L
Lexlegis.ai
Lexlegis.ai is an Indian legal AI platform operated by Lexlegis Solutions Private Limited of Mumbai, built on a curated closed corpus of Indian laws, circulars and judgments that the company traces to ninety years of legal documentation and knowledge-structuring work by its founding family. Its live product, Legal AID, has three parts. Ask answers legal questions on direct tax, indirect tax, corporate law and general laws in the IRAC format of issue, rule, application and conclusion, returning a list of the sources relied on that a user can open, download and read back with an explanation of why each was relevant. Interact lets a user upload contracts, orders and notices and query across them. Draft produces contracts, submissions, written arguments, oral argument outlines and authority bundles. Alongside these sits MIRA, a skill-configurable agent that reads a brief, selects from a library of 215 production skills, orders them into a plan, reasons through each step and checks its own reasoning against a verification layer before producing output. The company describes a reasoning chain in which every factual claim is traced back to its source and any claim whose trace is broken is flagged. Five deployment modes are published, running from multi-tenant software as a service hosted in India, through a single-tenant sovereign Indian cloud, deployment inside the customer's own AWS, Azure or Google estate with the customer's identity system and keys, to on-premises hardware and fully air-gapped installations, with data residency lockable to a chosen region on every mode above the first. The platform is ISO/IEC 27001 certified, reports SOC 2 Type 2 and ISO/IEC 20000-1, is audited by CERT-In empanelled assessors, and is aligned to India's Digital Personal Data Protection Act 2023 and the GDPR. Its terms of service commit that customer content is not used to train shared models, and its trust materials state that any fine-tuning on customer data happens only under a signed data processing addendum and inside the customer's own deployment boundary. Legal AID is sold on published per-user subscriptions with a seven-day free trial requiring no payment instrument, with eligibility limited to legal and compliance professionals. Named customers include KPMG, Dhruva Advisors, Greaves Cotton, Thermax, K Raheja Corp, Aurtus, MyGate, Artha Energy and the Government of India. The company showed the platform at NVIDIA GTC 2026 and opened a preview beta in the United States in March 2026.
Capability grades
All 15 axes, graded from public sources on the date shown. Hover a grade to see what the letter means on that axis.
AI Centrality
How much of the product is actually AI. Whether the machine learning is the mechanism the buyer is paying for or a feature layered onto conventional software, and whether the vendor is specific about which is which.
Every product sold is model work and the company describes itself as AI-first. Legal AID is three generative capabilities: Ask answers legal questions in the IRAC format with sources returned and explained, Interact reads uploaded contracts, orders and notices and answers across them, and Draft produces contracts, submissions, written arguments, oral argument outlines and authority bundles. MIRA sits on the same foundation as a skill-configurable agent that reads a brief, selects from a library of 215 production skills, orders them into a plan, reasons through each step and checks its own reasoning against a verification layer before producing output. The agreement defines the Service as the platform including Ask, Interact, Draft and MIRA, so the model products are the contracted thing rather than a layer over something else. The vendor states it trained its own model on its corpus and positions that against fine-tuning and retrieval-augmented generation, and its published model layer describes signed checkpoints, per-tenant guardrails and an inference audit trail, which are the controls of an organisation running its own inference rather than reselling someone else's. The non-model asset is real and is described honestly rather than credited: a curated closed corpus of Indian laws, circulars and judgments built over decades of legal documentation work. No separately purchasable non-AI product was located on any surface read, and the corpus is an input to the assistants rather than a search subscription sold on its own. Terms of service v2.0, trust page and product pages read 7 September 2026.
Citation Accuracy and Hallucination Disclosure
Whether the vendor publishes measured accuracy on citations and assertions, grounds output to primary sources, and says plainly what its system does when it does not know. Legal has a documented public record of fabricated citations reaching filed briefs, so an untested claim of accuracy is not evidence.
Real architectural controls sit under an unqualified accuracy claim that is never measured, which is the shape this band's middle grade exists for. The controls are documented and specific. Output is returned in the IRAC format with a list of the sources relied on, each openable and downloadable, with a summary of the source and a statement of why it was pertinent. Retrieval is from a closed curated database of laws, circulars and judgments rather than the open web, and the vendor describes the order as retrieve from the corpus, then reason, then verify. The published model layer names output hallucination detection and a meta reasoning gate, per-tenant guardrails and signed model checkpoints. The reasoning chain is stated in absolute terms: every factual claim in every output is traced back to its source, and if the trace is broken the claim is flagged, with no exceptions. MIRA is described as checking its own reasoning against a layer built specifically to catch its mistakes. What is absent is any measurement of the claim. The vendor markets research without hallucination and grounded, hallucination-free legal intelligence, and publishes no accuracy figure, no error rate, no hallucination rate, no test set, no benchmark and no evaluation result for any of it. An absolute negative claim with no measurement behind it is weaker evidence than a modest measured one, and it is recorded here because a reader should be able to see the gap between the claim and what supports it. The trust centre lists an item titled A note on Hallucinations and a report titled AI Flow for Ask; neither was opened on this channel and both are the named rebuttal route on this row. Trust page, product pages and terms read 7 September 2026.
