F

Fileread

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 product is the model work and nothing survives its removal. Fileread exists to read a document production and answer questions about it: users query the full collection in plain language rather than by keyword, and the vendor's own framing is that the point is asking the question you cannot turn into a keyword. Generation is the deliverable, not a feature bolted onto a repository: fact memos, chronologies, contradiction reports, character lists, org charts, deposition outlines and factual briefs are produced from the record, and a module called Workbench runs longer chains of research to assemble them. Retrieval spans formats that defeat keyword tooling, including images, handwritten notes, audio, video and spreadsheets, which is model work rather than indexing. The vendor's positioning sentence is an AI claim about its own architecture, that most legal AI is built to generate and this one is built to be checked, with every answer pinned to the location in the source. Strip the models out and there is no residual product: no case law database, no docket service, no document management system, only a customer's own production which the customer already holds. Homepage, product framing and the vendor's human oversight paper read 7 September 2026.

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.

Grounding is real, documented and architecturally central, and no accuracy figure exists. What is published: every answer returns pinned to the exact location in the source document, a claim traces to its page so checking takes a glance, and generated work product retains its connection to the underlying material so a reviewer can inspect where a finding came from. The vendor states the design intent plainly, that the system is built to be checked rather than to be trusted, and maintains a dedicated paper on source citation and verification. It also publishes an authored, dated and practitioner-reviewed taxonomy of seven ways legal AI fails, with a detection control for several of them: unsupported generation, correct source with incorrect interpretation, context loss, missed evidence, conflicting evidence presented as certainty, identity confusion, and date and sequence errors. That is an unusually candid document and one of the failure modes has a product answer, since contradiction reports exist to surface the conflicts that the fifth mode describes being flattened. What keeps this off the top grade is the distinction between naming the failure modes of a class and measuring your own. The taxonomy is written as guidance about legal AI generally, not as disclosure about this system: no accuracy figure, no error rate, no test set, no benchmark and no evaluation result is published for Fileread itself. A buyer can see that the vendor understands how such systems fail and cannot see how often this one does. Homepage, failure modes paper of 17 August 2026 and human oversight paper read in full 7 September 2026.

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.

Substantive published commitments with the autonomy boundary stated, short of anything the product enforces. The boundary is stated, which is what keeps this off the middle grade: Workbench performs longer chains of research and assembles chronologies, factual briefs, contradiction reports and deposition outlines, and the vendor states that the resulting work remains a draft for professional review and refinement. So the reader is told the system runs multi-step work unattended and is told where it stops. Around that sits a published oversight framework, authored, dated and practitioner-reviewed, which is more than most records on this axis offer: oversight is framed as beginning before the prompt, with the task, included and excluded materials, intended audience, acceptable level of uncertainty, claims requiring source support, the person responsible for final review and the record to be retained all settled in advance; review is scaled to risk, with exploratory output to be treated as a starting point, citations sampled, and results not represented as established fact; and a twelve-point review checklist runs from whether cited passages actually support the statements to whether a responsible professional approved the final use. What is absent is the product half. The framework is guidance offered to the buyer rather than a specification of the system's own controls, only one paragraph describes what Fileread itself does, no configurable autonomy level or override is described, and nothing states that the checklist is recorded or enforced anywhere in the product. Human oversight paper and homepage read in full 7 September 2026.

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.

Named customers with named people in named roles, and one quantified account, with no method behind any of it. Five attributed testimonials are published: a partner at Kellogg, Hansen, Todd, Figel & Frederick describing use on an expedited case that went to trial; a partner at Fields, Han & Cunniff describing the system retrieving a correct document from an inaccurate description; a senior vice president for innovation and strategy at JND Legal Administration assessing the product on maturity, defensibility and user experience against alternatives; a manager of knowledge and practice support services at Nutter McClennen & Fish on the Relativity integration; and the chief technology officer of Cimplifi on replacing Boolean search. That is a spread across a litigation boutique, a large firm practice-support function, a litigation-support provider and an eDiscovery services company, which tells a buyer who actually runs it. One quantified account is published with the arithmetic shown: a matter arrived as 4,000 text message screenshots, which the platform converted and deduplicated to 725 actual messages, work the vendor states would have consumed more than 20 hours of attorney time. Independent coverage is cited, including a Law360 piece on a firm's generative AI experiments and reporting of a 6 million dollar seed round. What is absent is measurement: no aggregate figures, no recall or precision on retrieval, no customer count, and the time saved rests on a single anecdote rather than a study. Homepage, customer testimonials and cited coverage read 7 September 2026.

