L

Lexbe

Under the AI sits a full eDiscovery stack. It processes more than one million documents an hour on over 10,000 processors, then indexes, deduplicates, runs OCR, groups near duplicates, threads email, redacts, Bates stamps and produces with privilege logs. Generative AI runs through three tools layered on that base. CoPilot works inside the document viewer on one document or a selected group, Pilot answers questions across a whole case, and AutoPilot codes documents for responsiveness and privilege at volume.

Older machine learning sits beside them in Bayesian predictive coding, sentiment tagging, speech transcription and image recognition. Take the language models away and the processing, review and production platform still runs, which is how customers used it for well over a decade before CoPilot arrived.

Every CoPilot result links back to the document it came from, and Pilot ends each answer with links to the case documents it drew on, which open in the document viewer. Pilot uses retrieval augmented generation over the case repository, and its Strategic Insights setting widens how much of the case it brings into an answer. AutoPilot returns a written rationale with each coding call. Lexbe describes Pilot as unable to hallucinate because grounding language limits each answer to the text sent to the model, and says AutoPilot's precision and recall beat traditional technology assisted review in internal testing. No precision, recall or error figure has been released, so neither claim can be tested from outside.

Each of the three tools is framed around a supervising attorney. CoPilot follows a human in the loop model in which reviewers keep control of coding decisions and quality checks. AutoPilot is a first pass layer that returns a confidence rating from 1 to 10 and a rationale for every document. Its recommended workflow validates output against a human coded control set drawn before the run, tests a sample of the documents it did not flag, and leaves final responsiveness and privilege calls to attorneys.

AutoPilot is not self serve, since Lexbe's professional services team engineers the prompts and helps with quality control. Pilot is limited to fact questions and gives incomplete answers to requests for lists across a whole case. The master services agreement calls AI output probabilistic and tells customers to apply human review to it. The confidence level at which a document goes to a person is chosen per matter rather than fixed by the product.

Lexbe says more than 8,795 legal professionals have relied on the platform and quotes named attorneys at named firms. A litigation partner at Logan & Lowry says it helped a three lawyer team win a verdict of nearly $9 million against Greenberg Traurig. Partners at Richman Greer, Curry & Friend, Vladeck, Raskin & Clark, The Brodsky Law Firm and Heslin, Rothenberg, Farley & Mesiti praise its pricing, speed and support.

The legal director of Disability Rights Ohio describes using it with co counsel on large document cases. An insurance defense case study covers a construction liability matter shared by five defense firms, with a Thomson Coe attorney quoted on the savings. The 70 percent cut in first level review time comes from a modeled 100,000 document example and internal testing rather than a named customer. None of the quotes is dated or carries a measured before and after figure.

Section 3 of the master services agreement treats customer data as confidential and acknowledges it may be privileged or attorney work product, which Lexbe holds as agent of the attorneys on the proceeding. Lexbe discloses it only to the customer or on the customer's instruction, subject to law or a valid order, and promptly notifies the customer of a subpoena or investigative demand. Matter separation comes from seven fixed user types, case by case access and a Limited User who sees only tagged documents.

Joint defense groups can share one repository while each defendant keeps its own work environment. Pilot runs on Anthropic's Claude with zero data retention, and Lexbe states that no user data from it trains any model. That promise sits in the help center rather than the agreement, and section 9 has customers accept the OpenAI, Anthropic and AWS terms on retention and training when they use the AI services. No retention period for prompts or case data is stated.

The agreement describes the platform as a supplemental tool for experienced lawyers and litigation professionals, not for people who cannot manage the litigation process unaided. Customers must be licensed attorneys or law firms, or have retained an attorney who advises them on proper use, and they rely on their own skill rather than on Lexbe for legal advice. Section 9 places every decision, piece of advice and action based on AI output with the customer, and tells it to apply human review to output that may be probabilistic.

Marketing pitches Pilot as a digital lead investigator for pressure testing case themes, yet every tool is sold to litigation teams rather than to the public. No ethics opinion or jurisdiction limit is named.

Lexbe refers to a Responsible AI framework that keeps the attorney in control, with output tied to source documents and a person making the final call. It also describes how AutoPilot was built. Its professional services team and an AI research and development team developed and tested a prompt engineering method that, by Lexbe's account, exceeds the precision and recall benchmark for older technology assisted review.

Customers get guidance on validating output with control sets and tests on unflagged documents. The framework itself has not been released as a document, and no one inside Lexbe is named as accountable for AI. The results behind the benchmark claim are not shared, and nothing addresses uneven output across languages, custodians or document types.