Autonomy and Oversight Model
What the system decides on its own, what a lawyer must approve, and whether the vendor documents where the review point sits. A tool that drafts under review and a tool that files without one are different products and different risks.
The oversight position is contractual and an automated checkpoint is described, with the threshold at which the agent acts alone left unstated. Clause 8 of the terms is headed AI outputs and professional responsibility and is unusually direct: outputs are tools to support legal or compliance work and not legal advice, the customer remains responsible for reviewing and validating all outputs before relying on them, filing them or acting upon them, no output creates a lawyer client relationship with the vendor, and the customer's professional judgment is not replaced by the Service. Clause 7 adds a conduct obligation most agreements omit, that the customer will not use outputs to falsely attribute authority or to deceive any court or tribunal. On the product side an automated gate is described rather than merely asserted: the model layer publishes a meta reasoning gate and output hallucination detection through which every input and output passes, with the vendor stating that failure of any layer opens a ticket rather than a breach, and MIRA is described as verifying its own reasoning before producing output. Marketing states the position plainly, that the routine work is taken and the judgment stays with the user, end to end under supervision. What is missing is the limb this band names as commonly absent. Nothing states what MIRA does without a person, at what point a skill chain completes unattended, whether any step requires human approval before the next runs, what a user sees when the meta reasoning gate rejects a claim, or whether autonomy is configurable. Terms of service and trust page read in full 7 September 2026.
Operational and Outcome Evidence
Named, dated evidence that the product works in production at real firms or legal departments. Case studies with figures and identified customers count. Unattributed testimonials and launch announcements do not.
A named customer roster with real institutional weight, and no outcome measured or attributed. The roster is published on the homepage and spans professional services, industry and government: KPMG, Dhruva Advisors, Greaves Cotton, Thermax, K Raheja Corp, Aurtus, MyGate, Artha Energy and the Government of India, with the vendor stating more than 100 enterprises. A government body and a Big Four firm as named users are meaningful signals in this market. Public milestones are dated and checkable: the platform was showcased at NVIDIA GTC 2026 in March 2026, where the company announced a preview beta launch in the United States, and the skill library is quantified at 215 production skills. What is absent is everything past the logo. No case study is published, no customer is quoted by name and role, no deployment is described, and no outcome is measured: no time saved, no volume processed, no adoption figure, no accuracy or quality result from any customer. The efficiency claims that do appear are unattributed and unquantified, being research memos in hours rather than days and authority bundles built automatically. A reader can establish who is said to use it and nothing about what it did for them. Homepage, about page, trust page and the March 2026 announcement read 7 September 2026.
Privilege and Confidentiality Posture
How client confidences are handled: attorney client privilege and work product treatment, segregation of one client matter from another, whether client data trains any model, and what the vendor commits to in writing rather than in marketing.
Substantive published commitments across every limb this axis tests except the one that decides the top grade. What is published, and most of it contractually: clause 10 of the terms binds each party to hold the other's confidential information in confidence and use it only for the purposes of the agreement, names Customer Content as the Customer's confidential information, and survives termination for five years or in perpetuity for trade secrets. Clause 5 leaves all rights in customer content with the customer and grants the vendor a licence to process it solely to provide the Service. The privacy policy states that content data is governed by strict access controls and is accessible only to the account it belongs to. The trust page publishes matter level isolation on multi-tenant software as a service, which is a direct answer to the question this axis asks and one most records cannot give, alongside role based access control with custom roles, single sign-on required on all enterprise deployments, least privilege, access log management, and session recording and replay on enterprise. Customer managed keys with hardware security module storage are available on the non-SaaS modes, and retention is configurable to zero on enterprise and on the on-premises and air-gapped modes. What is absent is express privilege and work product treatment. The words privilege and work product appear nowhere on any surface read; clause 8's statement that no lawyer client relationship arises with the vendor addresses the vendor's own position, not the protection of the customer's privilege. That limb is required, so the record holds below the top grade. Terms, privacy policy and trust page read in full 7 September 2026.
UPL and Professional Responsibility Posture
Whether the vendor is clear that it supplies a tool rather than legal advice, who its audience is, and how it addresses unauthorized practice of law, competence and supervision duties, and jurisdiction limits. ABA Formal Opinion 512 is the reference point.
The boundary is drawn in the agreement, backed by an access gate, and no governing professional instrument is named. What is published: clause 8 states that outputs are tools to support legal or compliance work and not legal advice, that the customer remains responsible for reviewing and validating outputs before relying on, filing or acting on them, that no output constitutes a lawyer client relationship with the vendor, and that professional judgment is not replaced. Clause 7 carries a prohibition rarely seen in a legal AI agreement and directly on this axis: the customer will not use outputs to falsely attribute authority or to deceive any court or tribunal. Clause 2 gates the product itself, restricting paid plan eligibility to legal or compliance professionals or those employed in a role requiring legal research and drafting tools, which is an access control rather than a disclaimer and is a real answer to the unauthorised practice question. The vendor also states that every skill is designed, tested and validated by senior lawyers. What is absent is any named authority. No rule of the Bar Council of India, no provision of the Advocates Act, no state bar council guidance and no court practice direction is cited anywhere, in a jurisdiction where the Supreme Court has publicly taken up the consequences of AI-generated authority in filings. Terms of service and about page read in full 7 September 2026.
