Hebbia vs Legora: how they compare in 2026

Matrix is described by the vendor as a platform for building AI agents that complete whole 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 names 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 beneath the models; without them there is nothing left to sell.

Everything Legora sells is generative. The Agent plans and executes legal work across the platform's tools, Tabular Review extracts answers from each document into a grid, research answers questions across jurisdictions, and drafting runs in Word, Outlook and the built in editor. An orchestration layer handles tool routing, memory, model selection and guardrails beneath them. Without the models there is no product left to sell.

The security page states that Hebbia is creating the standard for trustworthy large language models that never hallucinate and give correct, verifiable answers. Section 7.1 of the Main Services Agreement 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, test set, evaluation or described retrieval method appears on the security page, the legal page or in the agreements. Outputs are described as verifiable and structured step by step, over documents the customer uploaded and can open. Hebbia also publishes a Matrix product page.

Each cell in Tabular Review expands to its reasoning and the source documents behind it, and Portal answers carry links back to the exact source. Legal research works several sources in parallel, ranks them by authority and checks each finding against primary law before the answer returns. The rebuilt research layer runs on an ontology of the law that maps how authorities rank and relate, including amendments and temporal validity, and went live in the United States on 29 October 2026. Legora says three people check every word of that data to reach 99.997 percent accuracy, and that the ontology makes the Agent 40 percent faster. US case law is updated every 24 hours. Its AI native citator, built to a standard set and audited by former publisher attorney editors, is still to come. Legora publishes its BAR benchmark of 5,161 cases across 28 practice areas, last updated in July 2026, and says it has cataloged more than 50 ways AI fails at legal research without publishing the list. No hallucination rate for the Agent's answers is published.

The product is sold as agents that complete whole tasks rather than answering questions, running analysis in several steps across whole document sets and carrying the work through to a memo, a deck or a model. The vendor calls it an operating system for complex work. Sourcing is shown for each cell of the grid so a user can check where an answer came from, and the structured format is presented as the guarantee of quality. Nothing published 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.

The Agent proposes a plan before it acts, carries out the work across documents, data sources and tools, checks whether it met the goal, and delivers work product into the project workspace for the lawyer's review and approval. It can also act without being asked, watching email, documents and projects and starting work when something needs attention. Tabular Review adds Mark as Reviewed, Lock Cells and Review Mode for team review, and full audit trails record every tool call, file access and agent action. ISO 42001 covers the AI management system, with human oversight named as a principle. No threshold at which the Agent stops for a person, and no route after a wrong output, is described.

The legal page displays the marks of Latham & Watkins, Cleary Gottlieb, Ropes & Gray and Seyfarth with no statement of what any of them does with the product, no attributed quotation and no result. The quantified claims are on third party pages. OpenAI's case study on Hebbia states that law firms reduce credit agreement review time by 75 percent, saving $2,000 an hour in legal fees, and that investment bankers save 30 to 40 hours a deal. 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.

Legora's October 2026 Series E release puts use at more than 140,000 legal professionals at more than 2,400 firms and teams in over 80 markets, with more than half of the AmLaw 50 as customers and in house teams above 40 percent of new customers in the third quarter. Named rollouts carry dates. Jones Day took Legora firm wide across more than 2,500 lawyers in 40 offices (October 2026), Hogan Lovells Cadwalader began a pilot with more than 1,300 lawyers in 18 practice teams (October 2026), and Stoel Rives adopted it for about 350 attorneys (October 2026). Linklaters reports that 97 percent of its weekly active users worked in the Agent within a month, and that monthly active use among partners rose from 65 percent to almost 80 percent. Other named customers include White & Case, Crowell, Accor, Dentons Europe, BMW Group, Salesforce and Veolia. The customers page reports that 25 percent of users see better quality work, that experienced users save 80 percent of non billable hours weekly, and that firms bill $6.4 million more per 100 lawyers a year, without a date, sample or method for those figures.

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). 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 DPA names OpenAI, Anthropic, Google and Microsoft as processing user prompts and files. No published document addresses segregation between users, matters or clients inside a customer tenant, for a product sold to four named AmLaw firms. Privilege and work product are not addressed. The agreement also permits improving machine learning models with Customer Materials where those models are solely for that customer.

The general terms bar Legora from using a subscriber's confidential information to train or fine tune AI models and require the same of its subcontractors. The terms, the data processing agreement and the security measures annex are published in versions for the EU and the United States, and the annex covers least privilege access, personnel confidentiality, authorization controls and retention on customer instruction. Ethical walls and role based access are stated to keep confidential information from crossing matters or projects, on access control built on the Zanzibar authorization model, and engineers reach customer data only with the customer's written approval. Privilege and work product handling are not addressed.