Case data sits on encrypted servers in the United States on Amazon Web Services, with AES-256 encryption at rest and in transit. Access runs through seven fixed user types, optional two factor authentication and single sign on, added in February 2026. Administrators can export a log of who viewed or downloaded each document, kept for two months. Account administrators delete cases themselves, and recovery from backup may be possible for a fee, depending on the backup delete cycle.

At termination Lexbe may delete customer data, and a paid case archive is available if requested beforehand. The AI integrations run with zero data retention at the model providers. The privacy policy, last uploaded in September 2021, is a general one, and no retention period during the subscription, breach notice practice or subprocessor list beyond the named AI providers is offered.

Lexbe indemnifies customers against claims that the platform infringes a copyright, a United States patent or a trademark. It also indemnifies against claims arising from a breach of its section 17 warranties, which cover its authority to contract and commercially reasonable efforts to keep the platform available. A 99.9 percent monthly availability commitment carries pro rata credits as the sole remedy. Everything else is disclaimed.

The service is provided as is, and AI output is used at the customer's sole risk. Lexbe takes no liability for failures in the third party libraries and services it relies on, indirect and consequential damages are excluded, and claims must be brought within two years. No liability cap figure or insurance is stated, and nothing covers a wrong coding call or summary.

The Clio integration pulls matter folders from Clio into a case. A user signs in to Clio once through the Third Party Applications page and picks a matter or folders within it. Up to 3,000 documents upload at a time, and Clio folder paths are kept in the Source FilePath metadata field. Single files cannot be picked, and the link is removed from the same page. CoPilot can also be tied to a customer's own OpenAI account by pasting an API key and choosing a model.

Productions received from other parties load with their load files through a .dat metadata mapping step, and outgoing productions carry load files for other review platforms. No connector to iManage, NetDocuments, Microsoft 365, Outlook or a matter management system is offered, so Clio is the only practice system in the list.

The platform is cloud only and runs on Amazon Web Services, reached through a browser on Windows, Mac or a mobile device with nothing to install. All customer electronically stored information is kept on servers in the United States, and customers in the European Union and Switzerland are covered by Standard Contractual Clauses provided on request. Where the OpenAI, Anthropic and Amazon models process documents is not stated, and no tenancy model, region choice or private deployment is offered.

No independent security attestation is claimed for Lexbe. Its security is described as AES-256 encryption at rest and in transit, data access controls and active management of the AWS environment, with optional two factor authentication and single sign on. No SOC 2 report, ISO 27001 certificate, trust center or penetration test summary is offered.

The agreement names OpenAI, Anthropic and Amazon Web Services as the providers behind Lexbe's AI services, with Amazon Transcribe, Amazon Translate and Amazon Rekognition behind audio transcription, translation and image recognition. Release notes name the language models. GPT-4.1 and GPT-4.1 Mini were added to CoPilot in September 2025, Claude 4 Sonnet and Claude 4 Opus to Pilot and AutoPilot in June 2025, and five more OpenAI models to AutoPilot in April 2025.

Pilot runs on Claude with zero data retention. Customers learn of model changes through release notes after the fact, with no advance notice commitment, and where each model runs is not stated.

Lexbe prices by data volume, the gigabytes uploaded and hosted, billed monthly in arrears on the highest usage in the month. Plans include pay as you go accounts with no commitment, flat monthly rate plans, subscriptions of a year or more for a set capacity and hibernation for inactive cases, with volume and subscription discounts. There are no user or case fees, ESI processing is free on subscription accounts, users are unlimited, and CoPilot and Pilot come with every hosting plan.

The billing policy spells out what counts toward storage, including the text index, capped at 35 percent of document size for billing. AutoPilot, professional services and forensics are billed separately at per gigabyte, per page, per document or hourly rates. No rate is published, and every plan routes to an estimate request.

Lexbe is built for boutique and mid sized law firms with document heavy cases, and also sells to corporate legal departments, managed review providers, disability rights organizations and Mac users. Practice pages cover construction, environmental, insurance defense, intellectual property, labor and employment, securities and telecommunications litigation, and the platform handles third party subpoenas, internal investigations and compliance reviews.

Forensic collection covers phones, computers, email, cloud and collaboration accounts such as Slack and Microsoft Teams, social media, websites and drones. A few limits are stated. Use is meant for experienced litigation professionals working with a licensed attorney. Native upload batches should stay near 20 GB with mail files under 50 GB, and Pilot is not built for whole case list requests. No practice area, jurisdiction or firm size is excluded.