DeepJudge vs Fileread: how they compare in 2026

No content or conventional product sits underneath DeepJudge's AI. It licenses and publishes nothing of its own. The material it works on belongs to the customer and stays in the customer's systems. What DeepJudge supplies is the retrieval: search by intent that understands content, context and relevance, automatic sorting into a taxonomy, and detection of near duplicates and redlines across what it describes as billions of data points. Without the models, what remains is connectors and a permissions sync, plumbing a firm would not buy on its own. The founders say search, rather than models, is their wedge. That is a claim about which machine learning matters, not that machine learning is peripheral.

The AI is the whole product. Fileread reads a document production and answers questions about it in plain language rather than by keyword; the vendor's own phrase is asking the question you cannot turn into a keyword. The deliverables are generated from the record: fact memos, chronologies, contradiction reports, character lists, org charts, deposition outlines and factual briefs, with Workbench running longer research chains to assemble them. Retrieval covers formats that keyword tools miss, including images, handwritten notes, audio, video and spreadsheets. The vendor says the design is built to be checked, with every answer pinned to its place in the source. Without the models nothing would remain: there is no case law database, docket service or document management system behind it, only the customer's own production, which the customer already holds.

Grounding is built into the product, but no accuracy is measured. Every result is a document the firm already holds, in its own system, which the user can open and read in full, so checking is direct. The retrieval method is described: search by intent, automatic taxonomy sorting, detection of near duplicates and redlines, and linking of client, matter, document and people data. Section 4(j) of the subscription terms requires the customer to verify the accuracy, reliability and fitness of AI features, including through human review. No accuracy figure, evaluation, test set or benchmark appears anywhere. Nothing addresses hallucination in the workflow and agent layer, which is where new text, rather than retrieved documents, is produced. The one nearby figure is satisfaction: 90 percent of users say they find results faster than with existing tools, which is not an accuracy claim.

Every answer is tied to its source, and no accuracy figure is published. Each answer is pinned to the exact place in the source document, so a claim traces to its page, and generated work product keeps its link to the underlying material so a reviewer can see where a finding came from. The vendor says the system is built to be checked rather than trusted and keeps a paper on source citation and verification. It also publishes a dated, authored, practitioner reviewed list of seven ways legal AI fails, with a detection control for several: unsupported generation, correct source with wrong interpretation, context loss, missed evidence, conflicting evidence presented as certainty, identity confusion, and date and sequence errors. Contradiction reports address the fifth by surfacing conflicts. The list is guidance about legal AI in general, though, not a disclosure about this system. No accuracy figure, error rate, test set, benchmark or evaluation result is published for Fileread itself. The paper shows the vendor knows how such systems fail, not how often this one does.

Human review is a contract requirement, and the excluded uses are named. Section 4(j) requires the customer to verify the accuracy, reliability or fitness of AI features through human review. The same clause bars use in judicial decision making by judges, in the work of lawyers serving as arbitrators, mediators or other dispute resolution neutrals, and in any use restricted under the EU Artificial Intelligence Act. Naming who may not use the product is more useful than the usual disclaimer. What is not published is the operating boundary. The workflow layer is described as agents that carry out complex tasks in several steps, retrieving from many sources and taking actions. The customer builds those workflows itself with a low code builder, so the point at which a workflow acts without a person is set by the firm and described nowhere. Governance of agents is offered as a feature with no account of the controls.

The limits on autonomy are stated and backed by a published oversight framework, but the product itself enforces none of it. Workbench runs longer research chains and assembles chronologies, factual briefs, contradiction reports and deposition outlines, and the vendor says the result stays a draft for professional review and refinement. So the system runs multistep work on its own, and the vendor says where it stops. Around that sits an oversight framework that is authored, dated and practitioner reviewed. It says oversight starts before the prompt, by settling the task, the materials included and excluded, the audience, the acceptable uncertainty, the claims that need source support, the person responsible for final review, and the record to keep. Review scales with risk, and a twelve point checklist runs from whether cited passages support the statements to whether a responsible professional approved the final use. The framework is guidance for the firm, not a description of the product's own controls. Only one paragraph describes what Fileread itself does, no configurable autonomy level or override is described, and nothing says the checklist is recorded or enforced in the product.

