DeepJudge

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.

There is no content asset and no conventional product underneath. Unlike the other research records in this lane, DeepJudge licenses nothing and publishes nothing of its own: the material it works on belongs entirely to the customer and stays in the customer's systems. What DeepJudge supplies is the retrieval itself, described as intent-based search that understands content, context and relevance, with automatic classification into a taxonomy and near-duplicate and redline detection across collections the vendor describes as billions of unstructured and structured data points. Strip the models out and what remains is a set of connectors and a permissions synchroniser, which is plumbing rather than something a firm would buy. The founders' own framing, that search rather than models is the wedge, is a claim about which machine learning matters, not a claim that machine learning is peripheral. Verified 2 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 structural rather than asserted: every result is a document the firm already holds, in its own system, which the user can open and read in full, so the verification path is the shortest of any record in this pull. The retrieval method is described at a working level, covering intent-based search, automatic taxonomy classification, near-duplicate and redline detection, and multi-level combination of client, matter, document and people data. Section 4(j) of the subscription terms places a contractual obligation on the customer to verify, including through human review, the accuracy, reliability and appropriateness of the AI features. What is absent is measurement. No accuracy figure, evaluation, test set or benchmark appears anywhere, and nothing addresses hallucination in the workflow and agent layer, which is where generated text rather than retrieved documents is produced. The one published number nearby is a satisfaction measure, that 90 per cent of users report finding results faster than with existing tools, which is not an accuracy claim. Verified 2 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.

The human review requirement is contractual and the excluded uses are unusually explicit. Section 4(j) requires the customer to verify the accuracy, reliability or appropriateness of AI features through human review, and the same clause prohibits use in judicial decision-making by judges, in the professional activities of lawyers serving as arbitrators, mediators or other alternative dispute resolution neutrals, and in any use restricted under the EU Artificial Intelligence Act. Naming the roles that may not use the product is rarer and more useful than the usual disclaimer. What is not published is the operating boundary. The workflow layer is described as agentic reasoning that executes complex multi-step tasks, retrieving from multiple sources and taking actions, and the customer builds those workflows itself using a low-code builder, so the threshold at which a workflow acts without a human is set by the firm and described nowhere. Governance of agents is offered as a capability without any account of what the governance controls are. Verified 2 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.

Sixteen firms are named on the home page, spanning global elite practices including Freshfields, Greenberg Traurig, Holland and Knight and Gunderson Dettmer, United States mid-market firms including Cozen O'Connor, ArentFox Schiff and Vorys, Swiss and Austrian practices including Homburger, Lenz and Staehelin, Wenger Vieli, Advestra, Schoenherr and CMS Switzerland, and one in-house function at SBB. Three named individuals speak on the record with title and firm, including two chief innovation officers. Figures are published: 65 hours saved per user per year on searching, 85 per cent adoption within two months, 23 queries per user per day, fourfold return in the first year, and 90 per cent reporting faster results. One figure is properly attached, with a named M&A partner at Homburger stating that more than 80 per cent of the firm's legal professionals had incorporated it into their workflow. What holds this below A is that the headline figures are aggregate and carry no method or measurement basis, and no dates appear on the home page; the individual customer story pages were not opened this pass. Verified 2 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.

Four of the five limbs are met, and the segregation limb is met better here than by any other record in this pull. Section 5(b) of the subscription terms provides that DeepJudge does not use customer data to train the AI or machine learning models underlying the services, which is a contractual prohibition naming the thing rather than a policy statement. Segregation is the product's architecture rather than a claim about it: access permissions are continuously synchronised with the firm's source systems so that existing ethical walls are maintained, which is the matter-level requirement a firm buyer needs, and section 5(e) adds that customer authored workflow logic will not be disclosed or incorporated into the service provided to any other customer. Retention and deletion are stated, with customer data deleted within sixty days of termination. The position on model providers is answered structurally, since the firm selects the models and the hybrid deployment clause discloses exactly when processing leaves the firm's own infrastructure. The limb that fails is privilege and work product handling: neither term appears anywhere, and ethical walls address conflicts rather than privilege. Verified 2 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.

