Hebbia vs Legora: how they compare in 2026
Hebbia's Matrix and Legora's Tabular Review are the same idea, a grid that runs AI down a set of documents with each answer linked to its source, and both are sold to law firms for diligence and review. The grid does not separate them on the totals: each sits in the top two bands on ten of fifteen axes. It does separate them on the two axes a legal buyer weighs hardest, in opposite directions. On liability Hebbia holds an A to Legora's B, because Hebbia's published agreement sets out an intellectual property indemnity, a twelve month fee cap that includes it, and 99.9 percent availability backed by service credits. On citation accuracy Legora holds a B to Hebbia's C, because Legora has published how its rebuilt research layer checks authority, with an AI native citator in limited beta audited by former publisher attorney editors, while Hebbia publishes no method or measurement and its agreement states that the platform is not warranted accurate or complete. Hebbia also names every model provider behind its product, and Legora names none.
At a glance
All 15 axes, side by side
The same grid applied to every vendor in the index, graded from public sources. 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.
The artificial intelligence is the product. Matrix is described by the vendor as a platform for building AI agents that complete end-to-end tasks rather than chatting, decomposing a question into structured steps across thousands of uploaded documents and returning a grid where each row is a document and each column a question, with sourcing shown per cell; Max is the agent product sold alongside it. The subprocessor register makes the architecture concrete, naming four large language model providers that process user prompts and files and five separate services that host parts of the inference infrastructure. There is no underlying document system, workflow tool or system of record that survives the removal of the models: strip them out and there is nothing left to sell. Legal solution page, security page, Matrix 2.0 announcement and DPA Annex III read 7 September 2026.
The models are the product. The workspace is AI native rather than a document system with a model attached, and every surface the vendor sells, review, drafting, research and workflows, is a generative capability. Remove the models and nothing remains to sell.
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.
Accuracy is asserted and never measured, and the vendor's own agreement contradicts the marketing claim. The security page states that Hebbia is creating the standard for trustworthy large language models that never hallucinate and give correct, verifiable answers. The Main Services Agreement, section 7.1, states in capitals that Hebbia does not warrant that the platform is accurate or complete and that it is provided as is. No accuracy figure, no test set, no evaluation and no described retrieval method appears on any first-party surface read, and the grounding that does exist is real but only asserted: outputs are described as verifiable and structured step by step, over documents the customer uploaded and can therefore open. The bottom limb does not fire, because a hallucination claim standing alongside a real document-grounded architecture is not a bare claim, but a claim that a system never hallucinates, published beside a term disclaiming accuracy, is the marketing-versus-agreement gap in one page. The unread Matrix product page is the rebuttal route and would be the place a described method appears. Security page, MSA and Matrix 2.0 post read 7 September 2026.
Grounding is documented in vendor material, short of any figure an outsider can test. Review output links each cell to its source. On 14 Sep 2026 the vendor published how its rebuilt research layer works: an ontology of the law that maps how authorities rank and relate, including amendments, holdings against dissents and temporal validity, and an AI native citator whose standard is set and audited by a team of former publisher attorney editors, built on technology from its Qura and Wexler acquisitions. The vendor also states it has catalogued more than 50 distinct ways AI fails at legal research, without publishing the list. The research layer is in limited beta, with general availability planned for the fourth quarter of 2026. Not located as of 24 Sep 2026: a published accuracy measurement, an evaluation framework or a hallucination rate, which is what separates this from an A.
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.
Autonomy is claimed at length and oversight is asserted without a described control. The product is sold as agents that complete end-to-end tasks rather than answering questions, running multi-step analysis across whole document sets and carrying the work through to a memo, a deck or a model; the vendor's framing is an operating system for complex work. What sits on the oversight side is transparency of output rather than a control structure: sourcing is shown for each cell of the grid so a user can check where an answer came from, and the structured step-by-step format is presented as the guarantee of quality. No located material states what an agent may do without a person, at what point it stops, what review the vendor expects before an output is used, or what happens when it is wrong beyond the warranty disclaimer. Legal solution page, security page and Matrix 2.0 announcement read 7 September 2026.
Human oversight is asserted as a governing principle in vendor material and is covered in the abstract by an ISO 42001 certification of the AI management system. What is not published is the mechanism: where the review point sits, what an agentic workflow does on its own, at what threshold it stops, and what a supervising lawyer must approve. Searched the vendor site, blog and trust center on 29 Aug 2026. Oversight appears as a stated principle rather than a described control.
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.
Logos stand in for evidence on the vendor's own surfaces, and the figures live somewhere else. The legal page displays the marks of Latham & Watkins, Cleary Gottlieb, Ropes & Gray and Seyfarth with no accompanying statement of what any of them does with the product, no attributed quotation and no result. The quantified claims that exist are on third-party pages and are described here rather than credited: OpenAI's case study on Hebbia states that law firms reduce credit agreement review time by 75 per cent, saving $2,000 an hour in legal fees, and that investment bankers save 30 to 40 hours a deal, and Andreessen Horowitz's investment announcement reports unnamed customers saying analyses that took two to three hours now take two to three minutes. None of that is published by the vendor, none names the firm behind the figure, and no method is stated for any of it. A named, dated deployment on Hebbia's own surface is the route to a higher grade. Legal page read and third-party material reviewed 7 September 2026.