The platform is sold to law firms for due diligence, capital markets drafting and matter work. The Main Services Agreement addresses use restrictions, warranties, indemnities, liability and arbitration without touching advice, competence or supervision, and 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 page describes lawyers using the product to advise better and become the trusted adviser clients rely on, with no statement that the output is not legal advice. The audience is professional.

Lawyers, firms and in house teams are the buyers, and the product is described as working with lawyers rather than replacing them, with the Agent's output delivered for a lawyer's review and approval. Legora's Acceptable Use Policy, dated 4 May 2026, and clause 6.1.2 of its EU general terms state that Legora is not a law firm and does not practise law, and the Global terms warrant that its agentic tools include human in the loop review checkpoints. No treatment of competence and supervision duties and no jurisdiction limit appears in the policy, the terms or the product pages, though Legora publishes a list of supported countries.

The legal 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. The certificate, its scope and its date are not published. No governance policy is published, nor any statement of who inside Hebbia is accountable for model behavior, what is tested before an agent ships, or any finding about uneven output across matter types or populations. A trust center runs on Vanta.

Legora holds ISO 42001 for 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 as the operating principles. No internal owner of AI governance is named, no results of testing on model behavior before release are published, and nothing covers uneven output across matter types, parties or populations.

Annex II of the Data Processing Agreement is a full table of technical and organizational measures. It lists AES-256 encryption at rest through keys managed by AWS, 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. Customers may request deletion at any time and delete through self service, all personal data is deleted following termination, and a subject access form is published for erasure and portability. 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 notifying any new one ten days before it touches personal data, with an objection right. The incident clause requires Hebbia to inform the customer without undue delay rather than within a fixed number of hours.

The security measures annex sets least privilege and role based access, and centrally stored logs traceable to unique usernames, with security logs kept at least 12 months. It adds data integrity signing, background checks and confidentiality agreements for staff, and retention as the customer instructs. Subscriber environments are logically separated and full production backups run every four hours. The data processing agreement has Legora report a personal data breach within 36 hours and delete or return personal data at the subscriber's choice when it ends. The subprocessor list, last updated 6 July 2026, names each provider with its purpose, data and residency, and a new subprocessor comes with 30 days to object. Penetration tests cover the full platform twice a year.

Section 8.2 gives a defense and indemnity against third party claims that Hebbia's technology infringes a US patent, copyright or trade secret, with three named exclusions and the option to procure, replace, or terminate and refund. 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 the indemnity sits inside the cap. Section 9.1 excludes indirect loss and data inaccuracy for both parties, and the Excluded Claims in 9.3 lift 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 service level agreement commits to 99.9 percent monthly availability, with a service credit of one day's fees for each hour of downtime, capped at one week of fees a month and claimable within 24 hours. No insurance is stated.

General terms and conditions are published openly in an EU version, last updated May 2026, a version for the United States and Global terms dated September 2026. Clause 6.3.2 of the EU terms caps liability at the greater of twelve months of fees or €250,000, and clause 6.3.3 raises it for Enhanced Claims, breaches of confidentiality or the data processing agreement, to the greater of twice the fees or €500,000. The US version has the same structure in dollars, and the Global terms cap at twelve months of fees and twice the fees with no floor. Clause 5.1.1 gives an intellectual property indemnity only, clause 6.2.6 gives no warranty on output, and the Global terms exclude liability for beta features. No indemnity for third party claims arising from output or insurance position is stated.

The Data Processing Agreement names Merge API as an opt in subprocessor that applies only to customers who configure a Salesforce or Outlook connector. It names Google LLC as providing email and document management capabilities through Google Workspace Enterprise alongside its model role. The Matrix 2.0 announcement adds Daloopa, a financial data feed, as a connected source. No description of direction, sync behavior or configuration is published. No document management system, matter management, electronic billing or practice management integration is named on the legal, security or Matrix pages or in the agreements, so a firm's documents reach the platform by upload or through a connector set built for finance.

Document management systems are integrated rather than replaced. Legora reads access rights from the source system each time a user queries rather than caching them at setup, and Intapp Walls for AI and iManage Security Policy Manager enforce matter restrictions. The October 2026 ecosystem release covers reading from and saving back to iManage, NetDocuments, Box and SharePoint, and names Datasite, Intralinks, DiliTrust, Everlaw, Docusign, Ironclad, Agiloft, Slack, Salesforce and LawVu among connected systems. Laurel, Intapp Time and PointOne turn work done in Legora into time entries by client and matter. Add ins bring the Agent and playbooks into Word and Outlook, EDGAR sits among the research sources, and MCP connectors link Gemini Enterprise, Blue J and ChatGPT Enterprise. Partnerships with Sirion, Conga, Avvoka and Ansarada were announced in the second half of 2026, several rolling out in phases or in beta, and an iManage MCP server was announced as planned. Field level mappings for each connector are not documented.