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.

Real controls are published and nothing addresses privilege, which matters more here than on most records because the material this product ingests is an entire litigation production. What exists: a statement that nothing a customer uploads trains a shared model; data encrypted and hosted in Azure and processed through what the vendor calls private large language models; and a trust centre control inventory naming least privilege, access monitoring, access log management and data asset classification, alongside HIPAA compliance which bears on the sensitive material productions contain. Those are corporate and infrastructure controls, and they are credited on the stewardship axis rather than counted twice here. What is absent is everything this axis asks. No privilege or work product treatment is published at any level. Nothing states who at the vendor may read customer matter material or under what conditions. No segregation position is published between matters held by the same customer or between customers, and the question is sharpened by the Relativity integration, since nothing states whether the platform inherits the workspace permissions a firm has already built there or maintains a second permission model of its own. The customer agreement that would carry a confidentiality clause is not readable: a Master Services Agreement is listed in the trust centre behind an access request, and the website terms render client-side and returned no body on this channel. The vendor's own oversight paper names protecting privileged material as an attorney duty, which is an acknowledgement of the risk rather than a commitment about the product. Homepage, security page and trust centre read 7 September 2026.

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.

Substantive published engagement with where professional responsibility sits, short of a located disclaimer or a named authority. Several limbs of this band do not bite and are named rather than penalised: the product analyses the customer's own document production rather than advising on law, its outputs are factual findings cited to the record rather than legal conclusions citing authority, and it is sold to lawyers and litigation support professionals rather than to the public, so the unauthorised practice risk this axis was written for is structurally remote. On what does bite, the vendor engages directly and in its own authored material. Its oversight paper states that artificial intelligence can assist with parts of litigation work but does not possess responsibility for the representation; that attorneys must interpret contested facts, assess credibility, understand procedural posture, protect privileged material and apply legal standards; that a fluent response may obscure uncertainty, missing context or an incorrect assumption; and that an exploratory question for internal orientation does not require the same process as a factual representation intended for a client, a witness examination or a court filing. Its review checklist ends with the output being reviewed under applicable legal and professional standards and a responsible professional approving the final use. What is missing is the formal half: no disclaimer language was located, because the terms of service render client-side and returned no body, and no bar rule, ethics opinion or court standing order is named anywhere. Human oversight paper and failure modes paper read in full 7 September 2026.

AI Governance and Bias Disclosure

Published governance over model behaviour: who owns it inside the vendor, what is tested before release, and what is disclosed about disparate output across matter types, parties, or populations.

An AI-specific governance apparatus is disclosed by name, and its contents are gated. The trust centre lists two instruments in its policy library that most records in this corpus do not have at all: an AI System Development and Evaluation Policy, and an Artificial Intelligence Management System Plan, the latter using the term of art for the management system that ISO/IEC 42001 defines. It carries a separate AI section with overview, security and monitoring entries, so AI is treated as its own control domain rather than folded into general security, and the whole trust centre is described as continuously monitored with automated compliance monitoring and automated alert response. That is real substance, and it is the reason this is not the middle grade. What keeps it off the top is that nothing behind the titles was read and nothing is attested. Both policies sit behind an access request, so their existence and titles are established and their contents are not, and crediting a document by its title is not available here. No ISO/IEC 42001 certification is claimed, in contrast to ISO/IEC 27001:2022 which is claimed with its version. No evaluation result, benchmark or model card is published. Nobody is named as accountable for the system's behaviour. And no bias, fairness or representativeness disclosure exists at any level, which is worth naming for a product whose identity confusion failure mode, published by the vendor itself, involves combining information belonging to people with similar names, addresses or roles. Trust centre and failure modes paper read 7 September 2026.