Named customers appear in vendor material, including a published Grant Thornton UK forensic investigations story, and the vendor states more than 1,000 customers across 50 plus markets. Additional named users including Cleary Gottlieb, Goodwin, Linklaters, White and Case, Dentons and Barclays appear in vendor recruiting material. Not located as of 29 Aug 2026: dated outcome figures with a method a reader could assess, which is what separates this from an A.
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, short of the segregation and privilege limbs. What is published and binding: Customer Materials are the customer's Confidential Information and the customer owns them (MSA 6.1, 3.3); Hebbia will not access or use Customer Materials to provide technical support without the customer's documented consent (1.2), which is a sharper commitment than most agreements in this corpus make; Customer Materials are deleted within 30 days of termination (5.4(c)) and personal data is deleted on request at any time (DPA Annex II); the position on model providers is explicit, with OpenAI, Anthropic, Google and Microsoft named in the DPA as processing user prompts and files. Two limbs are missing. Nothing located addresses segregation between users, matters or clients inside a customer tenant, which matters for a product sold to four named AmLaw firms. Privilege and work product are not addressed anywhere, which R33 treats as a required limb of the top band. The training position is also adverse and is graded on its own signal: the agreement permits improving machine learning models with Customer Materials where those models are solely for that customer. MSA and DPA read in full 7 September 2026.
Substantive published commitments, and unusually for this market they sit in openly published contract documents rather than only on a trust page: general terms and conditions in EU and US versions, a data processing agreement, and a security measures annex covering least privilege access, personnel confidentiality obligations, authorization controls and retention on customer instruction. The security page states the vendor will not use customer data to train or fine tune models. Two gaps hold this off an A. Attorney client privilege and work product handling is not addressed directly in located material, and matter level segregation between users is not documented.
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. Where the advice line is not the duty a product raises, the axis is read through the nearest professional duty it does raise: judicial conduct rules and the reviewing duty for products sold only to courts, and the duty to bill for time actually spent for products that draft time entries.
No position on the advice line was located, for a platform sold to law firms for due diligence, capital markets drafting and matter work. The Main Services Agreement was read in full and addresses use restrictions, warranties, indemnities, liability and arbitration without touching advice, competence or supervision; its disclaimer at 7.1 is a warranty disclaimer rather than a statement of what the product is and is not. The acceptable use policy governs misuse, infringement and harassment, not the practice of law. The legal solution page describes lawyers using the product to advise better and become the trusted adviser clients rely on, without a corresponding statement that the output is not legal advice. The audience is unambiguous and professional, which is recorded rather than credited. This records what is establishable on the date; a product disclaimer inside the application, if one exists, is the rebuttal route. Surfaces checked 7 September 2026.
The intended audience is unambiguously lawyers, firms and in house teams, and vendor material describes the product as working with lawyers rather than replacing them. Searched the vendor site, the published acceptable use policy, the general terms and the blog on 29 Aug 2026 and located no published position on the advice line, no treatment of competence and supervision duties, and no statement of jurisdiction limits.
AI Governance and Bias Disclosure
Published governance over model behavior: who owns it inside the vendor, what is tested before release, and what is disclosed about disparate output across matter types, parties, or populations.
A published, independently audited governance framework, short of testing results or a named owner. The legal solution page displays ISO/IEC 42001:2023, the management system standard for artificial intelligence, alongside SOC 2 Type II, and the security page repeats an ISO mark. A certification against an AI management standard is real substance and carries more weight than a self-published principles page, which is the position this index has taken since the Corlytics and Ontra records. What is not published is what the standard's certification would sit on top of: no governance policy, no statement of who inside Hebbia is accountable for model behaviour, no description of what is tested before an agent ships, and no disclosure of any finding about uneven output across matter types or populations. The certificate itself, its scope and its date were not located, and the Vanta-hosted trust centre that would carry them returns page metadata with no body on this channel, which is recorded as a retrieval limit and named as the rebuttal route. Legal page, security page and trust centre attempted 7 September 2026.
Holds an ISO 42001 certification covering its AI management system, independently audited with ongoing surveillance, and publishes what the certification covers: how AI is designed, deployed, supervised and monitored, with human oversight and structured governance named as the operating principles. That is a published governance framework with real substance and independent validation, which is rare in this market. Not located as of 29 Aug 2026: a named internal owner of AI governance, published pre release testing results for model behaviour, or any disclosure about uneven output across matter types, parties or populations.
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.
Retention, deletion, access control, subprocessors and incident practice are all published and specific, and the whole set sits in documents a buyer can read before signing. Annex II of the Data Processing Agreement is a full technical and organisational measures table: AES-256 encryption at rest through AWS-managed keys, TLS 1.3 in transit, multi-factor authentication and single sign-on with two-factor required on all production access, hourly backups of production datastores that are periodically tested, monitored security logging with escalation, change management automated through CI/CD, and annual SOC 2 Type II audits covering the security criteria. Deletion is stated twice over: customers may request deletion at any time and delete through self-service, all personal data is deleted following termination, a subject-access form is published for erasure and portability, and MSA 5.4(c) deletes Customer Materials within 30 days of termination. Annex III names every subprocessor with its processing location, and the DPA commits to listing and notifying any new one ten days before it touches personal data, with an objection right. The one softness is the incident clause, which requires Hebbia to inform the customer without undue delay and cooperate to the timescales the law requires rather than committing to a fixed number of hours. DPA and MSA read in full 7 September 2026.