AI Governance and Bias Disclosure
Published governance over model behaviour: who owns it inside the vendor, what is tested before release, and what is disclosed about disparate output across matter types, parties, or populations.
Technical controls over the model are published in detail and no governance apparatus is described. What is genuinely governance-adjacent and credited here rather than on the security row: per-tenant custom guardrails, which put configuration of model behaviour in the customer's hands, and the vendor's statement that every skill in the library is designed, tested and validated by senior lawyers before it ships, which is a human review gate on the product's content. The trust centre also lists a code of ethics, anti-bribery and anti-modern-slavery items, and a product security entry titled A note on Hallucinations. What is missing is everything that would make this governance of the AI rather than security around it. No AI governance framework or standard is published or claimed and no ISO/IEC 42001 or equivalent is asserted, in contrast to the five security and privacy frameworks that are asserted with precision. Nobody is named as accountable for model behaviour. No pre-release evaluation protocol, benchmark, red-team exercise or published result exists for any model or skill. And no bias, fairness or representativeness disclosure appears at any level, which is worth naming for a product whose corpus is Indian judicial and tax material and whose outputs are used in a jurisdiction where courts have begun scrutinising AI-derived authority. Trust page, about page and trust centre index read 7 September 2026.
AI Safety and Data Stewardship
Retention, deletion, access control, and what happens to prompts and documents after they are processed. Whether the vendor states its subprocessors and its incident practice, or leaves the buyer to assume.
Published policy across every limb this axis asks about, in the agreement and in a structured trust architecture, with one gap named. Certifications are asserted with precision: ISO/IEC 27001 certification stated as the 2022 revision, SOC 2 Type 2, ISO/IEC 20000-1:2018, plus GDPR, CCPA and India's DPDP alignment, with CERT-In empanelled audits and annual audit stated where applicable. Encryption is specified as TLS 1.3 in transit and AES 256 at rest, with customer managed keys and hardware security module key storage available on the non-SaaS modes. Access control is detailed: SAML 2.0, OIDC and SCIM 2.0 provisioning, single sign-on required on enterprise, role based access control with custom roles, multi-factor enforced through the customer's identity provider, least privilege and access log management. Network controls include virtual private cloud isolation, private link, web application firewall, denial of service protection, IP allow listing and geographic restriction, with no inbound traffic at all on the on-premises and air-gapped modes. On the model itself, prompt injection detection and sanitisation and model poisoning defence via signed checkpoints. Training, retention and deletion are all committed: no customer content in shared training corpora, retention configurable and reducible to zero on enterprise and the isolated modes, data erasure documented, and data exportable at any time with return or destruction at customer election on termination, with attestation. Incident practice exists on both sides, a data breach notification item in the trust centre and a coordinated disclosure programme committing acknowledgement within 24 hours, triage within 72 and remediation to severity-based service levels, with safe harbour for good faith researchers. The gap, named because every partial names its limitation: the privacy policy states that a current sub-processor list is published at the trust centre, and no such list appears on that page. Privacy policy, terms, trust page and trust centre read 7 September 2026.
AI Liability and Recourse
What the vendor stands behind contractually when its output is wrong. Indemnities, caps, carve outs, insurance, and whether any of it is published or only reachable through a negotiated agreement.
More of what this axis asks for than most records carry, with one absence that keeps it off the top. What is published: an affirmative warranty, clause 11 warranting that the Service will be provided with reasonable care and skill, which is a positive obligation rather than the pure as-is disclaimer that is the norm in this corpus. A service level with a number, clause 9 carrying a 99.5 per cent monthly uptime target on the SaaS mode and bespoke service levels on enterprise and non-SaaS modes set out in the order form, with a named customer success manager for enterprise. Insurance disclosed rather than absent, the trust centre listing both cyber insurance and professional indemnity insurance in its legal section. A liability cap at clause 12 set at fees paid in the twelve months preceding the claim, mutual exclusion of indirect, incidental, consequential and punitive damages, and express carve-outs preserving liability for gross negligence, wilful misconduct and anything that cannot be limited by law. Termination at clause 13 on 30 days uncured material breach, with customer content returned or destroyed at the customer's election and, per the trust page, with attestation. What is absent is indemnity. No indemnity of any kind runs to the customer: no intellectual property indemnity, no defence obligation, no remedy ladder, nothing. And the warranty that does exist covers the manner of provision rather than the product's output, clause 11 stating expressly that outputs are not warranted to be accurate, complete, error free or fit for purpose, so on the question this axis exists to ask the customer still bears the loss when an output is wrong. Terms of service and trust centre read in full 7 September 2026.
Practice Systems Integration Depth
How deeply the product reaches into the systems legal work already lives in: document management such as iManage and NetDocuments, Word and Outlook, contract lifecycle management, matter management, e-billing, and court filing systems.