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.

The published control inventory is the broadest located in this pull, and one limb is missing. 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, with data encrypted and hosted in Azure and processed through private large language models. Subprocessors are named on the public page rather than on request, being OpenAI, Baseten, Pinecone, Langfuse and Microsoft Azure, with Amazon Web Services and Azure listed as infrastructure, and a subscribable feed publishes changes to that list with dated entries recording the addition of Langfuse for prompt use and of OpenAI. Incident practice is present, with a documented incident reporting process and designated response personnel. Access control is present, with least privilege, access monitoring, access log management and data asset classification. Around those sit data loss prevention, application and network penetration testing, code analysis, credential management, backups with a backup policy, business continuity and disaster recovery including contingency plan testing and failover, change management with change notification and verification, endpoint controls including mobile device management and threat detection, anti-DDoS and firewalls, and security, privacy and phishing training. The gap, named because every partial names its limitation, is retention and deletion: no retention period, no deletion commitment and no end-of-matter or end-of-engagement disposal position was located anywhere, which is the question a litigation buyer asks about a production it has uploaded. Trust centre, security page and homepage read 7 September 2026.

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.

Nothing establishable on this date states who bears the loss when the product is wrong, and the cause is two different things which the reader should be able to tell apart. The agreements exist and the vendor says so: a Master Services Agreement and a Data Processing Agreement are both listed by name in the trust centre's legal section. Neither is published; both sit behind an access request, which is a disclosure choice about how the terms are made available rather than an absence of terms. Separately, the website terms of service linked in the site footer render client-side and returned the hosting platform's shell with no body on this channel, and a search on clause language did not recover the text, so that document is a limit of this reading rather than a vendor decision. The consequence for a buyer is the same either way and is what the grade records: no indemnity, no liability cap, no warranty, no service level commitment, no exclusive remedy and no insurance position can be read before entering a commercial conversation. The exposure worth naming on this product class is that its outputs become fact statements and pleadings a lawyer signs, and its own published failure taxonomy describes seven ways such outputs go wrong, with nothing published about what the vendor stands behind if they do. Trust centre read and terms of service attempted 7 September 2026.

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.

One integration, into the system this product class actually runs on, documented on that platform's own marketplace. Fileread is available as a native integration inside Relativity, with a listing on the Relativity app hub, and the integration is described by a customer rather than only by the vendor, a practice support manager at Nutter McClennen & Fish attributing easier adoption to it. That is the deepest integration available in this segment: Relativity is where litigation productions already sit, and the vendor's design position follows from it, that the platform works on the productions a team already has with nothing to migrate. A second distribution surface exists on the Microsoft commercial marketplace, and the trust centre lists integrations as a product security control. Cimplifi, an eDiscovery services provider, describes offering the product through its own ecosystem, which is a channel rather than an integration and is recorded as such. What is absent is breadth and documentation. No API or developer documentation was located anywhere in the site navigation or footer. No document management system, case management platform, word processor or e-signature product is named. No integrations page or connector directory exists, and nothing describes what moves in which direction between Fileread and Relativity or how permissions travel across that boundary. Homepage, footer, Relativity app hub listing and trust centre checked 7 September 2026.

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.

Neither co-equal limb is stated, and what is published in their place is infrastructure. The cloud providers are named clearly and in more than one place: data is stated to be encrypted and hosted in Azure and processed through private large language models, and the trust centre lists both Amazon Web Services and Microsoft Azure under infrastructure and Microsoft Azure again among the subprocessors. Naming the cloud is where a workload runs, not the tenancy model and not the region, and it is credited on the supply chain and stewardship rows rather than here. On tenancy, nothing states whether customers are separated logically or physically, whether a matter is isolated from another matter, or what single-tenant options if any exist; the phrase private large language models points at dedicated inference rather than at tenancy of the data store and is not treated as a tenancy statement. On region, nothing states a country or region of storage or processing, no region selection is offered, and no data residency commitment appears despite GDPR compliance being claimed, which ordinarily brings a transfer position with it. Two gated artefacts in the trust centre would answer this directly and were not read, being a Data Flow Diagram and a Network Diagram, and they are the named rebuttal route on this row. Security page, homepage and trust centre checked 7 September 2026.