The published security measures annex covers access on a least privilege and role based model, centrally stored logs traceable to unique usernames with security logs retained at least 12 months, data integrity signing, personnel background checks and confidentiality agreements, and retention set by customer instruction. The published data processing agreement commits the vendor to assist with the customer's own breach notification obligations, so incident practice is addressed. ISO 27001:2022 is audited yearly. Not located as of 29 Aug 2026: a current named subprocessor list, which is the remaining element of the A bar.
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.
What the vendor stands behind is published and specific, including the places where it stands behind nothing. Section 8.2 gives a defence and indemnity against third-party claims that Hebbia's technology infringes a US patent, copyright or trade secret, with three named exclusions and the procure, replace or terminate-and-refund mitigation ladder; section 8.3 makes the indemnities the only remedy for third-party intellectual property claims. Section 9.2 caps aggregate liability at fees paid or payable in the twelve months before the event, expressly including indemnity claims, so a buyer can see that the indemnity is inside the cap rather than outside it, which is less generous than several peers and is stated rather than hidden. Section 9.1 excludes indirect loss and data inaccuracy for both parties, with Excluded Claims at 9.3 lifting that exclusion for the intellectual property indemnities and for a customer breach of the use restrictions. Section 7.1 disclaims all warranties in capitals and states that the platform is not warranted accurate or complete. The invocable remedy is the service level agreement: 99.9 per cent monthly availability with a day of pro-rated fees credited for each hour of downtime, capped at one week of fees a month and claimable within 24 hours. No insurance is stated and no warranty of performance exists. MSA read in full 7 September 2026.
A real published position, which is uncommon here. General terms and conditions are published openly in EU and US versions and carry numbered liability clauses, aggregate caps that apply across the subscriber and its affiliates, a separate cap of 100,000 Euro on beta features, and a carve out concept the terms call an Enhanced Claim. A buyer can read the allocation of loss before entering a sales process. Not located as of 29 Aug 2026: indemnity scope for third party claims arising from output, any warranty on output, and any insurance position, so the full picture is short of an A.
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.
Real integrations are named in a binding document, short of depth and short of the systems legal work actually lives in. The Data Processing Agreement names Merge API as an opt-in subprocessor applicable only to customers who elect to configure a Salesforce or Outlook connector, and Google LLC as providing email and document management capabilities through Google Workspace Enterprise alongside its model role, so the connectors are documented with what they carry and when they apply. The Matrix 2.0 announcement adds a financial data feed, Daloopa, as a connected source. What an implementer would need is not published: no description of direction, sync behaviour or configuration was located. The gap a legal buyer should notice is what is absent from the list rather than what is on it: no document management system, matter management, e-billing or practice management integration was located on any surface read, so a firm's documents reach the platform by upload or through a finance-oriented connector set. DPA, Matrix 2.0 post and product pages read 7 September 2026.
Real and named integrations: an iManage technology partnership working through iManage APIs with an announced expansion, SharePoint, Box, a Microsoft Word add in for drafting and redlining, Outlook, EDGAR, and import from virtual data rooms and contract lifecycle systems. On 17 Sep 2026 the vendor released a plugin for ChatGPT Enterprise that brings its grounded answers into the ChatGPT interface. The vendor is explicit that it integrates with document management systems rather than replacing them. Not located as of 29 Aug 2026: implementer level documentation describing what each integration moves, in which direction, and what an administrator must configure, which is what the A bar asks for.
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.
Residency is published in unusual detail and the deployment tiers are not. Annex III of the Data Processing Agreement gives a processing location for every subprocessor: Amazon Web Services down to the named regions us-east-1, us-west-1, eu-west-1 and eu-central-1, and for each model provider and inference host either all US locations or EU regional processing, with document parsing and telemetry recorded as regional based on tenant. That answers where processing happens as distinct from where data is stored, provider by provider, which is more than most records on this axis publish. Transfers out of the EEA and the United Kingdom run on the 2021 Standard Contractual Clauses with the UK Addendum, completed in the DPA down to the module, the docking clause and Irish governing law. DPA 3.5 states that processing may occur in any country where Hebbia, its affiliates and its authorised subprocessors maintain facilities, subject to law. What is not published is a deployment choice: no single-tenant, private-cloud or on-premise option is described, no statement of tenancy separation was located beyond encryption and logical measures, and nothing states what a customer can select or what changes between tiers. DPA read in full 7 September 2026.
The claim is made and the detail is not published. Vendor material states flexible storage options matched to data sensitivity, and separate EU and US contract documents indicate region specific arrangements, with the technical team in Sweden operating under GDPR. Searched the vendor site, the security pages and the published legal documents on 29 Aug 2026 and located no list of available regions, no tenancy model, and no statement of where processing happens as distinct from where data is stored.
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.