This is the most complete professional responsibility position located in the pull, and all of it sits in the operative agreement. Section 4(j) states that DeepJudge is not a law firm and does not engage in the practice of law and that the customer acknowledges it is not relying on the services for legal advice. It states who may use the product and, unusually, who may not, prohibiting use in judicial decision-making or other activities performed by judges, and in the professional activities of lawyers serving as arbitrators, mediators or other alternative dispute resolution neutrals. It reaches competence and supervision through a positive obligation rather than a disclaimer, requiring the customer to verify the accuracy, reliability or appropriateness of the AI features including through human review. And the legal boundary is stated twice: section 4(a) confines use to compliance with the obligations applicable where the customer and its users are located, and section 4(j) excludes any use prohibited or restricted under the EU Artificial Intelligence Act, naming the instrument. The coverage limb that asks about jurisdictional reach of legal content does not bite on a product that searches the firm's own documents rather than any jurisdiction's law, and is recorded here as inapplicable rather than failed. No bar or law society guidance is engaged, which is recorded on the signal rather than here. Verified 2 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.

One binding governance commitment is published and everything around it is missing. Section 4(j) of the subscription terms makes compliance with the EU Artificial Intelligence Act a contractual restriction on use, naming the instrument and excluding the administration-of-justice applications the Act treats as high risk, which is more than a principles page and is enforceable by the customer. The workflow layer is also sold partly on governance, offering to build, deploy, orchestrate and govern AI agents. What no published material supplies is any content behind either. No responsible AI page exists, no governance framework, no individual or function named as accountable for model behaviour, no account of what is tested before a release ships, no certification such as ISO 42001, and nothing at all about uneven output across matter types, practices or populations. ISO 27001 and SOC 2 Type II are security attestations, which this axis treats as a different subject. Checked the home page, product page, security page, subscription terms and privacy policy on 2 September 2026. Verified 2 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.

Amended 2 September 2026 under R42, after the data processing addendum version 1.1 of May 2026 was supplied by the operator; it had defeated both direct fetch and search on the original pass and this row was then graded B on three of five elements with two expressly not credited. All five are now established and specific, so the grade moves to A. Retention and deletion: section 3(a) of the subscription terms deletes all customer data including training data and customer models within sixty calendar days of termination, and DPA clause 11 adds prompt destruction or return on request, extending to data held by subprocessors, with confirmation of destruction. Access control: permissions continuously synchronised from source systems on least-privilege principles, access restricted to authenticated users, configurable audit logging, indexing controls over what is crawled at all, and DPA Annex 2 adding multi-factor authentication, role-based access control, centralised event logging and segregated pre-production and production environments. Subprocessors: Annex 3 names five entities with purpose and location, and clause 5 requires equivalent written terms, due diligence, vendor liability for their acts, thirty days notice before a new subprocessor processes anything, and an objection right with termination without liability. Incident practice: clause 7 commits to notice without undue delay and in any event within seventy-two hours of discovery, covering breaches at subprocessors as well as at DeepJudge, with the definition at 1.13 expressly including unauthorised internal access, plus cooperation so the customer can 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 controls audited for SOC 2 Type II. Clause 3.4 adds a purpose limitation binding subprocessors: no selling or sharing, no processing for other purposes, no retention or disclosure outside the direct business relationship, and no combining customer personal data with data from other sources.

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.

The whole allocation of loss is published, specific, and mutual, which is rare in this corpus. Section 7(a)(ii) caps each party's total aggregate liability at the fees paid to DeepJudge in the twelve months before the event giving rise to it, and names four express exceptions that sit outside the cap: the customer's payment obligations, breach of the use restrictions at section 4(e), fraud or wilful misconduct, and infringement of intellectual property rights. Section 7(a)(i) excludes indirect and consequential damages symmetrically for both parties, with the same carve-outs. Section 7(b) gives the customer a real indemnity running the right way, with DeepJudge defending third-party claims that the services or underlying software infringe intellectual property rights, and setting out the remedies in order: procure the right to continue, modify or replace to become non-infringing, or terminate the affected portion. There is also a service commitment a buyer can hold: section 1(a)(iii) sets a 99 per cent quarterly uptime target with downtime defined and six named exclusions. What is not offered is any warranty on output, since section 6 disclaims accuracy and completeness, and no insurance is mentioned. Verified 2 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.