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.

A real trust centre with a named artefact inventory, and the reports themselves behind a request whose terms the portal does not state. What is ungated and readable: five compliance positions, being SOC 2 Type 2, ISO/IEC 27001:2022 with the standard version given rather than the bare number, HIPAA, GDPR and CCPA; a control inventory spanning roughly 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 centre updates carrying dated subprocessor changes. That breadth of ungated detail is more than most records in this corpus publish at all. What sits behind the access flow: 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, to view and download sensitive information and to request access, and does not state whether access is granted instantly on an email address, by a self-service non-disclosure click-through, or by vendor approval after a sales conversation. Under the standing rule for gated artefacts the lower tier is taken where the portal does not say which it is, and that is why this is not the top grade, alongside the absence of any named auditor, audit period or scope statement. No request was submitted. Trust centre and security page read 7 September 2026.

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 fullest provider-level disclosure located in this lane, and the models are still not named. The trust centre publishes the subprocessors on its public page rather than behind the document request, and the list is legible as an AI stack rather than as generic infrastructure: OpenAI, Baseten, Pinecone, Langfuse and Microsoft Azure, with Amazon Web Services and Azure also listed as infrastructure. A reader can therefore establish that a named model provider is in the path, that inference serving and vector storage sit alongside it, and that a prompt observability layer is involved, the update feed recording in terms that Langfuse was added as a subprocessor for prompt use and that OpenAI was added as a subprocessor. Change notification exists as a published mechanism rather than a promise, the trust centre carrying a dated subprocessor changes feed with a subscribe option, which is a limb most records on this axis fail outright. Two things keep it off the top grade. No model or model family is named: the site says data is processed through private large language models, and under the standing rule naming a provider is not naming the models underneath. And nothing states which parts of the platform route customer content to which provider, so a reader knows OpenAI is in the stack without knowing what it sees. Trust centre and security page read 7 September 2026.

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.

No pricing information is published at any level, including the unit of charge, and the site's own navigation establishes that no pricing page exists rather than that one could not be reached. The published inventory of the site is Product, Solutions with three segment pages, Customers, Security and Resources, with a footer adding Workflows, Company, Privacy and Terms of Service. There is no pricing entry in either the header or the footer. Every conversion path on 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 or edition is named, no unit is identified whether by seat, matter, document, gigabyte or production, no band or range appears, no minimum or term is stated, and no free tier or trial is described. The two documents that would carry the commercial mechanics are unavailable: the Master Services Agreement is listed in the trust centre behind an access request, and the website terms of service render client-side and returned no body. Where the only route is an invitation to contact sales, no pricing row is owed and none is written. Site navigation, footer, security page and trust centre checked 7 September 2026.

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.

The buyer, the workflow and the material are each described with real specificity, and no boundary is stated. Three buyer segments have their own surfaces: litigation teams, eDiscovery, and corporate legal. Five workflows are published individually, being early case assessment, discovery document review, evidence analysis, deposition preparation, and factual briefs and case narratives, and the lifecycle claim is explicit, from first production to trial. The named customer set matches that spread rather than sitting in one corner of it, running from a Washington litigation firm and a boutique through a large firm's knowledge and practice support function to a litigation administration provider and an eDiscovery services company. Coverage of the material itself is unusually concrete and is the strongest part of this row: text, images, handwritten notes, audio, video and spreadsheets in the same matter, with the vendor claiming the formats that defeat other tools, illustrated by a production that arrived as 4,000 text message screenshots. What is absent is any limit. No jurisdiction is named, and none is excluded, though the customer set and the vocabulary are United States civil litigation throughout. No matter type, case size or practice area is named as in or out of scope. No volume ceiling is published despite scale being the product's central claim. Homepage, solutions and workflow navigation and customer testimonials read 7 September 2026.