DeepIP vs ipQuants: how they compare in 2026

The models are the product and the Word add-in is only the delivery surface. The vendor states proprietary generative models fine tuned for patent law, custom models trained on patent data to reduce errors and to mimic previously successful filing styles, and agentic AI handling complex multi step tasks across drafting and portfolio analysis. Founders are described as having built AI systems for Airbus, IBM and SAP before this, so the model work is the founding competence rather than an added capability. Every function is generated or model driven: drafting, office action response suggestions, prior art surfacing, patentability analysis, style matching and portfolio insight. Remove the models and what remains is a Word sidebar with nothing in it. Third consecutive A on this axis in ip-and-patents, and the category now matches the plaintiff pattern at three for three.

The models are the engine of a core capability and the agreement itself shows what sits underneath them. The terms divide the platform into three components: the Qthena Digital Cockpit, a feature set for document review, annotation and management of published patent documents; analytics and big data features that let a user query ipQuants' proprietary datasets and statistics without uploading their own content; and generative AI functionality, currently AskQthena, for analysis, summarization, drafting and interactive queries. Only the third is generative. The first two are a document workspace and a statistics product that would still work with the models removed, and the company's original distinguishing claim was data rather than generation, being first to publish EPO examiner, opposition and appeal member insights. The generative layer is plainly where the recent product work has gone, with named skills for drafting, prior art comparison, invention disclosure and drawing generation, but this is a machine learning layer on a working platform rather than a product that disappears without it. Checked 4 September 2026.

Error reduction is claimed repeatedly and no grounding mechanism or measurement is published. Stated: custom models trained on patent data reduce errors, the platform is engineered to reduce errors and support legal rigor, and relevant prior art is surfaced automatically. Those are claims about outcome rather than descriptions of architecture, and nothing located describes how a generated passage is tied to a source, whether citations to prior art or case law are produced, or whether a user can trace an assertion back. No accuracy figure, no evaluation, no hallucination disclosure and no confidence signal were located. Recorded against the record rather than glossed: an independent three month review of this product states that an AI hallucination issue exists and is manageable with proper review processes. That is a third party observation rather than a vendor disclosure and is not credited as one, and it is noted here because the vendor addresses the risk nowhere. Compare Solve Intelligence at B in this category for publishing a continuous human evaluation methodology, and Patlytics at B for color coded confidence indicators.

The published limits are unusually candid and none of them is a measurement. The company states that Qthena's drafting skill does about 80 percent of the work with the user finalizing, that drawing generation reliably completes around 90 percent, that it deliberately withheld both skills until quality was adequate, and that it expects a higher figure as underlying models improve. It also names the failure modes it says competitors ship with, listing broken geometry, wrong perspective, missing edges and incorrect part definition. Those are completion estimates rather than accuracy figures: no test set is described, no evaluation method is published, no error rate against a benchmark appears, and nothing states how the percentages were arrived at. The terms disclaim accuracy directly, stating that generative output is probabilistic, may not always be accurate, complete or appropriate, and may not be relied on as a sole source of truth. Grounding is real in the sense that the corpus is published patent documents and the company's own datasets, but no retrieval method is described, which is what the band above requires.

A clear positioning statement sits alongside an autonomy claim and neither is reconciled. The vendor states that its tools are designed to augment attorney capabilities rather than replace them and are built through continuous feedback from leading IP firms, which is the right posture. It separately states that its agentic AI takes on complicated multi step jobs, whether drafting patents or extracting insights from an entire portfolio, and that it manages office action deadlines and generates response suggestions. Deadline management on a prosecution docket is an area where an unattended error has consequences that cannot be undone. Nothing published reconciles the two: no statement of what an agent may complete without attorney approval, no confidence threshold, no escalation behavior, and no description of what review the vendor expects before a generated response is filed. Held at C on that basis, with the observation that a product embedded inside Word places the attorney at the point of authorship structurally, which is oversight by architecture rather than by published policy.

A written commitment that the human decides, backed by real constraints in the agreement, short of published thresholds. The company states plainly that it will always ensure the human is in charge and decides on each step, and that the drafting skill produces a first draft the user must finalize. The terms carry that into obligations rather than leaving it as marketing: the customer must evaluate output for accuracy and appropriateness including through human review as needed, must not rely on output as a sole source of truth or as a substitute for professional advice, and must not use output relating to an individual to make decisions with a legal or material effect on that individual. Review surfaces are real, with the cockpit holding the documents and drawings correctable conversationally so a user can adjust reference signs and line work. What is missing is the threshold. The company says it shows customers when AI-generated drawings can be used, when they need human correction, and when they should not be used at all, but that guidance is delivered by its team during rollout rather than published, so a buyer cannot read it before committing.

