L

Lexlegis.ai

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, arguments and authority bundles. MIRA, on the same foundation, selects from a library of 215 production skills, orders them into a plan and checks its own reasoning against a verification layer.

The agreement defines the Service as the platform including Ask, Interact, Draft and MIRA, so the model products are the contracted thing. The vendor states it trained its own model on its legal collection, and its published model layer describes signed checkpoints, guardrails for each tenant and an inference audit trail. The asset outside the models is real: a curated closed collection of Indian laws, circulars and judgments built over decades.

No product without AI is sold separately, and the collection feeds the assistants rather than being sold as a search subscription.

Real architectural controls sit under an accuracy claim that is never measured. Output comes in the IRAC format with a list of the sources relied on, each openable and downloadable, with a summary 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, in the order retrieve, reason, verify. The model layer names output hallucination detection, a meta reasoning gate, guardrails for each tenant and signed model checkpoints.

The vendor states that every factual claim in every output is traced back to its source and that a claim whose trace is broken is flagged, with no exceptions. MIRA is described as checking its own reasoning against a layer built to catch its mistakes. The vendor markets research without hallucination and grounded, hallucination free legal intelligence. No accuracy figure, error rate, hallucination rate, test set, benchmark or evaluation result is published for any of it, so an absolute claim stands without measurement. Its trust center lists items titled A note on Hallucinations and AI Flow for Ask.

The oversight position is contractual and an automated checkpoint is described, but the point at which the agent acts alone is not stated. Clause 8 of the terms, headed AI outputs and professional responsibility, says outputs support legal or compliance work and are not legal advice. The customer must review and validate all outputs before relying on, filing or acting on them, no output creates a lawyer client relationship with the vendor, and the Service does not replace the customer's professional judgment.

Clause 7 adds that the customer will not use outputs to falsely attribute authority or to deceive any court or tribunal. On the product side, the model layer publishes a meta reasoning gate and output hallucination detection through which every input and output passes. The vendor states that failure of any layer opens a ticket rather than a breach, and MIRA is described as verifying its own reasoning before producing output.

Nothing states what MIRA does without a person, when a skill chain completes unattended, whether any step needs human approval, what a user sees when the gate rejects a claim, or whether autonomy is configurable.

A named customer roster with institutional weight, and no outcome measured or attributed. The vendor's homepage names KPMG, Dhruva Advisors, Greaves Cotton, Thermax, K Raheja Corp, Aurtus, MyGate, Artha Energy and the Government of India, and the vendor states more than 100 enterprises. Public milestones are dated and checkable: the platform was shown at NVIDIA GTC 2026 in March 2026, where the company announced a preview beta in the United States, and the skill library is quantified at 215 production skills.

No case study is published, no customer is quoted by name and role, no deployment is described, and no outcome is measured. There is no time saved, volume processed, adoption figure, or accuracy or quality result from any customer. The efficiency claims that appear are unattributed and unquantified: research memos in hours rather than days and authority bundles built automatically. The record shows who is said to use it and nothing about what it did for them.

Substantive commitments are published, most of them in the contract, and privilege is not addressed. Clause 10 of the terms binds each party to hold the other's confidential information in confidence and use it only for the agreement. It names Customer Content as the customer's confidential information and survives termination for five years, or indefinitely for trade secrets. Clause 5 leaves all rights in customer content with the customer and licenses the vendor to process it solely to provide the Service.

The privacy policy says content data is accessible only to the account it belongs to. The trust page publishes matter level isolation on multi tenant SaaS, alongside custom roles, single sign on required on enterprise, and session recording on enterprise. Customer managed keys with hardware security module storage are available on the other modes, and retention can be set to zero on enterprise and the on premises and air gapped modes.

The words privilege and work product appear nowhere in the terms, the privacy policy or the trust pages. Clause 8's statement that no lawyer client relationship arises with the vendor addresses the vendor's own position, not the customer's privilege.

