and are not part of any current legal NLPbenchmark.

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 09 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072

1.3. Research Questions and Scope

Three research questions structure this paper. RQ1: which types of risk clauses appear most frequentlyandcarrythegreatestlegalconsequencesin Indian commercial contracts? RQ2: which NLP and LLM techniques perform best when detecting those risks? RQ3: what are the principal difficulties in applying LLMs to Indian legal language? The scope is limited to English primary commercial contracts governed by Indian law vendor agreements, employment contracts, technology and SaaS agreements, real estate transactions, and M&A documentation.Criminal,constitutional,andfamilylaw contextsfalloutsidethisscope.

2. LITERATURE REVIEW

2.1. NLP for Legal Contract Analysis Global Work

The Contract Understanding Atticus Dataset (CUAD), introduced by Hendrycks et al. [2], set the foundational benchmark for contract risk detection, covering 510 contracts annotated across 41 clause categories.

LegalBERT, developed by Chalkidis et al. [3] by continuing BERT pre training on 12 gigabytes of legal text court opinions, legislation, and contracts showed convincingly that domain adapted language models outperform general purpose models on legal classification tasks. ContractNLI [4] recast clause risk detection as a Natural Language Inference problem, makingitpossibletouseentailmentbasedmethodsto check whether a given clause satisfies or violates a normativecondition.Collectively,thesepapersshowed that while models are able to perform interpretable clause level semantic role labeling of legal text, this is only the case when models are trained on domain specificdata.

2.2. LLMs for Legal Reasoning

The use of instruction tuned LLMs on legal tasks hasmovedquicklysince2022.BommaritoandKatz[5] showedGPT-4scoringatthe90thpercentileontheUS UniformBarExaminationwithoutanytaskspecificfine tuning.

Commercial legal AI platforms, including Harvey AI and CasePilot, have since deployed GPT-4 based systems for contract review, document drafting, and case research. One limitation that becomes especially serious in the Indian legal context is hallucination models fabricating plausible sounding legal citations, getting the jurisdiction of statutory provisions wrong,

or generating legally incorrect interpretations while appearingfullyconfident[6].Thisfailuremodecarries higher stakes in legal work than in most other domains,sincebothpractitionersandclientsarelikely toactonwhateverthemodelproduces.

2.3. Indian Legal NLP Existing Work and the Gap

Indian legal NLP research has concentrated mainly on case law, with commercial contracts receivingfarlessattention.Pauletal.[7]introducedIn LegalBERT, a BERT model pre trained on Indian Supreme Court judgments, which outperformed LegalBERT trained predominantly on US and EU legal corpora on Indian legal text classification tasks. Research groups at IIT Bombay and NLSIU Bangalore have built corpora for Indian case law analysis, judgment summarization, and statutory citation extraction. What is missing and what this paper directlytakesonisanydedicatedannotatedcorpusfor Indian commercial contracts. Nothing comparable to CUAD exists for the clause types, risk taxonomies, and statutory references that characterize Indian commercial agreements. As a result, every existing model applied to Indian contracts is working under a distribution shift, having been trained on US or UK contract language and then applied to Indian legal draftingconventionsthatdifferinmaterialways.

3. RISK TAXONOMY FOR INDIAN CONTRACTS

3.1.

Definition and Classification Framework

Contractual risk is evaluated at the level of the analysis. clause level along three dimensions: ambiguity (how far a clause admits multiple legally significant readings). imbalance (how far a clause tilts obligations liability or rights towards one party) and enforceability gap (how. According to Indian law a clause may be void voidable or unenforceable. Compared to straightforward risk/non-risk labeling this three part framework is more practically useful and naturally maps. onto the framework for risk scoringdescribedinSectionVII.

3.2. High-Risk Clause Types in Indian Contracts

With their roots in Indian statutory law eight clausecategoriesstandoutashavingthehighestriskin the Indian commercial context. Indian vendor and technology agreements frequently contain indemnity clauses that subject the service provider to unlimited one sided indemnity obligations requiring them to protect the client against any and all losses resulting from any breach without defining the breach limiting

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 09 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072

exposure or distinguishing between direct and consequential loss. Since there is no statutory cap on contractual indemnity under the ICA unlimited indemnityclausesarelegallypermissibleunder Indian lawbutcouldbedisastrousinreallife.Duetoseatand venue confusion arbitration clauses governed by the Arbitration and Conciliation Act 1996 carry a high degreeofrisk.