Certification is real and stated in a binding annex, short of accessible evidence. Annex II of the Data Processing Agreement records that Hebbia has undergone a SOC 2 Type 2 audit covering the security trust service criteria and undergoes annual SOC 2 Type II audits, and the customer-facing pages display SOC 2 Type II and ISO/IEC 42001:2023 marks. The DPA also gives a real audit right: the customer or its appointed third-party auditors may audit compliance and Hebbia must make relevant information, policies, records and staff available, once in any twelve months and more often after a security incident. What is missing is the accessible half. No auditor is named, no audit period or report date appears, no scope statement is published, and the trust centre at trust.hebbia.ai is a Vanta-hosted portal that returns page metadata with no body on this channel, so whether any report is self-serve could not be established. That is recorded as a retrieval limit and the portal is the rebuttal route; the lower tier is graded and the reason stated. DPA read in full and trust centre attempted 7 September 2026.
Three current independent certifications, ISO 27001:2022 audited yearly, SOC 2 Type II, and ISO 42001 for AI management, with a public trust center at security.legora.com carrying a resources section and a data flow diagram, plus security whitepapers and a security measures annex published openly with no gate at all. The open publication of the security annex is full credit rather than a request flow. The certifying auditor is not named in located material, which is the one thing a peer in this category does publish.
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 supply chain is disclosed further than any other record in this lane, and stops short of naming the models. Annex III of the published Data Processing Agreement names OpenAI, Anthropic, Google and Microsoft as providing large language model capabilities through their APIs, states for each that the subject matter processed is user submitted prompts and files, and gives the processing location as all US locations or EU regional processing. It goes a layer deeper than most: Cerebras, Groq, Baseten, Fireworks AI and Modal Labs are each named as hosting a subset of Hebbia's own inference infrastructure, with Elasticsearch for search and indexing over prompts, files and generated artefacts, MongoDB for embedding and re-ranking, and Reducto for document parsing. Change notification is contractual, with any new subprocessor added to the list and notified ten days before it accesses personal data, and an objection right attached. What is absent is model naming: no specific model or version is identified on any first-party surface, and provider naming is not model naming. Third-party material fills that gap and is described rather than credited, since OpenAI's own case study names o3-mini, o1 and GPT-4o inside Matrix and Anthropic's names Claude as a model customers can select. DPA Annex III read in full 7 September 2026.
Vendor material refers to how it implements, supervises and evolves AI without identifying what sits underneath. Searched the vendor site, the trust center, the published general terms, the data processing agreement and the security measures annex on 29 Aug 2026 and located no named model provider, no subprocessor list and no commitment to notify customers when the supply chain changes. Third party sources describe a multi model approach running on Microsoft Azure, which is not vendor material and does not move this axis.
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.
There is a page called Pricing and it contains no price. It publishes positioning, a product name and three industry links, and every button on it requests a demo; no figure, band, tier or unit appears. What lifts this off the floor is the agreement. Fees are set in an Order Form and payable in US dollars unless the Order says otherwise; the licence is per seat, since administrators may provision Authorized Users up to a maximum stated in the Order Form; orders renew automatically for one-year terms unless either party gives 30 days' notice; Hebbia may notify different renewal pricing at least 45 days in advance, which tells a buyer that renewal increases are unilateral and time-boxed; payment is due within 30 days, late amounts carry 1.5 per cent a month, and invoices must be disputed within 60 days. So the unit and the mechanics are readable before a sales call and the number is not. Third-party seat prices circulating on review sites are not evidence and are not recorded. No VendorPricing row is written, since a row belongs to vendors graded A or B on this axis. Pricing page and MSA read 7 September 2026.
Searched the vendor home page, product pages, newsroom and legal pages on 29 Aug 2026. No pricing page, no published rate, no stated unit of charge and no published tier structure located. The only commercial entry point is a demo request, which is sales gated and earns no credit. Several independent third party pricing analyses state the same, and one reports a consumption based tier that was not located on the vendor site. Third party per seat estimates are not vendor published and do not move this axis. Note that the vendor does publish its general terms openly, which is transparency of terms rather than of price and is graded on the liability axis instead.
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.
Coverage is described with real substance and one significant limit is published, short of the full boundaries. The legal solution page names six areas of legal work the product is built for: banking and finance documentation, buy-side and sell-side due diligence, capital markets drafting from filings and precedent transactions, private equity fund formation and portfolio governance, matter intelligence built from a firm's past work, and firm-tailored workflows. The wider platform serves investing, banking and corporate finance, so a legal buyer can see that the product spans transactional and financial work rather than litigation or contentious practice, and no dispute, court or regulatory practice area is claimed. The limit worth reading before purchase is contractual rather than marketing: MSA 2.4 with definition 12.13 prohibits the customer from submitting Restricted Information without Hebbia's prior written approval, and Restricted Information covers sensitive personal identifiers, protected health information and personal data as defined in the GDPR, with DPA Annex I separately prohibiting special categories. A firm running European matters, employment work or anything touching health records needs that approval first. Firm segments are not broken down, government and in-house use are not addressed, and what is unsupported is otherwise left open. Legal page, MSA and DPA read 7 September 2026.
