F

Fileread

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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 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.