The Supreme Court of India has held that the legal seat of arbitration determines the curial law (BharatAluminiumCo.v.KaiserAluminium,2012)but many Indian contracts only specify a "venue" without clarifying if that venue is the seat – a distinction that ultimately determines which High Court will have supervisoryjurisdictionovertheproceedings.

Limitation clauses attempting to exclude liability forfraudorwillful default maybevoidunderSections 23 and 28 of the ICA as they prohibit agreements to limit legal remedies. Any such clause that purports to deprive the aggrieved party of the right to approach a courtfor relief isunenforceable – but suchclausesare routinely included in technology outsourcing agreements drafted on foreign law templates and broughtintoIndiancontractswithoutchange. Penalty and liquidated damages clauses fall under ICA Section 74, which entitles an aggrieved party only to ‘reasonable compensation’ regardless of whether the sum named in the contract represents a genuine preestimate of loss or a penalty. This position differs sharply from English law and means that punitive penalty clauses are only partially enforceable. Data privacy clauses in contracts drafted before the DPDPA 2023 are systematically inadequate: they typically say nothing about data localization requirements, rights notificationsfordataprincipals,ortheDataProtection Officer designation that significant data fiduciaries are nowrequiredtoputinplaceundertheAct.

3.3. Contract Types and Domain-Specific Risks

Table-I: High-RiskClauseTypesinIndianCommercial Contracts

Indemnity Unlimited/one-sidednoICAcap Arbitration Seatvs.venuecuriallaw ambiguity

Limitation of Liability VoidunderICA§23/28if restrictscourtaccess

Liquidated Damages ICA§74only'reasonable compensation'awarded

Data Privacy DPDPA2023compliancegapsin

legacycontracts

Non-Compete LargelyvoidunderICA§27in employmentcontext

IP Assignment

Force Majeure

Broadworkforhirestripping employeeIPrights

COVIDgappandemic/cyber eventsoftenexcluded

Contracts for sale of immovable property under RERA 2016 have their own risks: the penalty clauses for delay of possession must be compliant with the RERA five year minimum period for delivery of possession of the property; and agreements entered into between a developer and a buyer of property in respect of any project by the developer in a State/Union Territory that has notified the date of completion of the project under RERA often include developer friendly force majeure provisions which RERAauthoritieshave,inthepast,struckdown.

4. CLASSICAL NLP TECHNIQUES

4.1.

Clause Segmentation and Multi-Label Classification

Indian legal documents present sentence boundary challenges that go well beyond anything encountered in US contracts. Commercial drafting in India routinely deploys extremely long sentences dense with nested sub clauses separated by semicolons, parenthetical definitions, and cross references. A single sentence in a complex indemnity clausecansprawlacross400wordsandsevendistinct logical conditions. CUAD style annotation adapted to the eight Indian clause risk categories identified in Section IV provides the target label space for multi label clause classification. In practice, a single clause may simultaneously carry indemnity risk (unlimited liability), data privacy risk (DPDPA obligations are triggered), and limitation of liability risk (the indemnified party’s rights are purportedly capped) making multi label architectures a necessity rather than a design choice. Fine tuned LegalBERT and In LegalBERT models outperform classical SVM baselines on this task, but performance drops sharply for clause typesthatwereneverrepresentedinCUAD.

4.2. Semantic Similarity and Anomaly Detection

One of the most practically useful techniques available to Indian law firms is comparing contract clauses against a curated library of market standard Indian templates the ‘standard’ indemnity clause endorsed by the Bar Council, the data processing

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 09 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072

agreement template circulated by MeitY, or the SEBI mandated disclosure clauses for listed company employment agreements. Embedding based cosine similarity between a given clause and the template library can surface provisions that stray materially from market norms, directing lawyer attention to exactly the clauses most likely to be out of the ordinary. This method is intended to be used in conjunction with classification based methods and not asareplacementforthem:riskisassessedbydeviation fromaknownbaselineandnotbylabelprediction,and it can deal with new clause structures that were not presentinthetrainingdata.