Segment coverage is described with substance: large law firms, in house legal departments, and professional services including a published forensic investigations customer story, spanning more than 1,000 customers across 50 plus markets with multi jurisdiction and cross border work as a stated strength. Practice coverage spans review and diligence, research, drafting and investigations. Not located as of 29 Aug 2026: any statement of the boundaries, meaning which firm sizes or practice areas the product is not built for.
The 12 legal signals, side by side
Recorded rather than graded. These are the questions a practitioner has to answer before a tool touches a client matter, and the answers are taken from public material only.
Client Data in Training
Can material a lawyer puts into this product be used to train a model?
The published agreement expressly permits training on customer content, in a bounded form, while the marketing says the opposite. Main Services Agreement clause 3.2 grants Hebbia a license to use Customer Materials to provide and support the platform and then adds that Hebbia may use Customer Materials to improve machine learning models that are solely for use by that customer, and may not license or otherwise make models improved with Customer Materials available to any third party.
No opt-out is located. Both halves belong in the record. The security page states that Hebbia is one of the only AI companies that never trains on customer data and the legal page carries a badge reading no training on user data, which is what a buyer sees first. The privacy policy is drafted more carefully than the badge and is consistent with the clause, saying that personal data in customer data is not used to train or improve generalized or third-party models.
So the permission is real, it names machine learning models, and it operates on customer content, but it is confined to a model only that customer uses and expressly barred from reaching anyone else, which is narrower than the shapes that carry this value elsewhere in the index. MSA, security page, legal page and privacy policy read 7 September 2026.
The vendor home page and security page both state that customer data is not used to train or fine tune any AI models. The vendor publishes its general terms and conditions and its data processing agreement openly, and a training prohibition was searched for in those documents on 29 Aug 2026 and not located, so the commitment as recorded rests on the security and marketing pages rather than on a located contract term. The full agreement text was not read end to end.
Prompt and Output Retention
How long does the product keep what a lawyer typed, and can that be set to zero?
The customer controls deletion and no retention period is fixed, with zero retention not offered. Annex II of the Data Processing Agreement states that customers determine what data they route through the service, that they may request deletion at any time, that Hebbia deletes personal data on written request where self-service deletion is not available, and that all personal data is deleted following service termination; a subject-access form is published for erasure and portability, and the Main Services Agreement adds that Customer Materials are deleted within 30 days of termination.
This record is more precise than most on what the retained material actually is: Annex III states for each model provider and inference host that the subject matter processed is user submitted prompts and files, so prompts and generated artifacts are squarely inside the regime rather than left to inference, and Elasticsearch is named as indexing prompts, files and generated artifacts. What is not published is an in-term retention period, any statement of how long prompts persist while a subscription runs, or a no-retention setting. DPA and MSA read in full 7 September 2026.
Section 17 of the published security measures annex, dated 31 Jan 2025, states that during the term of the data processing agreement personal data is subject to the retention requirements the subscriber instructs from time to time, and that after termination or expiry clause 11 of that agreement governs. Retention is therefore customer instructed and the commitment sits in a contract document rather than a policy page, which is the contractual form of control this value describes and is enforceable in a way a policy page is not.
Recorded at customer controlled rather than the top value because no retention period is published and no zero retention setting was located as of 29 Aug 2026. Also on the same document and worth factoring in as a retention floor: subscriber environments are logically separated at all times, and full production backups are taken every four hours.
Ethical Walls and Matter Segregation
Does retrieval respect the firm’s ethical walls, or can the model read across them?
No located public material addresses ethical walls or matter-level segregation. The agreement covers administrator provisioning of Authorized Users and account credentials, and the security measures annex covers encryption, authentication and logical protections between Hebbia's customers, but nothing read describes whether retrieval inside a customer's own workspace respects walls between matters, clients or teams, or whether an agent running across an uploaded corpus can reach material a particular lawyer should not see.
That is a live question for a platform whose legal page names four large law firms and whose core function is analysis across an entire document set at once. The unread Matrix and Max product pages and the gated trust center are the rebuttal routes. MSA, DPA and legal page checked 7 September 2026.
Searched the vendor site, the iManage partnership announcement, the security page, the published security measures annex and the trust center on 29 Aug 2026. No vendor material was located addressing whether retrieval enforces document management system permissions at query time per user, or how ethical walls and matter level segregation are handled. The security annex documents least privilege access for vendor personnel, which is a different question. A partner case study describes per query authentication, which is not vendor material.
Third Party Request and Subpoena Notice
If someone subpoenas the vendor for a firm’s data, does the firm hear about it first?
The agreement commits to prior notice where permitted, with assistance to contest. Main Services Agreement clause 6.3 permits either party to disclose the other's confidential information when required by law or regulation only on condition of giving prior notice of the compelled disclosure, to the extent permitted, and providing reasonable assistance at the disclosing party's cost to contest or limit it. Customer Materials are the customer's Confidential Information under 6.1, so the clause reaches uploaded documents rather than only account data.
The Data Processing Agreement adds that Hebbia will promptly inform the customer, with full details, of any request or complaint from a data subject, regulator or third party unless prohibited by law. No transparency report, request statistics or law-enforcement guidelines page was located, which is what separates this from the top value. MSA and DPA read in full 7 September 2026.
