DeepJudge vs Midpage: 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 artificial intelligence is the product. Midpage is a legal research and brief generation platform where retrieval, synthesis and drafting are all model driven; there is no underlying workflow or document management system that would remain if the models were removed. The company was founded in 2022, after the generative wave rather than before it, and the product is distributed both standalone and as an integration inside ChatGPT, which is a distribution model only an AI native product can have.

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.

Accuracy is independently measured, with a published method and published figures for each criterion. Midpage submitted its standalone research product to the 2025 Vals Legal AI Report legal research study, run by an outside evaluator. The evaluator published its rubric and weights in advance: accuracy 50 percent, authoritativeness 40 percent and appropriateness 10 percent. The study used 200 United States legal research questions from attorneys at named firms, including Reed Smith, Fisher Phillips, McDermott Will and Emery, Ogletree Deakins, Paul Hastings and Paul Weiss. Midpage scored 78 percent accuracy, 74 percent authoritativeness and 70 percent appropriateness, a 76 percent weighted aggregate. That beat the lawyer baseline by 7 points on accuracy and 6 on authoritativeness. The figures sit on the evaluator's own site, where anyone can check them. The evaluator also published the failures: three cases where Midpage returned no response because of technical issues, and eight where it explained it could not locate supporting documents. All evaluated systems, this one included, did worse than a general model when asked to survey all fifty states for a single statute, and struggled on questions spanning several jurisdictions.

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.

No published position is published on what the system does unaided, what a lawyer must review, where a workflow stops, or what the vendor commits to when an output is wrong. The evaluator's report records that the product sometimes explained it could not locate supporting documents rather than answering, which is observed behavior rather than a published oversight model, and it was not treated as one.

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.

Independent benchmark results stand where deployment evidence would go. The strongest evidence attached to this vendor is test set performance measured by an outside evaluator. That is real, but it is not a production outcome: it shows how the system performs on 200 constructed questions, not what changed at a firm that adopted it. The vendor site, product pages and third party coverage were searched. No named customer, case study pairing an organization with figures and a date, or assessable deployment method is published. The study questions came from attorneys at named firms, but those firms supplied test material and were not disclosed as customers.

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.

The vendor's security page answers two of the three questions precisely. Training is addressed at both layers: Midpage does not use customer data to train or fine tune AI models, and its agreements with model providers do not let them train on Midpage customer data. Retention is published with actual periods. Web app data, including queries, uploads and outputs, is held while an account is active and deleted within 60 days of account deletion or a valid deletion request. Plugins and integrations store no queries, uploads or outputs, though those workflows may share queries with model providers, who may keep them for up to 60 days. The page discloses the providers' retention window as well as its own. Encryption is AES-256 at rest and TLS 1.2 or higher in transit, subprocessors must maintain measures no less protective, and an annual SOC 2 Type II audit validates the controls. The third question is unanswered. No separation model between customers, users or matters is published, and attorney client privilege and work product are not addressed directly.

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 stated audience is litigators and law students, a population of lawyers and trainees rather than consumers, and the vendor is explicit about it. The student audience is worth noting, because it includes people who are not yet admitted, which raises a professional responsibility question of a different shape rather than an unauthorized practice problem. The vendor site, product pages and third party coverage were searched. No published position on the advice line, no treatment of competence or supervision duties, and no jurisdiction limits are published. The audience is stated clearly and the position is not.

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.

Submitting to independent evaluation is the governance act here, and it is a real one. Midpage was one of only four systems that agreed to be measured and named in the Vals legal research study. The evaluator and trade press both recorded that major vendors declined to take part, and at least one took part but withheld permission to publish its results. Agreeing to publication, including your own failure cases, is a meaningful accountability position. But it is a single evaluation, not a management system. The vendor site and third party coverage were searched. No AI governance framework, AI management certification, named owner of model governance or testing regime before release is published, and nothing on uneven output across matter types, parties or populations.

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 vendor's security page covers most of the ground. It publishes AES-256 encryption at rest and TLS 1.2 or higher in transit using industry standard algorithms. It states a retention and deletion practice with a 60 day outer limit after account deletion or a valid deletion request. Queries, uploads and outputs from plugin and integration workflows are not stored at all, and the page discloses that model providers may keep submitted queries for up to 60 days. Subprocessors are contractually required to maintain security measures no less protective than the vendor's own, and the whole control set is validated by an annual SOC 2 Type II audit. More detail on service providers and data handling is available on request. Three things are not published: a named subprocessor list, as distinct from the contractual requirement; a hosting provider or region; and an incident or breach notification practice.

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 indemnity, liability cap, carve out, warranty on output or insurance position is published, and no customer terms of service is published as published on the property. Recorded as a pure absence on the surfaces reached. The exposure is worth naming: this is a litigation research product whose output is authority bound for filings, in a market where the published sanctions record for fabricated citations is growing, and nothing published addresses who bears the loss.

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.

One integration is named and none is documented. The product is available inside ChatGPT through what the evaluator calls a bespoke integration, which is a distribution channel into a general assistant rather than a connection to a firm's own systems. The vendor site, product pages and third party coverage were searched. No document management connector such as iManage or NetDocuments, practice management or court filing integration, or Microsoft Word add-in is published, and nothing documents what the ChatGPT integration passes or what an administrator configures. For a product sold to litigators inside firms, the lack of any connector to firm systems is the main gap.

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.

The vendor site, product pages and third party coverage were searched. Nothing is published on the deployment model: no tenancy statement, hosting provider, region, residency option or statement of where processing happens, and the security page names no host or region. The product is evidently delivered from the cloud, through a browser and through ChatGPT, but that is inferred from how it is sold, not a published position.

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.

The vendor states directly on its security page that its controls are independently validated through an annual SOC 2 Type II audit, which supplies the cadence as well as the standard, and states that on request it can provide SOC 2 documentation, questionnaire responses and additional detail on service providers and data handling. That is a published request route to substantive diligence material rather than a sales gate. A customer of this vendor, GC AI, separately publishes that Midpage is SOC 2 Type II compliant, which corroborates the vendor's own statement. No coverage period, audit scope, report date or named auditing firm is published, and no trust portal exists; the route is a request to the vendor.

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.

No model, provider, hosting location or subprocessor is published, and no commitment to notify customers of supply chain changes. The evaluator's report distinguishes Midpage's standalone research product from its ChatGPT integration, which tells a reader the product runs independently of that channel, but nothing published identifies what powers the standalone product. A buyer cannot determine from published material which company processes their research queries.

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 page is published, no rate is published, no unit of charge is stated and no tier structure appears. No free trial or self serve entry point is published on the vendor's own property, and no third party pricing figure is published either. Noted for a reader: the product being reachable through a ChatGPT integration may mean some access runs on that platform's commercial terms rather than the vendor's, which published material does not clarify.

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).

Practice focus and jurisdictional coverage are both stated with substance. The practice position is narrow and explicit: the product is built for litigators, and the evaluator independently confirms it focuses on litigator workflows, with responses suited to the research questions litigation produces. Jurisdictional coverage is listed by court system: United States federal and state courts, tribal courts and military courts, plus statutes, regulations and selected administrative decisions. Naming tribal and military courts is precise. Two things are missing. The segment is described by practice type without addressing firm size. No statement of what the product does not cover is published; the measured weakness on fifty state surveys and questions spanning several jurisdictions comes from the evaluator, not the vendor.