Enterprise identity integration is specified in unusual detail and no legal practice system is named. What is established: the agreement defines the Service as including web, API and on-premise components made available under an order form, so a programmatic surface is contracted for rather than merely advertised; and the identity layer is published with named standards, being SAML 2.0, OIDC and SCIM 2.0 provisioning, with single sign-on required on all enterprise deployments and multi-factor enforced through the customer's own identity provider policies. On the isolated modes the platform runs inside the customer's own cloud estate using the customer's identity and access management and keys. Those are real integrations and they are integrations with the customer's security stack, not with the systems a legal team works in. What is absent is the substance of this axis. No document management system, case or matter management platform, word processor, email client or e-signature product is named on any surface read. No integrations page, connector directory or partner list was located in the site navigation. No API or developer documentation was located, so the API's existence is established from the agreement and its scope and depth are not. Nothing describes what data moves in which direction between the platform and any system a firm already runs. Terms of service, trust page, privacy policy and site navigation checked 7 September 2026.
Deployment Model and Data Residency
Where the software runs and where the data sits. Multi tenant cloud, single tenant, private deployment, on premises, and whether region of residence is a published option or an enterprise conversation.
Both limbs are published in full, per mode, and the record states what changes between tiers, which is the question this axis asks and which most records cannot answer. Five deployment modes are named and described. Mode A is multi-tenant software as a service hosted in India on ISO 27001 certified infrastructure, with matter level isolation stated for the multi-tenant case. Mode B is single tenant on the L&T Vyoma sovereign Indian cloud, with virtual private cloud isolation. Mode C runs inside the customer's own AWS, Azure or Google Cloud estate in a region of the customer's choosing, with the customer's own identity and access management and keys, virtual private cloud isolation and private link or private service connect. Mode D is on premises, including an NVIDIA DGX Spark unit on a desk or a full deployment in the customer's data centre. Mode E is air gapped, with the vendor stating there is no inbound traffic at all on modes D and E. Residency is addressed as architecture rather than a setting: data location is stated per mode in the privacy policy, and residency can be locked to a chosen region on modes B, C, D and E. Customer managed keys with hardware security module storage are available on modes C, D and E. Cross-border transfers for customers outside India are stated to be governed by standard contractual clauses and equivalent mechanisms. What a buyer cannot establish is a residency choice on the entry SaaS mode, which is fixed in India, and no sub-processor list was located to test the mode statements against. Privacy policy, trust page and homepage read in full 7 September 2026.
Security Certifications and Trust Center
Independent attestation a buyer can pull without a sales call: SOC 2, ISO 27001, penetration test summaries, a trust center with current reports and named scope rather than a badge image.
A real trust centre with a named artefact inventory, and the reports behind a request whose terms the portal does not state. What is ungated: six compliance positions in the trust centre, being ISO/IEC 27001, ISO/IEC 20000-1:2018, SOC 2 Type 2, SOC 2, GDPR and CCPA, with the vendor's own trust page adding DPDP and stating the ISO certification as the 2022 revision; CERT-In empanelled audits, which is the Indian government-recognised assessor regime and a meaningful local signal; a risk profile giving data access level, impact level and a four hour recovery time objective; a control inventory spanning product security, data security, data privacy, access control and environmental, social and governance items; and a coordinated disclosure programme with published response times of 24 hours to acknowledge and 72 to triage. What sits behind the access flow: the ISO/IEC 27001 certificate, the SOC 2 Type 2 report, a vulnerability assessment report, a penetration test report, a security compliance and deployment FAQ, and a document titled AI Flow for Ask. The portal offers to start a security review, to view and download sensitive information and to request access, and does not state whether access is instant on an email address, a self-service non-disclosure click-through, or approval after a sales conversation, so the lower tier is taken under the standing rule and that is why this is not the top grade, alongside the absence of any named auditor, audit period or scope statement. One discrepancy is recorded: the on-site trust page lists five frameworks including DPDP and omitting ISO 20000-1, while the trust centre lists six including ISO 20000-1 and omitting DPDP. No request was submitted. Trust page and trust centre read 7 September 2026.
Model Supply Chain Disclosure
Which models sit underneath, whose they are, where they run, and whether the vendor commits to telling customers when that changes. A legal buyer inherits every dependency it cannot see.
The architecture is described and nothing in it is named, and the one list the vendor says it publishes could not be found where it says it is. What is disclosed: the vendor states it built and trained its own model on its legal corpus and positions that against fine-tuning and retrieval-augmented generation, and its published model layer describes signed model checkpoints, per-tenant guardrails, prompt injection sanitisation and an inference audit trail, all of which are the controls of an operator running its own inference. So a reader learns that the model is claimed to be the vendor's own, which is a real answer to part of the question. What is not disclosed is any name or any boundary. No model or model family is named. No third-party model, API or inference provider is named, and nothing states whether any is used at all, so a reader cannot tell whether customer content reaches an outside model on any mode. The infrastructure that is named is either the vendor's own hosting in India, the L&T Vyoma sovereign cloud, or the customer's own AWS, Azure or Google estate, and a customer's own cloud is not a supply chain disclosure by the vendor. Change notification is committed and is the strongest element here: the privacy policy states that customers are notified of material changes to the sub-processor list, with a right to object for enterprise customers. But the list itself is stated to be published at the trust centre and does not appear on that page, so the commitment attaches to a document that could not be located. Privacy policy, terms, trust page and trust centre read 7 September 2026.