The boundary is drawn in the agreement and backed by an access gate, and no governing professional rule is named. Clause 8 states that outputs support legal or compliance work and are not legal advice, that the customer must review and validate outputs before relying on, filing or acting on them, that no output creates a lawyer client relationship with the vendor, and that professional judgment is not replaced. Clause 7 prohibits using outputs to falsely attribute authority or to deceive any court or tribunal.

Clause 2 restricts paid plans to legal or compliance professionals or people in a role requiring legal research and drafting tools, which is an access control rather than a disclaimer and answers the unauthorized practice question directly. The vendor also states that every skill is designed, tested and validated by senior lawyers. No Bar Council of India rule, provision of the Advocates Act, state bar council guidance or court practice direction is cited anywhere. The Supreme Court of India has publicly taken up AI generated authority in filings.

Technical controls over the model are published in detail, and no governance structure is described. Two items come close to governance. Guardrails can be set for each tenant, which puts configuration of model behavior in the customer's hands, and the vendor states that every skill is designed, tested and validated by senior lawyers before it ships. The trust center 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 governance of the AI itself. No AI governance framework or standard is published, and no ISO/IEC 42001 or equivalent is claimed, although five security and privacy frameworks are claimed precisely. Nobody is named as accountable for model behavior. No evaluation protocol before release, benchmark, red team exercise or published result exists for any model or skill. No bias, fairness or representativeness disclosure appears, for a product whose source material is Indian judicial and tax material.

Published policy covers the ground in the agreement and a structured trust architecture, with one gap. Certifications are stated precisely: ISO/IEC 27001 to the 2022 revision, SOC 2 Type 2 and ISO/IEC 20000-1:2018, with audits by CERT-In empanelled assessors. Encryption is TLS 1.3 in transit and AES 256 at rest, with customer managed keys and hardware security module storage on the modes other than SaaS. Access control covers SAML 2.0, OIDC and SCIM 2.0, single sign on required on enterprise, custom roles, multi factor authentication through the customer's identity provider, and access logs.

Network controls include private cloud isolation, private link, a firewall and IP allow listing. The model layer adds prompt injection detection and signed checkpoints against model poisoning. Customer content is excluded from shared training, retention can be reduced to zero on enterprise and the isolated modes, and data is exportable at any time and returned or destroyed on termination with attestation. A disclosure program commits to acknowledgment within 24 hours and triage within 72. The gap: the privacy policy says a current subprocessor list is published at the trust center, and none appears there.

A warranty, a service level, insurance and a cap are published, and no indemnity runs to the customer. Clause 11 warrants that the Service will be provided with reasonable care and skill. Clause 9 sets a 99.5 percent monthly uptime target on the SaaS mode, with bespoke service levels for enterprise and other modes in the order form. Its trust center lists cyber insurance and professional indemnity insurance. Clause 12 caps liability at fees paid in the twelve months before the claim and mutually excludes indirect, incidental, consequential and punitive damages, while preserving liability for gross negligence, willful misconduct and anything that cannot be limited by law.

Clause 13 allows termination on 30 days uncured material breach, with customer content returned or destroyed at the customer's election. No indemnity runs to the customer: no intellectual property indemnity and no defense obligation. The warranty covers the manner of provision, not the output. Clause 11 states that outputs are not warranted to be accurate, complete, error free or fit for purpose, so the customer bears the loss when an output is wrong.

Enterprise identity integration is specified in detail, and no legal practice system is named. The agreement defines the Service as including web, API and on premise components under an order form, so a programmatic surface is contracted for. The identity layer names SAML 2.0, OIDC and SCIM 2.0 provisioning, with single sign on required on all enterprise deployments and multi factor authentication enforced through the customer's identity provider.

On the isolated modes the platform runs inside the customer's own cloud estate using the customer's identity management and keys. Those are integrations with the customer's security stack, not with the systems a legal team works in. No document management system, case or matter management platform, word processor, email client or electronic signature product is named. No integrations page, connector directory or partner list is published.

No API documentation is published, so the API is established from the agreement and its scope is not. Nothing describes what data moves between the platform and any system a firm already runs.