4.3. Dependency Parsing and Obligation Extraction

Extracting obligation triples (party, action, condition) from Indian contract clauses requires dependencyparserscapable ofhandlingdeeplynested conditionalsyntax.Takeaclausesuchas:‘IftheVendor shall notdelivertheServices(ortheServicesshall not be delivered in accordance with the provisions of this Agreement) within the Delivery Period (or the period stipulated for the delivery of the Services) as specified in Schedule 3, and subject to Force Majeure Events as specified in Clause 19, the Client shall be entitled to levy liquidated damages at the rate of 0.5% of the Contract Value per week of delay (maximum up to 10%)’ – that single sentence has 6 different semantic compartments, each of which is relevant to the risk profile of that clause. Standard English dependency parsers trained on news text fail regularly on structures of this complexity. Extracting obligations with adequate coverage of Indian contract syntax requires either domain specific parser training or the LLMbasedextractionapproachdescribedinSectionVI.

5. LLMS FOR INDIAN CONTRACT RISK DETECTION

5.1. Why LLMs Outperform Classical NLP

LLMs bring three capabilities to Indian contract risk detection that classical NLP simply cannot match. First, in context learning lets a model identify risk in clause types it was never explicitly trained on given just two or three annotated examples of Indian arbitration seat/venue clauses, GPT-4 can spot structurallysimilarrisksin newcontractswithoutany retraining. This matters enormously for India, where annotated training data is scarce and new risk types emergewitheachnewstatute.

Second, with long context models like GPT-4 Turbo (having a 128,000 token context window) the entire 60 page vendor agreement can be processed in onego,enablingcrossclauseriskstobeidentified–for

instance the interaction of an indemnity clause in the vendor agreement and a limitation of liability clause –whichencoderonlymodelswithacontextwindowofa maximumof512or4,096tokenscannotevenattempt.

Third,LLMsgivehumanlanguageexplanationsof theriskstheyidentify,soalawyercanseenotonlythat a risk exists but whythatrisk isrenderedproblematic under Indian law, which section of the Indian law is engagedandwhatasaferalternativewouldbe.

5.2. LLM Architecture Selection

Encoder only models such as In LegalBERT are best suited to clause classification and NER tasks where speed and cost at scale are the primary concerns.Decoder only modelsare actuallyreallywell suited for the kind of reasoning heavy work that matters most in legal settings. Think about tasks like explaining why a clause is a red flag, suggesting safer alternativewording,orjustansweringstraightforward questions about where a contract's risks lie these modelshandlethatkindofthingwell.

Table-II: LLMArchitectureRecommendationsbyTask Type

Architecture Models Best For (Indian Contract Context)

Encoder Only In LegalBER T, LegalBER T NER,clause classification,fastbatch processing

Decoder Only GPT-4, LLaMA-3 Riskexplanation,clause Q&A,zero-shot detection

EncoderDecoder FLAN-T5, LegalT5 Clauserewriting, obligationextraction, summarization

RAGaugmented GPT-4+ vectorDB Statute-groundedrisk detection,caselaw retrieval

Encoder decoder models, on the other hand, shine when the job is more structured. Converting a messy obligation clause into a clean, organized format, or rewritingclausesthat'swheretheytendtooutperform the alternatives. Now, the RAG augmented setup (discussed in Section VII B) becomes really useful in a specificsituation:whenyouactuallyneedthemodelto referencerealIndianstatutesorcaselaw.Withoutthat

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 09 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072

grounding, you're kind of flying blind on jurisdiction specific questions. The RAG augmented architecture described in Section VII.B adds the most value when a task requires grounding in specific Indian statutes or caselaw.

5.3. Prompt Engineering for Indian Legal Risk Detection

The most important practical variable in LLM basedIndiancontractanalysisispromptdesign.There arethreepromptstrategiesthatareworthcontrasting. For well known risk types like indemnity and arbitration zero shot prompting which asks the model to identify all clauses creating unlimited indemnity obligations under Indian law and explain why each is risky works well. However it struggles with more subtle risks like the seat/venue distinction or DPDPA compliance gaps where the model simply lacks sufficientIndianlegalcontext.

Few shot prompting significantly increases accuracy on risks unique to India without requiring any fine tuning. It does this by providing two or three annotated Indian contract examples prior to presenting the target clause. Chain of thought prompting reduces hallucination by roughly 30% in controlled evaluations of legal reasoning tasks [6] and is particularly helpful for interactions between multipleclauses.

The model is asked to consider each part of the clause the relevant Indian law the pertinent case law and the conclusion. The most dependable method of passing risk is Retrieval Augmented Generation (RAG) which is based on the prompt and includes passages takenfromtheICADPDPA2023orpertinentSupreme Courtrulings.Thismethodisadvisedforanythingthat isnearactualuse.