DeepJudge names many customers and publishes figures, but most figures are not tied to a named customer. Sixteen firms are named on the home page. They include global firms such as Freshfields, Greenberg Traurig, Holland and Knight and Gunderson Dettmer, and US mid market firms such as Cozen O'Connor, ArentFox Schiff and Vorys. Swiss and Austrian firms include Homburger, Lenz and Staehelin, Wenger Vieli, Advestra, Schoenherr and CMS Switzerland, plus an in house team at SBB. Three named people speak on the record with title and firm, including two chief innovation officers. Published figures: 65 hours saved per user per year on searching, 85 percent adoption within two months, 23 queries per user per day, a fourfold return in the first year, and 90 percent reporting faster results. One figure is tied to a customer: a named M&A partner at Homburger says more than 80 percent of the firm's legal professionals use it in their work. The headline figures are aggregate, with no method, and no dates appear on the home page. The individual customer story pages were not opened.

Customers are named, with people and roles, and one account has numbers, though no method sits behind any of it. Five attributed testimonials are published. A partner at Kellogg, Hansen, Todd, Figel & Frederick describes use on an expedited case that went to trial. A partner at Fields, Han & Cunniff describes the system finding the right document from an inaccurate description. A senior vice president for innovation and strategy at JND Legal Administration assesses it on maturity, defensibility and user experience against alternatives. A manager of knowledge and practice support at Nutter McClennen & Fish speaks to the Relativity integration, and the chief technology officer of Cimplifi to replacing Boolean search. One account shows the arithmetic: 4,000 text message screenshots converted and deduplicated to 725 actual messages, work the vendor says would have taken more than 20 hours of attorney time. Independent coverage is cited, including a Law360 piece on a firm's generative AI experiments and reports of a 6 million dollar seed round. No aggregate figures, retrieval recall or precision, or customer count is published, and the time saving rests on one example.

Most confidentiality needs are met by contract and by design, but privilege is never addressed. Section 5(b) of the subscription terms says DeepJudge does not use customer data to train the AI or machine learning models behind the services, a contractual ban that names training. Separation is the product's design: permissions sync continuously with the firm's source systems so existing ethical walls hold at matter level. Section 5(e) adds that a customer's own workflow logic will not be disclosed to or used for any other customer. Customer data is deleted within sixty days of termination. The firm chooses the models, and the hybrid deployment clause says exactly when processing leaves the firm's own infrastructure. Privilege and work product are not mentioned anywhere, and ethical walls address conflicts, not privilege.

Real controls are published, and nothing addresses privilege, though the product takes in an entire litigation production. Published: a statement that nothing a customer uploads trains a shared model; data encrypted, hosted in Azure and processed through what the vendor calls private large language models; and trust center controls for least privilege, access monitoring, access log management and data classification, with HIPAA compliance bearing on the sensitive material productions contain. These are company and infrastructure controls. Nothing published covers privilege or work product, who at the vendor may read customer matter material and when, or segregation between matters or customers. The Relativity integration sharpens this, since nothing says whether the platform follows the workspace permissions a firm already built there or keeps a second permission model. The Master Services Agreement, where a confidentiality clause would usually sit, is behind an access request. The vendor's own oversight paper names protecting privileged material as an attorney duty, which acknowledges the risk without committing the product to anything.

DeepJudge sets out its professional responsibility position in the operative agreement. Section 4(j) says DeepJudge is not a law firm and does not practice law, and the customer accepts it is not relying on the services for legal advice. It says who may not use the product: judges in judicial decision making, and lawyers serving as arbitrators, mediators or other dispute resolution neutrals. It addresses competence and supervision through a positive duty, requiring the customer to verify the accuracy, reliability or fitness of AI features, including through human review. The legal boundary is stated twice. Section 4(a) limits use to what complies with the duties that apply where the customer and its users are located. Section 4(j) excludes any use prohibited or restricted under the EU Artificial Intelligence Act. The product searches the firm's own documents, not any country's law, so jurisdictional reach of legal content does not apply. No bar or law society guidance is engaged.

The vendor writes in depth about where professional responsibility sits, but publishes no disclaimer and names no authority. Some of the usual concerns are remote here: the product analyzes the customer's own production rather than advising on law, its outputs are factual findings cited to the record, and it is sold to lawyers and litigation support staff, not the public. On what does apply, the vendor's own oversight paper is direct. It says AI can assist with parts of litigation work but does not hold responsibility for the representation. It says attorneys must interpret contested facts, assess credibility, understand procedural posture, protect privileged material and apply legal standards, and that a fluent response may hide uncertainty, missing context or a wrong assumption. It separates an exploratory internal question from a factual statement meant for a client, a witness examination or a court filing. Its checklist ends with review under applicable legal and professional standards and approval by a responsible professional. No advice disclaimer is published, and no bar rule, ethics opinion or court standing order is named.