Tenancy and residency are both published for each mode, with what changes between them. Mode A is multi tenant software as a service hosted in India on ISO 27001 certified infrastructure, with matter level isolation. Mode B is single tenant on the L&T Vyoma sovereign Indian cloud, with private cloud isolation. Mode C runs in the customer's own AWS, Azure or Google Cloud estate, in a region it chooses, with its own identity management and keys.

Mode D is on premises, from a desk unit to a full data center deployment. Mode E is air gapped, and the vendor states there is no inbound traffic at all on modes D and E. The privacy policy states data location for each mode, and residency can be locked to a chosen region on modes B to E. Customer managed keys with hardware security module storage are available on modes C, D and E. Transfers for customers outside India are governed by standard contractual clauses.

The entry SaaS mode is fixed in India with no residency choice, and no subprocessor list is published to test the mode statements against.

A real trust center with a named document inventory, and the reports behind a request whose terms are not stated. Ungated, the vendor's trust center shows six compliance positions: ISO/IEC 27001, ISO/IEC 20000-1:2018, SOC 2 Type 2, SOC 2, GDPR and CCPA. The vendor's own trust page adds DPDP and states the ISO certification as the 2022 revision. Audits are by CERT-In empanelled assessors, the regime recognized by the Indian government.

A risk profile gives a four hour recovery time objective, and a disclosure program publishes response times of 24 hours to acknowledge and 72 to triage. Behind the access flow sit the ISO/IEC 27001 certificate, the SOC 2 Type 2 report, vulnerability assessment and penetration test reports, a security FAQ and a document titled AI Flow for Ask. The portal offers to request access but does not say whether access is instant, needs a nondisclosure click through, or follows a sales conversation.

No auditor, audit period or scope statement is named. The on site trust page has five frameworks, including DPDP and omitting ISO 20000-1, while the trust center has six, including ISO 20000-1 and omitting DPDP.

The architecture is described and nothing in it is named, and the subprocessor list the vendor says it publishes does not appear. The vendor states it built and trained its own model on its legal collection, rather than fine tuning or using retrieval augmented generation. Its model layer describes signed checkpoints, guardrails for each tenant, prompt injection sanitization and an inference audit trail. No model or model family is named.

No third party model, API or inference provider is named, and nothing states whether any is used, so nothing shows whether customer content reaches an outside model on any mode. The named infrastructure is 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. Change notification is committed: the privacy policy says customers are notified of material changes to the subprocessor list, with a right to object for enterprise customers. But the list is stated to be published at the trust center and does not appear there.

Published figures with a stated unit, a real free trial, and the commercial mechanics in the agreement. Two paid tiers carry figures on the vendor's 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 and confirmed at clause 4, requiring no payment instrument, with the vendor reserving the right to limit trial features and usage caps. Clause 2 restricts paid plans to legal or compliance professionals or people in a role requiring legal research and drafting tools.

Fees, billing frequency and currency are set in an order form or product plan, all fees exclude GST, VAT and equivalent taxes, and late payment carries interest at 1.5 percent per month or the statutory rate if lower. Clause 9 publishes a 99.5 percent monthly uptime target on the SaaS mode, and clause 13 covers termination and customer content on exit. Three gaps remain. The four other deployment modes and MIRA carry no published price, no usage quota, seat minimum or overage term is published for either paid tier, and trial usage caps are discretionary.

Practice scope is named specifically, and the underlying collection is not inventoried. Four practice domains are named: direct tax, indirect tax, corporate law and general laws, over a closed database of laws, circulars and judgments that the vendor calls India's largest repository of legal documents. Buyers named include general counsel, law firms and government bodies in the launch material, litigators and transactional and advisory counsel on the practitioner pages, and compliance professionals through the eligibility clause.

Capability is quantified at 215 production skills that can be combined into workflows. Currency of Indian law is addressed once in substance: the vendor distinguishes the Bharatiya Nyaya Sanhita from the Indian Penal Code as a failure of general purpose assistants, a real point in a jurisdiction that recently recodified its criminal law. No court or tribunal is named, no date range is given, nothing distinguishes full text from headnotes, and nothing states which parts of Indian law are thin or out of scope. The United States preview beta announced in March 2026 has no coverage description.