6. PROPOSED FRAMEWORK FOR INDIAN CONTRACT RISK DETECTION

6.1. Four-Stage Pipeline

The proposed framework is a four stage pipeline designedaroundthespecificconstraintsofIndianlegal tech deployment: limited annotated data, strict confidentiality requirements, and the need for outputs that are explainable and reviewable by a practicing lawyer.

Stage 1 : Ingestion and Preprocessing: PDF and DOCX contracts are parsed, clause boundaries are detected through a combination of heading pattern heuristics and sentence segmentation adapted for Indianlegalnumberingconventions,definedtermsare

indexed, and language detection is applied to flag any codeswitchedsegmentsforhumanreview.

Stage2 : ClauseClassification:InLegalBERT,fine tuned on the eight category Indian risk taxonomy set out in Table I, assigns each extracted clause to zero or more risk categories. This encoder only layer runs efficiently at batch scale and demands only modest computationalresources.

Stage 3 : LLM Risk Assessment: Stage 2 flagged provisions and any provisions that are indicated by cosine similarity to the standard provision library as semantically unusual are submitted to a Large Language Model (GPT 4 using a proprietary API or LlaMA3finetunedin-house)withaRAG-enhancedfew shotpromptbasedontheapplicableIndianprovisions. The model outputs four items: a risk classification (high / medium / low), a human understandable reasonfortheassignedriskclass,theIndianstatuteor casethroughwhichtheInternetsuggeststheprovision isgoverned,andarecommendedclausevariation.

Stage 4 : Output and Human Review: A contract levelriskdashboardshowseachflaggedclausewithits risk rating the model's explanation and a recommended edit. NoAIflaggedrisk resultsinaction until a lawyer signs off; the system is intended to be a tooltohelplawyersnotreplacethem.

6.2. Indian Legal Knowledge Integration

The RAG layer draws on a vector database containing: the full text of the Indian Contract Act 1872, Companies Act 2013, Arbitration and Conciliation Act 1996, RERA 2016, IT Act 2000, and DPDPA 2023; a curated library of market standard Indiancommercial contract templates;anda corpusof relevant Supreme Court and High Court judgments on contract clause disputes. Retrieval uses dense vector similarity the query clause is embedded and the top k most relevant statutory passages are retrieved followed by a re-ranking step that prioritizes recency and jurisdictional relevance. This architecture cuts hallucination risk substantially by constraining the LLM to reason over verified Indian legal text rather than relying on parametric knowledge that may be coloredbyUSorUKlegalnorms.

6.3. Evaluation Metrics and Deployment Considerations

Evaluatingtheframeworkcallsforfourcategories of metrics: clause level precision, recall, and F1 score

2026, IRJET | Impact Factor value: 8.315 | ISO 9001:2008

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 09 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072

measured against an expert lawyer annotation baseline; risk score calibration, assessed by the correlation between the framework’s risk scores and expert risk judgements; hallucination rate, defined as the proportion of LLM generated statutory citations that can be verified as accurate; and per-contract latency and cost, assessed in light of the economics of Indian law firm practice. Table III sets out reference performance benchmarks for the proposed pipeline stages, drawn from analogous evaluations in the literature and adapted to Indian legal context assumptions.

Table-III: ExpectedPerformanceBenchmarksby PipelineStage

*EstimatedfromanalogousRAGLLMevaluationson USlegalbenchmarks;Indianspecificvalidation pending.

6.4. Comparison with Existing Legal tech Approaches

There are three key ways in which the suggested framework is different from current commercial and academic legal technology systems. First the RAG layersexplicitanchoringinICA1872DPDPA2023and the Arbitration Act 1996 distinguishes it from platforms such as Kira Systems and Luminance which were trained primarily on contract corpora from the United States and the United Kingdom and exhibit documented performance drops when applied to Indian legal drafting conventions. Second the compound nature of Indian contract clauses that concurrently engage multiple statutory provisions is better captured by the multi label risk taxonomy anchored to Indian law (Section IV) which is significantly more granular than the binary risk flags derived from CUAD that the majority of academic

evaluations rely on. Third the frameworks explicit humanintheloopdesignclosesagapincurrentIndian legal tech tools such as SpotDraft and Leegality which offer template management and clause extraction but do not yet integrate statutory grounding with LLM basedriskreasoning.