Integration is not a feature of this product, it is the product, and the depth is documented accordingly. The systems named are the ones firm work actually lives in: the document management system, email, SharePoint, OneDrive, HighQ, intranets and experience management and metadata systems. What moves is described in both directions. Documents are crawled and indexed in place with no migration and no uploading, so the firm's data remains in its source system; permissions travel the other way, continuously synchronised from those systems into the index; and the firm decides through indexing controls what is crawled at all. What a firm must configure is set out contractually rather than left to implementation: section 1(d) requires continued access to the customer databases named on the order form, requires a named technical contact, and warns that withholding access degrades the service, while the on-premises clause makes the customer responsible for meeting stated hardware and configuration requirements. The platform also exposes governed access to other AI systems through the Model Context Protocol and publishes an open agent handoff specification. Verified 2 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.

Three deployment models are defined in the agreement itself, with the consequence of each spelled out, which is the most complete treatment of this axis in the pull. Cloud deployment hosts the services in a cloud hosting location named on the order form. On-premises deployment runs inside the customer's own infrastructure, with DeepJudge granted access to install and maintain, and the customer responsible for stated system requirements. Hybrid deployment is defined precisely as on-premises with certain processing, expressly including model processing, occurring outside the customer's infrastructure, and the clause states plainly that customer data may be processed in the cloud hosting location and that model processing may involve transfer of customer data outside the customer's systems. That is processing separated from storage, stated by the vendor rather than inferred, and it names the one circumstance in which an on-premises customer's data leaves. The security page adds that the customer runs its choice of model in its choice of residency. The limitation worth recording is that no specific regions are enumerated: residency is selectable and designated on the order form rather than published as a list. Verified 2 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.

Corrected 2 September 2026 under R42, after the data processing addendum version 1.1 of May 2026 was supplied by the operator. The original note stated that no penetration test or vulnerability assessment programme was described and called that unusual alongside two certifications. That was wrong and the document corrects it: Annex 2 records annual penetration testing, and states that controls are audited for SOC 2 Type II. The grade does not move. Two attestations are claimed and one carries its scope, with SOC 2 Type II stated as covering security, availability and confidentiality and ISO 27001 certification stated as covering the information security management system, which is more than most records in this pull offer. What the top band asks for is still absent from every surface a prospect can reach: no auditor is named, no certificate number, issue date, report date or coverage period is published, and there is no trust centre or portal. The route to evidence is a customer route rather than a public one. DPA clause 9 sets it out: certifications and audit reports or written questionnaires are made available first, and a customer audit is permitted only where those cannot demonstrate compliance, at the customer's expense, once per calendar year, with no access to systems, hosting sites or infrastructure. That is a real entitlement for a signed customer and not accessible evidence for a buyer still deciding, which is what keeps this at B.

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.

Amended 2 September 2026 under R42, after the data processing addendum version 1.1 of May 2026 was supplied by the operator. The original note said that not one provider was named and that no change notification commitment existed. Both are corrected: 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 customer selected, and 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 cannot be resolved. The grade does not move, and the reason is the shape of what is named. The three entities are hyperscaler platforms rather than model vendors: no discrete model house appears anywhere in the list, no model is named, and no default is stated, so a reader learns which cloud estates the inference runs in but not whose model produced the output. Provider-level naming does not reach the top band and platform-level naming reaches it less. What is genuinely well disclosed is the architecture and the geography: DeepJudge states that its architecture is model-agnostic and that the customer chooses from and may combine any commercial or open-source models, running its choice of model in its choice of residency, and the hybrid deployment clause in the subscription terms discloses that model processing may involve transfer of customer data outside the customer's own infrastructure. Customer-selected location on all three AI providers corroborates the deployment position rather than adding to it.

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.

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 per cent 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. Verified 2 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.

Who buys this is evidenced rather than asserted, and the roster does the work: global elite firms including Freshfields, Greenberg Traurig and Holland and Knight, United States mid-market practices including Cozen O'Connor, ArentFox Schiff and Vorys, a technology-sector specialist in Gunderson Dettmer, Swiss and Austrian firms including Homburger, Lenz and Staehelin, Wenger Vieli, Advestra and Schoenherr, and an in-house legal and compliance function at SBB, so both private practice and in-house use are demonstrated. Geographic reach is stated as teams across North America and Europe serving clients worldwide. Practice areas are not enumerated, and on this product that limb does not bite in the usual way, because the platform searches whatever the firm already holds rather than covering any defined body of law. What is genuinely missing is the boundary: no firm size band is named, nothing states a minimum data estate or user count below which the product is not sold, and no statement describes where it stops. The clearest limits published are the prohibited users at section 4(j) rather than any positive statement of scope. Verified 2 September 2026.