Commercial Transparency
Whether a buyer can learn what this costs without entering a sales process: published rates, the unit being charged, what sits behind an enterprise tier, and what implementation adds.
Published figures with a stated unit, a genuine free trial, and the commercial mechanics in the agreement rather than withheld. Two paid tiers carry figures on the vendor's own site: Professional from 9,000 rupees per user per month and Enterprise from 17,250. A seven day free trial of Legal AID is published on the site and confirmed at clause 4 of the terms, expressly requiring no payment instrument to start, with the vendor reserving the right to limit trial features and usage caps. Eligibility is stated rather than left implicit, clause 2 restricting paid plans to legal or compliance professionals or those in a role requiring legal research and drafting tools. The mechanics sit in the agreement: subscription basis, fees and billing frequency and currency set out in an order form or product plan, all fees exclusive of GST, VAT and equivalent taxes added at prevailing rates, and late payment attracting interest at 1.5 per cent per month or the statutory rate if lower. Clause 9 publishes a service level with a figure, a 99.5 per cent monthly uptime target on the SaaS mode. Clause 13 states the termination position and what happens to customer content on exit. Three gaps are named rather than smoothed: the four non-SaaS deployment modes and MIRA carry no published price and route to an order form; no usage quota, seat minimum or overage term is published for either paid tier; and trial usage caps are expressly discretionary. Homepage, pricing statements and terms of service read 7 September 2026.
Firm and Practice Coverage
Who the product is actually built for. AmLaw, midlaw, small firm and solo, in house departments, government and courts, and which practice areas are supported rather than merely claimed.
Practice scope is named with specificity and the corpus behind it is not inventoried. What is published: four practice domains addressed by name, being direct tax, indirect tax, corporate law and general laws, over a corpus described as a closed database of laws, circulars and judgments and characterised as India's largest repository of legal documents. Buyer types are named across the range this index cares about: general counsel, law firms and government bodies in the vendor's launch material, litigators and transactional and advisory counsel on the practitioner pages, and compliance professionals through the eligibility clause. Depth of capability is quantified where the vendor can quantify it, at 215 production skills composable into workflows. Currency of Indian law is addressed at least once in substance, the vendor distinguishing the Bharatiya Nyaya Sanhita from the Indian Penal Code as a failure mode of general-purpose assistants, which is a real coverage claim in a jurisdiction that has recently recodified its criminal law. What is absent is any inventory or boundary. No court or tribunal is named, no date range or historical start point is given, no depth statement distinguishes full text from headnote, and nothing states which parts of Indian law are thinly covered or out of scope. The United States is entered as a preview beta announced in March 2026 with no coverage description at all, so a reader cannot establish what the American corpus contains. Product pages, homepage, launch material and terms read 7 September 2026.
From ₹9,000 per user per month
- Legal AID is sold on two published per-user subscriptions: Professional from ₹9,000 per user per month and Enterprise from ₹17,250. Both give you Ask, Interact and Draft. You can try it free for seven days without giving a card, and you can only buy a paid plan if you are a legal or compliance professional or work in a role that needs legal research and drafting tools. Prices exclude GST and equivalent taxes, and late payment carries interest at 1.5 per cent a month. Enterprise adds a named customer success manager and a service level agreed in your order form; the standard cloud service carries a published 99.5 per cent monthly uptime target. MIRA and the four non-standard deployment options, running from a single-tenant sovereign Indian cloud through your own cloud account to on-premises and air-gapped installations, are all quoted individually rather than listed at a price. Either side can end the agreement for a serious breach that is not fixed within 30 days, and your content is returned or destroyed on exit as you choose.
Two published per-user subscription tiers with figures, quoted in Indian rupees on the vendor's own site: Professional from 9,000 rupees per user per month and Enterprise from 17,250. The unit of charge is per user per month and is stated. A seven day free trial of Legal AID is published and carried into the agreement at clause 4, expressly requiring no payment instrument, with the vendor reserving discretion to limit trial features and usage caps. Eligibility is a published condition rather than an assumption: clause 2 restricts paid plan eligibility to legal or compliance professionals or persons employed in a role requiring legal research and drafting tools. The commercial mechanics sit in the terms of service v2.0 rather than being withheld: subscription basis with fees, billing frequency, currency and payment terms set out in the applicable order form or product plan; all fees exclusive of GST, VAT and equivalent taxes added at prevailing rates; late payment attracting interest at 1.5 per cent per month or the statutory rate, whichever is lower; termination on 30 days uncured material breach with customer content returned or destroyed at the customer's election. A service level accompanies the price rather than being sold separately, clause 9 publishing a 99.5 per cent monthly uptime target on the standard SaaS mode, bespoke service levels on enterprise and non-SaaS modes set in the order form, and a named customer success manager for enterprise. The figure recorded here is the Indian rupee amount as published, not a converted United States dollar figure, following the corpus convention for non-USD price cards. Three gaps are named: MIRA carries no published price; the four deployment modes above standard SaaS, being the sovereign Indian cloud, customer-cloud, on-premises and air-gapped options, are all quoted by order form with no indicative figure or uplift published; and no usage quota, seat minimum or overage term is published for either paid tier. Homepage pricing statements and terms of service read 7 September 2026.