The approach primary expense is computational: atcurrentAPIpricing Stage 3GPT-4inferenceona 60 page contract is estimated to cost between $1. 50 and $3. 00 per contract. That amount may be too high for high volume low value processing situations like consumer contracts or mass employment agreements but it fits well within the economics of mid tier Indian corporatelegalwork.

7. CONCLUSIONS AND RECOMMENDATIONS

7.1. What We Found

Starting with the first research question after going through Indian contract law carefully, eight types of clauses stand out as genuinely high risk: indemnity, arbitration, limitation of liability, liquidated damages, data privacy, non compete, IP assignment, and force majeure. What makes this interesting is that these aren'tjustriskyinageneralsenseeachonehasitsown specific risk character under Indian law that's actually quitedifferentfromhowsimilarclausesworkinUSor UK contracts. So you can't just borrow Western contractanalysistoolsandexpectthemtoworkhere.

Forthesecondquestion,theshortansweris:nosingle AItechniquedoesthejobwellenoughonitsown.What actually works is combining LegalBERT based clause classification which is great for quickly processing large batches of contracts with RAG augmented few shot LLM prompting, which handles the deeper, context sensitive risk reasoning. Together, they form thehybridpipelinedescribedearlierinSectionVII,and thatlayeredapproachisreallythebestwecandoright nowgivenhowlimitedtheavailabletrainingdatais.

The third research question gets at something that comes up a lot when people talk about using LLMs in specialized domains hallucination, data scarcity, code switching between Hindi and English (and regional languages),andclientdataprivacy.Allfouroftheseare serious,andtheyall pointtothe sameconclusion:you can't just plug in a generic LLM and call it a day. Any deployment in the Indian legal context needs to be built specifically for that context, not borrowed wholesalefromsomewhereelse.

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 09 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072

7.2. What Should Actually Be Done About It

For Indian law firms, the practical takeaway is fairly straightforward adopt the hybrid pipeline, but make sure there's always a human reviewing the output before anything goes to a client. AI assisted review is useful, but it's not ready to work unsupervised. Also, wherever possible, keeping LLMs hosted on premises rather than sending data to third party cloud servers would go a long way toward protecting client confidentiality.

Legal tech startups working in this space have a real opportunity here. Building a high quality annotated benchmark dataset of Indian commercial contracts somethinglikewhatCUADhasdonefortheUSmarket would benefit the entire ecosystem. It's the kind of infrastructure that no single company can justify building alone, but everyone would benefit from, so treatingitasanindustrypublicgoodjustmakessense.

On the policy side, the Bar Council of India and the Ministry of Corporate Affairs really do need to step in andsetsomeclearguidelines.Rightnowthere'salotof ambiguity around who's responsible when AI assisted legal advice turns out to be wrong, and that ambiguity isn't good for anyone lawyers, clients, or the tools themselves.

And finally, for researchers multilingual contract NLP deserves more attention than it's currently getting. Building systems that can actually handle the kind of HindiEnglishandregionallanguagemixingthatshows up in real Indian contracts is a hard problem, and empirically validating any proposed framework on actual Indian contract data (not just synthetic or Westerndatasets)shouldbeaprioritygoingforward.

7.3. Limitations and Future Research

The framework presented here is conceptual; empirical validation on a real Indian commercial contractcorpusisthe essentialnextstepandthemost significant limitation of this analysis. The scope is also restrictedtoEnglishprimarycontracts,leavingentirely unaddressedthesubstantial bodyofregional language agreements Hindi, Marathi, Tamil, and Bengali commercial contracts that are commonplace in domestic procurement and real estate transactions. The LLM landscape is moving fast enough to create a genuineriskofanalyticalobsolescence:thecapabilities gapbetweenGPT4andInLegalBERTthatexiststoday could narrow considerably as smaller, locally fine tuned models continue to improve. Four future research priorities stand out: building India Contract Bench as a fully open, labelled benchmark for Indian commercial contracts with multi-label annotation

across the eight risk categories defined in this paper; developing multilingual LLMs capable of processing HindiEnglishhybridlegaltextwithcrosslingualclause alignment; extending the framework from risk detection to risk remediation through automated, jurisdiction-aware clause rewriting that proposes DPDPA and ICA compliant alternatives; and exploring LLMassistedlegalreasoningtoolscalibratedforIndian commercial courts and arbitral tribunals, where procedural requirements and judicial interpretation differ in important ways from common law equivalents.