One binding governance commitment exists, with nothing around it. Section 4(j) of the subscription terms makes compliance with the EU Artificial Intelligence Act a contractual limit on use, naming the law and excluding the uses in the administration of justice the Act treats as high risk. That is more than a principles page, and the customer can enforce it. The workflow layer is also sold partly on governance, offering to build, deploy, orchestrate and govern AI agents. No content sits behind either. There is no responsible AI page, governance framework, accountable person or team, account of testing before release, or certification such as ISO 42001. Nothing addresses uneven output across matter types, practices or populations. ISO 27001 and SOC 2 Type II cover security, a different subject.

An AI governance program is disclosed by name, and its contents sit behind an access request. The trust center's policy library lists an AI System Development and Evaluation Policy and an Artificial Intelligence Management System Plan, the latter using the term ISO/IEC 42001 defines. A separate AI section has overview, security and monitoring entries, so AI is treated as its own control area, and the trust center is described as continuously monitored with automated compliance monitoring and alert response. Both policies sit behind the access request, so their titles are public and their contents are not, and nothing is attested. No ISO/IEC 42001 certification is claimed, unlike ISO/IEC 27001:2022, which is claimed with its version. No evaluation result, benchmark or model card is published, and nobody is named as accountable for the system's behavior. No bias or fairness disclosure exists, which matters because the vendor's own published failure modes include identity confusion, combining information about people with similar names, addresses or roles.

DeepJudge publishes all five elements in specific terms, in its subscription terms and data processing addendum version 1.1 of May 2026. Retention and deletion: section 3(a) deletes all customer data, including training data and customer models, within sixty days of termination. DPA clause 11 adds prompt destruction or return on request, including data held by subprocessors, with confirmation. Access control: permissions sync continuously from source systems on least privilege principles, access is limited to authenticated users, audit logging is configurable, and the firm controls what is crawled at all. DPA Annex 2 adds multifactor authentication, role based access control, central event logging and separate test and production environments. Subprocessors: Annex 3 names five entities with purpose and location. Clause 5 requires equivalent written terms, due diligence, DeepJudge's liability for their acts, thirty days' notice before a new subprocessor processes anything, and an objection right with termination without liability. Incidents: clause 7 commits to notice without undue delay and within 72 hours of discovery, covering breaches at subprocessors too. The definition at 1.13 expressly includes unauthorized internal access, and DeepJudge must help the customer meet its own regulator and data subject duties. Encryption is TLS 1.2 or later in transit and AES-256 at rest, with annual penetration testing and SOC 2 Type II audited controls. Clause 3.4 bars subprocessors from selling or sharing data, processing it for other purposes, keeping or disclosing it outside the direct business relationship, or combining it with data from other sources.

The published control list is broad, with one gap. Certifications are named with the standard version: SOC 2 Type 2, ISO/IEC 27001:2022, HIPAA, GDPR and CCPA. Training is addressed: nothing a customer uploads trains a shared model, and data is encrypted, hosted in Azure and processed through private large language models. Subprocessors are named on the public page: OpenAI, Baseten, Pinecone, Langfuse and Microsoft Azure, with Amazon Web Services and Azure listed as infrastructure. A subscribable feed publishes dated changes to that list, recording the addition of Langfuse for prompt use and of OpenAI. Incident practice is documented, with a reporting process and designated response staff. Access control covers least privilege, access monitoring, access log management and data classification. Around those sit data loss prevention, penetration testing, code analysis, credential management, backups, business continuity and disaster recovery, change management, endpoint and network controls, and security, privacy and phishing training. The gap is retention and deletion. No retention period, deletion commitment or disposal position at the end of a matter or engagement is published, which matters for an uploaded litigation production.

Loss is allocated specifically and mutually. Section 7(a)(ii) caps each party's total liability at the fees paid in the twelve months before the event. Four exceptions sit outside the cap: the customer's payment duties, breach of the use restrictions at section 4(e), fraud or willful misconduct, and intellectual property infringement. Section 7(a)(i) excludes indirect and consequential damages for both parties, with the same exceptions. Section 7(b) gives the customer a real indemnity: DeepJudge defends third party claims that the services or software infringe intellectual property. The remedies are set in order: secure the right to continue, modify or replace the software, or terminate the affected part. There is also a service commitment. Section 1(a)(iii) sets a 99 percent quarterly uptime target, with downtime defined and six named exclusions. No warranty covers output, since section 6 disclaims accuracy and completeness, and no insurance is mentioned.