Searched the published EU and US general terms and conditions, the data processing agreement, the security measures annex and the acceptable use policy on 29 Aug 2026. No clause addressing government or law enforcement requests for customer data was located, and no transparency report was located. This records a search that did not surface the clause rather than a reading of the full agreements end to end.
Primary Law Corpus Provenance
Where does the law in this product come from, and does the vendor have the right to use it?
The sources are identified and no license basis is stated for the ones Hebbia supplies. The product analyses documents the customer uploads, and the agreement puts the rights basis on the customer, who warrants at 2.4 that it has the necessary rights, licenses and permissions to provide the materials. Two connected sources sit outside that: Google Workspace Enterprise reaches email and documents through a connector, and the Matrix 2.0 announcement adds Daloopa, described as a financial data feed now available to Hebbia customers, with no statement of the license, coverage or update cadence behind it.
No corpus of law, filings or public authority is claimed anywhere, and the signal's law-corpus limbs do not bite for a platform whose corpus is the customer's own data room. Legal page, MSA, DPA and Matrix 2.0 announcement read 7 September 2026.
Research coverage is described by jurisdiction, reported at twelve, and the vendor announced the acquisition of Qura, a Stockholm legal database covering case law, legislation and regulation, which it is extending to larger markets. What is not identified is the corpus itself: which publishers or public sources the law comes from, the license or public domain basis for each, and the update lag. Searched the vendor site, newsroom and product pages on 29 Aug 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. The product analyses credit agreements, filings, offering memoranda, fund documents and a firm's own past matters, and cites back to the documents the customer supplied; nothing read describes a citator, a treatment signal or a verification prompt for reported authority, and the vendor does not claim to retrieve case law. The signal's limbs do not bite for a transactional document-analysis platform, and that is recorded rather than graded around. Legal page, security page and Matrix material checked 7 September 2026.
The vendor published on 14 Sep 2026 that it is building its own AI native citator on an ontology of the law that captures the hierarchy of authority, how sources relate to each other and temporal validity, so that an amended rule or an overruled case is recognized as such. The standard is set and audited by a team of former publisher attorney editors, and the technology comes from its Qura and Wexler acquisitions. Recorded at the own treatment value because the method is described and the check is computed by the vendor rather than licensed from a commercial citator.
The citator is in limited beta with general availability planned for the fourth quarter of 2026, so it may not yet be switched on for a given account, and no treatment coverage figure or error rate was located as of 24 Sep 2026.
Refusal and Uncertainty Behavior
What does the product do when the answer is not in the corpus?
No located public material describes what the product does when it cannot ground an answer. The vendor's position is the opposite claim, that it is building large language models that never hallucinate and give correct, verifiable answers, which asserts the problem away rather than describing an abstention path. Nothing read sets out a no-answer state, a confidence or grounding score, or any published evaluation in which the system declines.
The same page's claim sits against the agreement's disclaimer that the platform is not warranted accurate or complete, and that tension is recorded on the citation accuracy axis rather than counted twice here. Security page, MSA, legal page and Matrix 2.0 announcement checked 7 September 2026.
Searched the vendor site, blog, newsroom and trust center on 29 Aug 2026. No published material was located describing what the product does when it cannot ground an answer, whether an explicit no answer path exists, or whether any confidence or grounding signal is exposed to the user.
Fabricated Citation Record
Does a public court record exist addressing fabricated or hallucinated legal citations in output from this product?
No court order, opinion or disciplinary record naming Hebbia or Matrix was located as of 7 September 2026. The AI Hallucination Cases database maintained by Damien Charlotin was searched on both names alongside a general search of sanctions coverage; the decisions that name legal-specific products name other vendors. This is a statement about the public record, not a finding about the product. The exposure profile is worth stating: the platform analyses documents a customer supplies and cites back to them rather than generating legal authority, so the fabricated-citation risk runs through what a user does with an output rather than through invented case law.
No court order, opinion or disciplinary record naming this product has been located as of 29 Aug 2026. Instrument searched: the AI Hallucination Cases database maintained by Damien Charlotin, which tracks court decisions worldwide involving AI generated hallucinated content and records the AI tool implicated where it is known. Also checked published 2026 sanctions summaries and secondary sanctions trackers. The entries located name filers, and in some rows other products, rather than this one.
This is a statement about the public record on the date shown and not a clearance, and it is bounded by what that database covers.
Bar Guidance Alignment
Has the vendor engaged in public with the ethics opinions its buyers are bound by?
No located public material engages with bar or ethics guidance, and none engages with lawyers' professional obligations even in general terms. The legal solution page speaks to winning and keeping clients, closing deals faster and becoming the trusted adviser, which is commercial rather than professional-responsibility framing. The security page's responsibility section commits to responsible use of the platform, and the acceptable use policy addresses illegal, deceptive or infringing use; both are about misuse of software rather than about a lawyer's duties.
Nothing names an ABA formal opinion, a state bar opinion, a regulator's AI guidance or a court standing order. The lower value was tested before being taken: a generic reference would require some engagement with professional responsibility, and none was located. Legal page, security page, acceptable use policy and MSA checked 7 September 2026.