7.4. Policy and Regulatory Implications

Beyond the technical and professional dimensions, the findings of this paper raise regulatory questions that Indian policymakers need to tackle proactively. The DPDPA 2023 creates data localization obligations that directly constrain the use of foreign hosted LLM APIs for contracts containing Indian personal data yet the Data Protection Board of India has issued no specific guidance on AI assisted legal services. The Bar Council of India’s existing conduct rules say nothing about AI tools in legal practice, leavingliabilityambiguityunresolvedwhenAIassisted contract review misses a material risk. A coordinated regulatory response BCI guidance on AI assisted legal practice, MeitY standards for legal AI data handling, and MCA guidelines on corporate use of AI contract review would provide the governance foundation that responsible adoption requires. Without such a framework,thefieldislikelytodevelopinapiecemeal way driven by commercial pressure, with liability consequences falling on clients rather than on the technology providers or practitioners who deployed the tools. India’s national AI strategy and the forthcoming Digital India Act offer a genuine opportunity to embed these governance principles before large scale commercial deployment normalizes unregulatedAIlegalpractice.

REFERENCES

[1] World Commerce & Contracting Association, "The CostofPoorContractManagement,"WCCReport, 2019.

[2]D.Hendrycks,C.Burns,A.Chen,andS.Ball,"CUAD: An Expert-Annotated NLP Dataset for Legal ContractReview,"inProc.NeurIPS,2021.

[3] I. Chalkidis et al., "LEGAL-BERT: The Muppets Straight Out of Law School," in Findings of EMNLP,2020,pp.2898–2904.

[4] M. Koreeda and C. D. Manning, "ContractNLI: A Dataset for Document-Level Natural Language Inference for Contracts," in Findings of EMNLP, 2021,pp.1907–1919.

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 09 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072

[5] M. Bommarito and D. M. Katz, "GPT Takes the Bar Exam,"arXiv:2212.14402,2022.

[6] OpenAI, "GPT-4 Technical Report," arXiv:2303.08774,2023.

[7] S. Paul, A. Goyal, and P. Goyal, "Pre-training Transformers on Indian Legal Text," arXiv:2209.06049,2022.

[8] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding," in Proc.NAACL,2019,pp.4171–4186.

[9] T. Brown et al., "Language Models are Few-Shot Learners,"inProc.NeurIPS,2020.

[10] A. Vaswani et al., "Attention is All You Need," in Proc.NeurIPS,2017.

[11] P. Lewis et al., "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks," in Proc. NeurIPS,2020.

[12]I.Chalkidisetal.,"LexGLUE:ABenchmarkDataset for Legal Language Understanding," in Proc. ACL, 2022,pp.4310–4330.

[13] Supreme Court of India, Bharat Aluminium Co. v. KaiserAluminiumTechnicalServiceInc.,(2012)9 SCC552.

[14] Ministry of Law and Justice, Government of India, "The Arbitration and Conciliation Act, 1996," No. 26of1996.

[15] Ministry of Law and Justice, Government of India, "TheIndianContractAct,1872,"No.9of1872.

[16] Ministry of Electronics and Information Technology, Government of India, "The Digital Personal Data Protection Act, 2023," No. 22 of 2023.

[17] Ministry of Housing and Urban Poverty Alleviation,GovernmentofIndia,"TheRealEstate (Regulation and Development) Act, 2016," No. 16 of2016.

[18] E. Lippi et al., "CLAUDETTE: An Automated Detector of Potentially Unfair Clauses in Online Terms of Service," Artificial Intelligence and Law, vol.27,no.2,pp.117–139,2019.

[19] A. Branting et al., "Semi-Supervised Methods for Explainable Legal Prediction," in Proc. 17th Int. Conf.Artif.Intell.Law(ICAIL),2019.

[20] S. Barocas and A. D. Selbst, "Big Data's Disparate Impact,"CaliforniaLawReview,vol.104,no.3,pp. 671–732,2016.

[21]R.Susskind,Tomorrow'sLawyers,3rded.Oxford: OxfordUniversityPress,2023.

[22] C. Manning et al., "The Stanford CoreNLP Natural Language Processing Toolkit," in Proc. ACL SystemDemonstrations,2014,pp.55–60.

[23] N. Holzenberger, A. Blair-Stanek, and B. Van Durme,"ADatasetforStatutoryReasoningin Tax

Law," in Proc. Natural Legal Language Processing Workshop,2020.