Annex III of the Data Processing Agreement gives a processing location for every subprocessor. Amazon Web Services is placed in the named regions us-east-1, us-west-1, eu-west-1 and eu-central-1, and each model provider and inference host is placed in all US locations or in EU regional processing, with document parsing and telemetry regional by tenant. That separates where processing happens from where data is stored, provider by provider. 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 authorized subprocessors keep facilities, subject to law. No single tenant, private cloud or on premise option is described, no tenancy separation is stated beyond encryption and logical measures, and nothing states what a customer can select.

Legora runs separate arrangements for the EU, the United States and Asia Pacific. Its subprocessor list sets out hosting and AI model providers by region, Microsoft, AWS and Google among them, with EU customers' data held in the EU and EEA. For Asia Pacific it separates hosting in Australia from AI model processing in the EU and Asia Pacific, with OpenAI processing in the EU. The security page says enterprise customers control where their data is stored. Storage is tiered to data sensitivity, technical teams in the EU and the United States support local clients, and subscriber environments are logically separated. The regions open to each package and the processing location for every feature are not set out in one place.

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 pages display SOC 2 Type II and ISO/IEC 42001:2023 marks. The DPA gives an audit right: the customer or its appointed auditors may audit compliance once in any twelve months, and more often after a security incident, and Hebbia must make relevant information, policies, records and staff available. No auditor, audit period, report date or scope statement is published. A trust center runs at trust.hebbia.ai, hosted on Vanta.

ISO 27001:2022 with a yearly audit, SOC 2 Type II covering the security criteria and ISO 42001 for the AI management system are all current. The trust center at security.legora.com carries a resources section and a data flow diagram, and the security whitepapers and security measures annex are published without a gate. The security page says penetration tests run twice a year under an assume breach method, while the security measures annex of January 2025 describes them as annual. Customers can bring their own encryption keys, set retention periods and have data deleted when the contract ends. The certifying auditor is not named.

Annex III of the published Data Processing Agreement names OpenAI, Anthropic, Google and Microsoft as providing large language model capabilities through their APIs. For each it states that the subject matter processed is user submitted prompts and files, and gives the processing location as all US locations or EU regional processing. Cerebras, Groq, Baseten, Fireworks AI and Modal Labs are each named as hosting part of Hebbia's own inference infrastructure, with Elasticsearch for search and indexing over prompts, files and generated artifacts, MongoDB for embedding and reranking, and Reducto for document parsing. Change notification is contractual: any new subprocessor is listed and notified ten days before it accesses personal data, with an objection right. No specific model or version is identified by the vendor. OpenAI's own case study names o3-mini, o1 and GPT-4o inside Matrix, and Anthropic's names Claude, but those are third party statements.

The subprocessor list, dated 6 July 2026, names Microsoft, AWS and Google for hosting, infrastructure and AI models and OpenAI for AI models, with DeepL for optional translation, Linkup Labs and Exa Labs for web results, turbopuffer for search infrastructure and Intercom for support, each with its residency. New subprocessors are announced by updating the list, with 30 days for the subscriber to object. An orchestration layer handles model selection across tasks. No model or model version is named, and nothing states which provider serves which feature.

Hebbia publishes no figure, range, plan or unit, and prices are quoted after a demo. The agreement supplies the mechanics. Fees are set in an Order Form and payable in US dollars unless the Order says otherwise. The license 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, and Hebbia may notify different renewal pricing at least 45 days in advance, so renewal increases are set by the vendor within a fixed notice period. Payment is due within 30 days, late amounts carry 1.5 percent a month, and invoices must be disputed within 60 days.

No pricing page or rate is published, and the commercial route is a demo request. The Global terms, dated September 2026, set the charging structure: a Platform Fee plus Credits, with AI features consuming Credits (clauses 1.3 and 2.1.2), and 45 days' notice of a Platform Fee change at renewal (clause 3.1.3). The Credit terms they refer to are not linked from the site. The general terms are published openly, so the structure can be read before a sales conversation even though the figures cannot.

The legal page names six areas of legal work: 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 workflows tailored to a firm. The wider platform serves investing, banking and corporate finance, so the product spans transactional and financial work rather than litigation, and no dispute, court or regulatory practice area is claimed. One limit is contractual. MSA 2.4 with definition 12.13 prohibits the customer from submitting Restricted Information without Hebbia's prior written approval. Restricted Information covers sensitive personal identifiers, protected health information and personal data as defined in the GDPR, and DPA Annex I separately prohibits special categories, so European matters, employment work or anything touching health records need that approval first. Firm segments are not broken down, and government and in house use are not addressed.

Large law firms, small and mid sized firms, in house legal departments, the public sector and professional services are all served, with solution pages for M&A, litigation, banking, insurance, private equity and tax. Multi jurisdiction and cross border work are stated strengths, with research sources across Europe and federal and state law in the United States. No firm size or practice area is named as outside its scope.