No published terms say who bears the loss when the product is wrong. The agreements exist: a Master Services Agreement and a Data Processing Agreement are listed by name in the trust center's legal section, but both sit behind an access request. That is a choice about how the terms are shared, not an absence of terms. No indemnity, liability cap, warranty, service level commitment, exclusive remedy or insurance position is published. The product's outputs become fact statements and pleadings a lawyer signs. The vendor's own failure taxonomy describes seven ways such outputs go wrong, and nothing published says what the vendor stands behind if they do.

For DeepJudge, integration is the product, and the depth is documented. It connects to the systems firm work lives in: the document management system, email, SharePoint, OneDrive, HighQ, intranets, and experience and metadata systems. Data flow is described both ways. Documents are crawled and indexed in place, with no migration or upload, so the data stays in its source system. Permissions flow the other way, synced continuously from those systems into the index, and the firm decides what is crawled at all. Setup duties are in the contract. Section 1(d) requires continued access to the customer databases named on the order form and a named technical contact, and warns that withholding access degrades the service. The on premises clause makes the customer responsible for stated hardware and configuration requirements. The platform also gives governed access to other AI systems through the Model Context Protocol and publishes an open specification for handing off between agents.

There is one integration, into the system this kind of product runs on, and it is documented on that platform's own marketplace. Fileread is a native integration inside Relativity with a listing on the Relativity app hub, and a customer, a practice support manager at Nutter McClennen & Fish, credits it with easier adoption. Relativity is where litigation productions already sit, and the design follows from that: the platform works on productions a team already has, with nothing to migrate. A second distribution channel exists on the Microsoft commercial marketplace, and Fileread's trust center includes integrations among its product security controls. Cimplifi, an eDiscovery services provider, offers the product through its own ecosystem, which is a sales channel rather than an integration. No API or developer documentation is published. No document management system, case management platform, word processor or electronic signature product is named, there is no integrations page, and nothing describes what moves between Fileread and Relativity or how permissions carry across.

Three deployment models are defined in the agreement, with where data flows spelled out for each. Cloud deployment hosts the services in a location named on the order form. On premises deployment runs inside the customer's own infrastructure, with DeepJudge given access to install and maintain it and the customer meeting stated system requirements. Hybrid deployment is on premises with some processing, expressly including model processing, outside the customer's infrastructure. The clause says customer data may then be processed in the cloud hosting location and that model processing may move customer data outside the customer's systems. That separates processing from storage, stated by DeepJudge rather than inferred, and names the one case where an on premises customer's data leaves. The security page adds that the customer runs its choice of model in its choice of residency. No regions are listed: residency is chosen and set on the order form rather than published.

Neither the tenancy model nor the region is stated; what is published is infrastructure. Data is encrypted and hosted in Azure and processed through private large language models, and Fileread's trust center lists Amazon Web Services and Microsoft Azure as infrastructure and Azure again as a subprocessor. Naming the cloud says where a workload runs, not how tenants are separated or which region holds data. Nothing says whether customers are separated logically or physically, whether one matter is isolated from another, or whether single tenant options exist. The phrase private large language models points to dedicated inference, not to how the data store is shared. No country or region of storage or processing is stated, no region choice is offered, and no residency commitment appears, though GDPR compliance is claimed. A Data Flow Diagram and a Network Diagram are listed in the trust center behind the access request.

Two certifications are claimed with their scope, but no evidence a prospect can reach is offered. SOC 2 Type II is stated to cover security, availability and confidentiality, and ISO 27001 to cover the information security management system. The data processing addendum, version 1.1 of May 2026, records annual penetration testing in Annex 2 and says controls are audited for SOC 2 Type II. Missing: no auditor, certificate number, issue date, report date or coverage period is published, and there is no trust center or portal. Evidence is for customers. DPA clause 9 says certifications and audit reports or written questionnaires come first. A customer audit is allowed only where those cannot show compliance, at the customer's expense, once a year, with no access to systems, hosting sites or infrastructure. That helps a signed customer but gives a buyer still deciding nothing to read.

A real trust center with a named document list, and the reports sit behind a request whose terms are not stated. Open to anyone: five compliance positions (SOC 2 Type 2, ISO/IEC 27001:2022 with its version, HIPAA, GDPR and CCPA); a control list of about twenty categories, from continuous monitoring and product security through access control, endpoint, network and application security to incident response, risk management, business continuity, asset management, change management and training; a named subprocessor list; and a subscribable feed of trust center updates with dated subprocessor changes. Behind the request sit the SOC 2 report, an ISO 27001 report, a penetration test report, a data flow diagram, a network diagram, the policy library and the agreements. The portal offers to start a security review and to request access. It does not say whether access is granted instantly, through a self service nondisclosure click through, or by vendor approval after a sales conversation. No auditor, audit period or scope statement is named.