Legal Signals
What each signal meansA signal records what public sources say on the date shown. It is not a grade and it is not a recommendation. Where a signal reads Not addressed, it means the index did not locate the material in public sources on that date, which is a statement about disclosure rather than about the product.
Client Data in Training
Can material a lawyer puts into this product be used to train a model?
The published terms prohibit training on customer content. Not a policy page, the agreement.
The prohibition is in the agreement, and the vendor is the only record located so far that states the qualifier and then explains what it leaves open. Clause 5 of the terms of service v2.0, effective 1 January 2026 and last updated 19 April 2026, provides that the customer retains all rights in documents, queries and other content submitted, grants only a limited licence to process that content solely to provide the Service, and states that Customer Content is not used to train shared models. The privacy policy repeats it twice, saying content data is not used to train shared models and that the vendor does not use personal data to train shared models. The trust page states that these commitments are contractual rather than aspirational. The qualifier is the word shared, and unlike the other records carrying that shape this vendor addresses what sits outside it rather than leaving a reader to infer: fine tuning on customer data, if offered, happens only under a signed data processing addendum and only inside the customer's own deployment boundary. That is a bounded carve-out with two conditions attached, and it is recorded here because it is the difference between a qualifier that hides an exception and one that defines it. Two structural facts reinforce the commitment rather than being counted twice: content retention is customer-configurable and reducible to zero on enterprise and the isolated modes, and on the on-premises and air-gapped modes the content never leaves the customer's estate at all. Terms, privacy policy and trust page read in full 7 September 2026.
Prompt and Output Retention
How long does the product keep what a lawyer typed, and can that be set to zero?
The customer sets the retention window and no retention is an available setting.
The customer sets the retention window and zero is reachable, with the tier at which it becomes reachable stated rather than hidden. The privacy policy separates three buckets and gives each its own rule. Content data, defined as documents uploaded, queries submitted and outputs generated, has its retention set by the customer, defaulting to the active life of the account plus ninety days for export. Usage data, being query logs, session identifiers, device metadata and telemetry, is retained for up to 24 months by default and is configurable on enterprise. Account data is held for the relationship plus seven years, which the vendor attributes to Companies Act and Income Tax Act record-keeping rules rather than leaving it unexplained. The trust page states the floor: retention windows are configurable to zero on enterprise and on the on-premises and air-gapped modes. So zero exists, and the honest limitation is that it is an enterprise and isolated-mode capability rather than something available on the entry subscription, where the default window governs. Exit is addressed alongside retention: all customer data is exportable at any time in standard formats, and on termination it is returned or destroyed at the customer's election with attestation, which clause 13 of the terms carries into the agreement. A data erasure control is listed separately in the trust centre. Privacy policy, terms and trust page read in full 7 September 2026.
Ethical Walls and Matter Segregation
Does retrieval respect the firm’s ethical walls, or can the model read across them?
The product maintains its own permission model, documented, requiring the firm to keep it aligned.
The vendor maintains its own segregation model, documents it, and states it at matter level rather than only at account level, which is rare on this signal. The data layer published on the trust page lists matter level isolation on multi-tenant software as a service as one of five data controls, alongside encryption, customer managed keys and hardware security module key storage. The privacy policy supports it from the other side, stating that content data is governed by strict access controls and is accessible only to the account it belongs to. The identity layer adds the machinery a firm would need to operate a wall: role based access control with custom roles, single sign-on required on all enterprise deployments, multi-factor enforced through the customer's own identity provider policies, SCIM 2.0 provisioning, and session recording and replay on enterprise. The isolated deployment modes take the question further by removing multi-tenancy altogether. What is not published is how the isolation is administered by the customer. Nothing states whether a matter boundary is created automatically or configured, whether a user can be screened from a specific matter, how a wall is applied to an existing matter, or what an administrator sees. Nothing inherits from a document or case management system, because none is integrated. So the model is the vendor's own, documented in outline, and a firm would have to keep it aligned with the walls it maintains elsewhere. Trust page and privacy policy read in full 7 September 2026.
Third Party Request and Subpoena Notice
If someone subpoenas the vendor for a firm’s data, does the firm hear about it first?
No located term or policy addresses third party requests for customer data.
No located public material addresses what happens when a third party demands customer data, and the omission stands out against how complete the rest of this estate is. The surfaces that would ordinarily carry it were read in full and none does. The terms of service run to fifteen clauses including a mutual confidentiality clause at clause 10, and that clause contains no compelled-disclosure carve-out and no notice obligation, which is unusual because the carve-out is standard drafting in exactly that position. The privacy policy names legal obligations as a lawful basis for processing under DPDP and GDPR and describes sharing with sub-processors, but has no law enforcement or government request section, no statement of what the vendor does on receipt of an order, and no notice commitment. The trust centre's open page lists a data privacy group containing cookies, data breach notifications and items, and no compelled disclosure item. No transparency report or law enforcement guidelines page exists on any surface. So nothing is established either way: neither a commitment to tell the customer nor a statement that disclosure would occur without notice. Two adjacent facts are recorded because they bear on the practical exposure without answering the question: on the on-premises and air-gapped modes the vendor does not hold the content, and a data processing addendum is named in clause 15 as forming part of the agreement but is not published. Terms, privacy policy and trust surfaces checked 7 September 2026.