Searched the vendor site, blog, newsroom and resource pages on 29 Aug 2026. No engagement with any named ethics opinion was located, including ABA Formal Opinion 512 and state or national bar guidance. The vendor publishes substantial governance and certification material, which addresses its own AI management system rather than the professional responsibility obligations its buyers are bound by.
Billing and Fee Posture
Does the vendor address what happens to the bill when the work takes an hour instead of six?
The product sits inside law firms' fee relationships with their clients and the published position is a savings claim. The legal page is built on three promises to a firm: close deals faster by speeding up due diligence and contract review, grow the book by freeing the team from repetitive document work, and go deeper on every deal. Third-party material puts numbers on the same claim, with OpenAI's case study reporting that law firms cut credit agreement review time by 75 percent and save $2,000 an hour in legal fees, which is a statement about billable work compressing.
Nothing located addresses what happens to the client's bill when it does: no guidance on recording AI-assisted work on a matter, no disclosure treatment, and no per-matter record of what the agents did. Legal page, MSA and third-party material checked 7 September 2026.
Vendor material is framed around speed and volume, describing analysis of thousands of documents in minutes and teams moving faster. Searched the vendor site, blog and legal pages on 29 Aug 2026 and located no per matter record of AI assisted work intended for fee purposes, and no published guidance on billing, fee or client disclosure treatment.
Outside Counsel Guideline Readiness
Can a firm get this vendor through a client’s AI clause without a bespoke negotiation?
A firm can answer a client's AI clause from published material without asking for anything. Annex III of the Data Processing Agreement is a complete subprocessor register: OpenAI, Anthropic, Google and Microsoft are each named as providing large language model capabilities through their APIs, with the subject matter recorded as user submitted prompts and files and the processing location given as all US locations or EU regional processing; Cerebras, Groq, Baseten, Fireworks AI and Modal Labs are named as hosting parts of the inference infrastructure; Amazon Web Services, Auth0, Elasticsearch, MongoDB, Reducto, Merge API and Datadog complete the list with their roles and locations.
The forwardable material is the same document: the DPA is published in full with the EU Standard Contractual Clauses and UK Addendum completed in its schedule, the security measures set out in Annex II, and a commitment to list and notify any new subprocessor ten days before it accesses personal data with an objection right. That satisfies the subprocessor list, the statement of which model providers see client content, and the client-facing artifact together.
One condition a firm should carry into the conversation: the agreement prohibits submitting personal data as defined in the GDPR without prior written approval. DPA read in full 7 September 2026.
Read against the artifacts this signal turns on, the picture is mixed and the value understates one half of it. Openly published with no gate and forwardable to a client today: general terms and conditions in EU and US versions, a data processing agreement, a security measures annex and an acceptable use policy. Not located as of 29 Aug 2026: a current subprocessor list, any statement of which model providers see client content, and any client facing consent or notification pack. The trust center carries a resources section that routes document access through a request.
Court Disclosure Support
If a judge’s standing order requires an AI disclosure, can the product produce one?
One element of a record exists and no export is built for disclosure. Outputs are described as verifiable, with sourcing shown for each answer in the grid so a user can trace a cell back to the document it came from, which is source attribution and is genuinely one of the three things a disclosure needs. The other two are absent from everything read: no per-document export covering which model produced an output, and no record of human verification.
Nothing addresses court disclosure obligations, standing orders on AI use, or a certification a filer could attach, and no template or guidance is published. The security annex's logging is monitoring of access to systems rather than a user-facing record of what an agent did. Security page, MSA, DPA and legal page checked 7 September 2026.
Searched the vendor site, product pages and published legal documents on 29 Aug 2026. Partner material describes citations being preserved when work is exported to Word, and the security annex documents security logging traceable to unique usernames, which is an infrastructure control rather than a record of AI assisted work. No per document record covering model used, sources retrieved and human verification was located.
The questions both sides leave open
Derived from the records above rather than written, so it cannot favor either vendor. Take these into both conversations and ask each side the same question.
- Ethical Walls and Matter Segregation
- Refusal and Uncertainty Behavior
- Bar Guidance Alignment
Which one fits
Choose Hebbia if
- Your clients' AI clauses ask which providers see their documents. Hebbia's data processing agreement names OpenAI, Anthropic, Google and Microsoft as processing user prompts and files, names Cerebras, Groq, Baseten, Fireworks AI and Modal Labs as hosts of its inference, gives a processing location for each, and commits to ten days' notice before any new subprocessor.
- You want to read the allocation of loss before a sales call. Hebbia's Main Services Agreement gives a defense and indemnity against third party intellectual property claims, caps each side at twelve months of fees with the indemnity inside that cap, and commits to 99.9 percent monthly availability with a day of fees credited for each hour of downtime.
- You need to know where processing happens, provider by provider. Hebbia's agreement names four AWS regions across the United States and Europe and records, for each model provider and inference host, whether it processes in the United States or regionally in the EU, with Standard Contractual Clauses completed for European transfers.
Choose Legora if
- Your work includes legal research, not only document review. Legora sells research inside the same workspace, reports coverage across twelve jurisdictions, and on 14 September 2026 described a rebuilt research layer with an AI native citator, in limited beta with general availability planned for the fourth quarter. Hebbia retrieves no case law and claims no citator.