AI providers are named at the platform level, but no model vendor or model. Annex 3 names Microsoft Ireland Operations Ltd, Google Cloud EMEA Ltd and AWS EMEA SARL, each described as AI provider and cloud infrastructure, with location chosen by the customer. Clause 5 commits to at least thirty days' notice before any new subprocessor processes customer data, with a right to object on reasonable data protection grounds and termination without liability if the objection is not resolved. These three are cloud platforms, not model vendors. No model house, model or default is named, so a reader learns which cloud estates run the AI but not whose model produced the output. The architecture and geography are well disclosed. DeepJudge says its architecture works with any model, and the customer chooses from and may combine any commercial or open source models, in its choice of residency. The hybrid deployment clause discloses that model processing may move customer data outside the customer's own infrastructure.

Providers are disclosed in full, and the models are not named. The trust center publishes its subprocessors on the open page, and the list reads as an AI stack rather than generic infrastructure: OpenAI, Baseten, Pinecone, Langfuse and Microsoft Azure, with Amazon Web Services and Azure also listed as infrastructure. So a named model provider is in the path, inference serving and vector storage sit alongside it, and a prompt observability layer is involved. The update feed records that Langfuse was added for prompt use and that OpenAI was added as a subprocessor. Change notice is a published mechanism, not a promise: the trust center carries a dated subprocessor changes feed with a subscribe option. Two gaps remain. No model or model family is named; Fileread's site says only that data is processed through private large language models. And nothing says which parts of the platform send customer content to which provider, so OpenAI is known to be in the stack without anything saying what it sees.

The unit and the structure are published in the subscription terms without any figure. Fees sit on an order form and are stated to be in United States dollars unless otherwise agreed. The term structure is set out: subscriptions renew automatically for successive one-year periods unless either party gives three months notice before the end of a term, invoices issue within ten days of the effective date and each subsequent term and are payable within thirty days, and late payment attracts the greater of five percent per annum or the statutory maximum. Professional services are separated cleanly, provided only under a mutually agreed statement of work that becomes part of the agreement, which tells a buyer that implementation is priced apart from subscription. Most usefully, the consumption unit is published: section 4(g) sets a hosting threshold of 500,000 documents per user or 250 gibibytes per user, above which DeepJudge may seek volume reduction or charge at standard rates. What is absent is any figure, any tier, and any pricing page; the only commercial route on the site is Book Demo.

No pricing is published at any level, including the unit of charge, and the site navigation has no pricing page. The header lists Product, Solutions (three segment pages), Customers, Security and Resources, and the footer adds Workflows, Company, Privacy and Terms of Service. Every page ends at the same call to action, an invitation to see the product on your matter, which is a demo request. No tier, edition or unit (seat, matter, document, gigabyte or production) is named, and no band, minimum, term, free tier or trial appears. The Master Services Agreement, which would carry commercial terms, sits behind an access request in the trust center.

The customer list shows who buys the platform. It includes global firms such as Freshfields, Greenberg Traurig and Holland and Knight, and US mid market firms such as Cozen O'Connor, ArentFox Schiff and Vorys. It also includes technology specialist Gunderson Dettmer, Swiss and Austrian firms such as Homburger, Lenz and Staehelin, Wenger Vieli, Advestra and Schoenherr, and an in house legal and compliance team at SBB. Teams are stated to be across North America and Europe, serving clients worldwide. Practice areas are not listed, which matters less here, because the platform searches whatever the firm already holds rather than a defined body of law. What is missing is the boundary. No firm size is named, nothing sets a minimum data estate or user count, and nothing says where the product stops. The clearest published limits are the barred users at section 4(j).

The buyer, the workflow and the material are each described specifically, and no boundary is stated. Three buyer segments have their own pages: litigation teams, eDiscovery and corporate legal. Five workflows are published: early case assessment, discovery document review, evidence analysis, deposition preparation, and factual briefs and case narratives, from first production to trial. The named customers match that spread, from a Washington litigation firm and a boutique to a large firm's knowledge and practice support team, a litigation administration provider and an eDiscovery services company. Coverage of the material is concrete: text, images, handwritten notes, audio, video and spreadsheets in the same matter, illustrated by a production that arrived as 4,000 text message screenshots. No limit is stated. No jurisdiction is named or excluded, though the customers and vocabulary are United States civil litigation throughout. No matter type, case size or practice area is named as in or out of scope, and no volume ceiling is published, although scale is the product's central claim.