Primary Law Corpus Provenance
Where does the law in this product come from, and does the vendor have the right to use it?
Sources are identified without stating the licence or rights basis.
The corpus is identified by type and the rights basis behind it is asserted only as to the vendor's own holding, with no cadence published. What is identified: a closed database of laws, circulars and judgments, described by the vendor as India's largest repository of legal documents, addressed to four named practice domains being direct tax, indirect tax, corporate law and general laws. The vendor distinguishes its corpus from open-web retrieval as a design position, and traces the collection to decades of legal documentation and knowledge-structuring work by named individuals in its founding lineage, which is more origin story than most records offer. Clause 6 of the terms states that the corpus, along with the software, models, skill library and metadata schemas, is owned by the vendor or licensed to it. That is an assertion about the vendor's own title and it does not state the basis on which the underlying legal materials are used: nothing identifies the reporters, publishers or official sources the judgments and circulars come from, and nothing states whether Indian primary law is treated as government material outside copyright or licensed from a publisher. Two further gaps are named. No update cadence is published anywhere, so a reader cannot establish how quickly a new judgment, circular or amendment reaches the corpus, which matters in a jurisdiction that has recently recodified core statutes. And no coverage inventory exists to test the size claim against. Terms, product pages and about page read 7 September 2026.
Good Law Verification
Does the product tell you when the authority it just cited has been overruled?
No located public material addresses whether authority is checked for subsequent history.
No located public material addresses whether authority is checked for subsequent history, and the verification the product does perform answers a different question. What the product verifies is existence and support: every response returns a list of the sources relied on, each openable and downloadable with a summary and a statement of why it was pertinent, and the reasoning chain traces every factual claim back to its source and flags a claim whose trace is broken. That is credited on the Citation Accuracy axis and is not counted again here. It establishes that a cited judgment is real and says what the vendor claims it says. It does not establish that the judgment still stands. Nothing published states whether a decision has been overruled, reversed on appeal, stayed, distinguished, declared per incuriam or superseded by statute, and no citator is licensed, no treatment or noting-up signal is computed or surfaced, and no currency flag attaches to any authority in an output. The vendor does show awareness of the underlying problem in a related form, contrasting general-purpose assistants that treat the Bharatiya Nyaya Sanhita as though it were the Indian Penal Code, which is a statutory currency point rather than a judicial treatment one, and no mechanism is described for either. In a jurisdiction where a judgment's standing turns on later benches, the absence is material and is recorded as an absence established on readable surfaces rather than a retrieval limit. Product pages, trust page and terms checked 7 September 2026.
Refusal and Uncertainty Behaviour
What does the product do when the answer is not in the corpus?
The vendor describes refusal or abstention behaviour in public materials.
What the system does when it cannot support a claim is stated, in terms, and without a demonstration behind it. The commitment is explicit: every factual claim in every output is traced back to its source, and if the trace is broken the claim is flagged, with the vendor adding no exceptions. That is a described behaviour on failure rather than a description of grounding, which is why it is graded here while the grounding architecture itself, being corpus-first retrieval, IRAC output and the sources list, is credited on the Citation Accuracy axis. Two further mechanisms sit alongside it. The published model layer names output hallucination detection and a meta reasoning gate through which every input and output passes, with the vendor stating that failure of any layer opens a ticket rather than a breach, so a detection event is treated as an operational signal that someone sees. And MIRA is described as verifying its own reasoning against a layer built specifically to catch its mistakes before it produces output. What is missing is the demonstrable half. No evaluation, benchmark, worked example or published result shows the flagging behaviour operating, no rate is given for how often a trace breaks, and nothing states what the user actually sees when a claim is flagged, whether the claim is withheld, marked, or returned with a caveat. A trust centre item titled A note on Hallucinations was not opened on this channel and is the named rebuttal route. About page, trust page and product pages read 7 September 2026.
Fabricated Citation Record
Does a public court record exist involving output from this product?
No court order, opinion or disciplinary record naming this product has been located as of the date shown. This is a statement about the public record, not a finding about the product.
No court order, opinion or disciplinary record naming Lexlegis.ai or Lexlegis Solutions Private Limited was located as of 7 September 2026. Searches were run on the company and product names against Indian and international sanction and hallucination language and against the AI Hallucination Cases database maintained by Damien Charlotin. The Indian decisions located name no tool. This is a statement about the public record and not a finding about the product. The environment is recorded because it bears on how the statement should be read rather than on the vendor: Indian courts are actively seized of the question, the Supreme Court having taken cognisance in a February 2026 order of a trial court that relied on non-existent AI-generated judgments, describing it as a matter of institutional concern bearing directly on the integrity of the adjudicatory process and indicating that such a decision would be misconduct with legal consequences to follow, and a separate bench having raised petitions drafted with AI citing non-existent authority. A vendor selling into that market on an express hallucination-free claim is exposed to the question in a direct way, and nothing on the public record to date names it.