- Your security review wants the AI management system certified and the evidence reachable. Legora holds ISO 27001:2022, SOC 2 Type II and ISO 42001, and runs a public trust center with a data flow diagram, while its security measures annex is published with no gate at all.
- Your matters carry European personal data. Legora publishes its general terms in EU and US versions and its data processing agreement openly, with retention set by customer instruction. Hebbia's agreement bars customers from submitting personal data as defined in the GDPR without its prior written approval.
In summary
Hebbia
Hebbia is a New York company whose platform, Matrix, runs multi step AI analysis across large document sets and returns a grid in which each row is a document, each column a question and each cell shows its source; Max is its agent product. It sells to asset managers, banks and law firms, naming due diligence, credit agreements, capital markets drafting and fund formation among its legal uses. The AI Legal Index grades it in the top two bands on ten of fifteen capability axes, with A grades on AI centrality, data stewardship and liability. Its published data processing agreement names four model providers and five inference hosts with a processing location for each. As of 7 September 2026 the index located no accuracy measurement, no published price and no position on the line between legal tooling and legal advice.
Legora
Legora is a collaborative AI workspace for law firms and in house legal teams, headquartered in Stockholm and founded in 2023, covering document review in a grid linked to sources, drafting in Word, agentic workflows and legal research. The AI Legal Index grades it in the top two bands on ten of fifteen capability axes, with A grades on AI centrality and security certifications: it holds ISO 27001:2022, SOC 2 Type II and ISO 42001, and publishes its general terms, data processing agreement and security annex openly. On 14 September 2026 it described a rebuilt research layer with its own AI native citator, in limited beta. It states that customer data does not train its models. As of 24 September 2026 the index located no published price, no named model provider and no subprocessor list.
Questions buyers ask
Hebbia vs Legora: which is better for law firm document review?
Neither on the totals: the AI Legal Index places both in the top two bands on ten of fifteen capability axes. The tie breaks on what a firm needs to read. Hebbia publishes more in its contracts, including an intellectual property indemnity, service credits and a full list of the model providers that process client files. Legora publishes more on accuracy and security, with a described citator in beta, ISO 42001 certification and an open trust center, and it adds legal research to review.
Does Hebbia train its AI on client data?
Hebbia's site says it does not train on customer data, and its Main Services Agreement sets the boundary precisely. Clause 3.2 permits Hebbia to use customer materials to improve machine learning models that are solely for that customer's use, and bars it from licensing or making those models available to any third party. Its privacy policy says personal data in customer data does not train generalized or third party models. No opt out from the customer specific improvement was located. Graded by AI Legal Index against 15 capability axes and 12 legal signals, including privilege handling and citation accuracy, from each vendor's own published materials, verified September 25, 2026. No vendor pays for placement.
Which AI models do Hebbia and Legora use?
Hebbia names its providers in its data processing agreement: OpenAI, Anthropic, Google and Microsoft process user prompts and files, and Cerebras, Groq, Baseten, Fireworks AI and Modal Labs host parts of its inference, with a location for each and ten days' notice before any new subprocessor. It does not name specific models or versions. Legora names no model provider and publishes no subprocessor list, and its trust center routes document access through a request. Graded by AI Legal Index against 15 capability axes and 12 legal signals, including privilege handling and citation accuracy, from each vendor's own published materials, verified September 25, 2026. No vendor pays for placement.
Do Hebbia and Legora work with iManage?
Legora does. It has an iManage technology partnership working through iManage APIs, connects to SharePoint, Box, Outlook and Word through an add in, and released a ChatGPT Enterprise plugin on 17 September 2026. Hebbia names no document management integration. Its published connectors are Salesforce and Outlook through Merge API, Google Workspace, and the Daloopa financial data feed, so a firm's documents reach it by upload or through those routes. Graded by AI Legal Index against 15 capability axes and 12 legal signals, including privilege handling and citation accuracy, from each vendor's own published materials, verified September 25, 2026. No vendor pays for placement.
What do Hebbia and Legora both leave unpublished?
A price, first: Hebbia's pricing page carries no figure and Legora has no pricing page. Neither publishes an accuracy measurement or says what the product does when it cannot ground an answer. Neither documents how ethical walls or matter separation work inside a firm's own workspace, which matters for tools that run across a whole document set at once. Neither names an ethics opinion, including ABA Formal Opinion 512, and both market speed without saying how saved time reaches a client's bill. Graded by AI Legal Index against 15 capability axes and 12 legal signals, including privilege handling and citation accuracy, from each vendor's own published materials, verified September 25, 2026. No vendor pays for placement.
Both training positions deserve a close read. Hebbia's agreement tells buyers it may use customer materials to improve machine learning models used solely by that customer and never made available to anyone else, a narrower permission than a general training right and the binding text behind the no training statements on its site. Legora states on its security page that customer data does not train its models, and that commitment was not located as a term in its published agreements. Legora's citator is in limited beta, so it may not yet be switched on for a given account. Hebbia was verified on 7 September 2026 and Legora on 24 September 2026. Neither vendor reviewed this page.
Neither vendor paid for inclusion, placement or a grade, and neither reviewed this page before it published. Everything above comes from public material on the dates shown. How the index grades.