Bar Guidance Alignment
Has the vendor engaged in public with the ethics opinions its buyers are bound by?
Public materials refer to professional responsibility in general terms without naming guidance.
Professional responsibility is engaged squarely and in the agreement, and no bar or court instrument is named. The engagement is more concrete than most records at this value. Clause 8 is headed AI outputs and professional responsibility and states that outputs are tools to support legal or compliance work rather than legal advice, that the customer remains responsible for reviewing and validating all outputs before relying on them, filing them or acting upon them, that no output creates a lawyer client relationship with the vendor, and that professional judgment is not replaced. Clause 7 goes further than the usual disclaimer and imposes a positive prohibition aimed at the courtroom, that the customer will not use outputs to falsely attribute authority or to deceive any court or tribunal, which addresses the precise conduct that professional guidance on AI has been written to prevent. Clause 2 restricts paid plan eligibility to legal or compliance professionals or those employed in a role requiring legal research and drafting tools, so the obligation attaches to a gated population. What is absent is any named authority. No rule of the Bar Council of India, no provision of the Advocates Act 1961, no state bar council circular and no court practice direction or standing order on AI-assisted filings is cited, mapped or linked anywhere on the surfaces read. That gap is sharper here than in most markets, because Indian courts have publicly begun addressing AI-generated authority and a vendor engaging the conduct question this directly might be expected to name the source of the duty. Terms of service read in full 7 September 2026.
Billing and Fee Posture
Does the vendor address what happens to the bill when the work takes an hour instead of six?
Public materials claim time savings without addressing billing or disclosure, and the product sits inside a fee relationship between a lawyer and a client where those savings would change the bill.
Time savings are marketed to a buyer who bills clients, and nothing addresses the client's side of the bill. The product sits inside a fee relationship for at least part of its buyer set: law firms and litigators are named buyers alongside general counsel and government bodies, and the work being compressed is legal research and drafting, which a firm bills. The savings claims are explicit and aimed at that work, promising research memos in hours rather than days, hearing preparation without the all-nighter, authority bundles built automatically, and answers in seconds. Against that, no published material addresses billing, fee treatment or client disclosure. Nothing tells a firm how to treat research hours that collapse, no per-matter record of AI-assisted work is offered as a basis for a fee narrative, and no guidance on disclosing AI assistance to a client was located on any surface. Clause 3 of the terms governs only what the customer pays the vendor, covering subscription basis, order forms, tax exclusivity and late payment interest, and never reaches the customer's own invoicing. The omission is worth naming because the vendor engages the lawyer's other professional obligations at length, gating eligibility to professionals and prohibiting the use of outputs to deceive a court, so this is not a general silence about the duties of the buyer. Practitioner pages, homepage and terms of service checked 7 September 2026.
Outside Counsel Guideline Readiness
Can a firm get this vendor through a client’s AI clause without a bespoke negotiation?
The material exists behind a sales conversation or an executed agreement.
The artefacts a firm would forward are named and governed, and the one that is said to be published could not be found. What is committed: clause 7 of the privacy policy states that the vendor relies on a small set of sub-processors, that a current list is published at the trust centre, that customers are notified of material changes to that list, and that enterprise customers have a right to object. A notification commitment with an objection right is stronger than most records carry. Clause 15 of the terms names a Data Processing Addendum and a Security Schedule as forming part of the entire agreement, so both exist and are contractually incorporated, and the trust centre carries a documents flow through which they and the certification reports can be requested. What is not available is the disclosure itself. The sub-processor list does not appear on the open view of the trust centre the privacy policy names as its location, so the commitment to notify changes attaches to a document a prospective buyer cannot read. No model provider is named anywhere, so a firm cannot tell its client whose model, if any beyond the vendor's own, processes matter content. The Data Processing Addendum and Security Schedule are named and not published. That places the forwardable material behind a request rather than in a firm's hands, which is what this value records. No access request was submitted. Privacy policy, terms and trust centre read 7 September 2026.
Court Disclosure Support
If a judge’s standing order requires an AI disclosure, can the product produce one?
Some elements of the record are available, short of a document level export.
Real elements of a record exist and none is framed or offered as something a lawyer could produce. The strongest element is specific to AI work rather than to data access, which is why it is graded here: the published model layer states an inference audit trail in which every call is logged and attributable. That is a record of what the model did and who invoked it, not a log of who opened a file, and it is materially different from the general audit logging controls that appear on most trust centres. Around it sit further elements. Every output carries the sources it relied on, openable and downloadable with a statement of why each was pertinent, and the reasoning chain flags any claim whose trace to a source is broken; those are credited on the Citation Accuracy and uncertainty rows respectively and are named here only because a reader assembling a record would reach for them. Clause 7 of the terms engages the courtroom directly by prohibiting use of outputs to falsely attribute authority or deceive a tribunal, which shows the scenario was considered. What is absent is the disclosure half. Nothing states that the inference audit trail is exportable, available to the customer at all rather than to the vendor, retained for any period, or presentable in a form a court would accept. No certification, template, declaration or guidance addressing a practice direction on AI-assisted filings is published, in a jurisdiction where courts have begun asking. Trust page, terms and product pages read 7 September 2026.