Company evidence is dated and specific, and no customer is identifiable. Checkable: $40m total funding across a $15m Series A announced March 2025 led by Resonance with Headline, Serena Capital and Balderton Capital participating, and a $25m Series B announced 2 March 2026; named founders with named prior employers; a named chief technology officer quoted describing the product architecture. Traction figures are unusually concrete for an early company: more than 8,500 applications supported and seven figure revenue within seven months of launch. Customer evidence is present but anonymous: a trial period quote reporting approximately 20 percent efficiency improvement in drafting and prosecution, and a security due diligence quote describing evaluation of the Azure deployment and certifications, neither attributed to a named firm or individual. An independent three month review reports 40 to 60 percent reduction in initial drafting time across patent types. Held at B rather than A because no customer is named anywhere in located vendor material, which is the difference between this record and Solve Intelligence at A in the same category.

Unattributed claims of scale stand in for deployment evidence on the surfaces that were read. The company states that Qthena is used by hundreds of corporate teams and law firms across the globe, and that many corporate teams and law firms trust it worldwide, without naming one. The single customer voice located is anonymous, quoted only as a valued customer saying the product is for the smart user, and it carries no firm, no role and no figures. No dated deployment, no measured outcome and no named reference was located on the pages read. One first-party partnership is documented and is a different kind of fact, recorded rather than credited as deployment evidence: Questel integrates Qthena into its own IP platform, announced by both parties. A dedicated case studies page exists in the site navigation and was not opened in this pass; it is named here as the limit and is the cheapest available upgrade on this record.

A coherent confidentiality set with one element no competitor offers, and no engagement with privilege as a professional concept. Published: SOC 2 Type II and ISO 27001 certification, GDPR compliance, a strict Zero Data Retention policy, complete data segregation, encrypted storage and end to end encryption, deployed on Microsoft Azure. The element that distinguishes this record is on premise deployment, which is graded on the Deployment axis and matters here because it is the only published answer in this category to a firm that cannot let an unpublished application leave its own infrastructure at all. A customer quote describes conducting extensive due diligence on the security infrastructure before selection, which indicates the claims survive procurement scrutiny. What is absent: no treatment of attorney client privilege or work product, no reference to the professional confidentiality obligation, and no acknowledgment that an unpublished application carries consequences beyond ordinary data sensitivity. Compare Patlytics at A and Solve Intelligence at A in this category, both of which engage the professional dimension directly.

Substantive commitments in the published agreement, including the one most vendors leave out, and no privilege treatment. Section 15 defines confidential information to expressly include customer content, customer information and generative AI inputs and outputs generated on behalf of the customer, so the material a patent attorney puts into the tool is contractually confidential rather than merely secure. Section 15.6 goes further than most: LLM providers are named, customer input and output are stated not to be used to train or improve those models, identifying metadata such as email and IP addresses is stated not to be transmitted, access is through paid enterprise APIs, and abuse monitoring is deactivated by default for at least one named provider so prompts are not scanned. Section 15.6.2 leaves content under customer control until the customer deletes it, and section 9 commits to permanent deletion of all customer data on termination. Two things hold it here. No privilege or work product treatment appears anywhere, and that limb is required rather than satisfied by a strong confidentiality regime. And nothing describes separation between customers or between matters inside a firm's own workspace, on a product built around shared project views and collaboration.

Not located. The product drafts patent applications and generates office action responses filed under a registered practitioner's signature, and it manages prosecution deadlines, all of which engage the duty of competence and the practitioner's responsibility for filed work. The vendor states its tools are designed to augment attorney capabilities rather than replace them, which is a product positioning statement and not a professional responsibility position. No reference to USPTO Rules of Professional Conduct, 37 CFR, duty of competence, or any bar or patent office guidance was located. Checked the home page, the law firm solution page, the product pages, the blog including the drafting guide, and the funding announcements on 29 Aug 2026. Compare Patlytics at B in this category, which names the specific rules directly.

A real published position on the advice line, with no engagement with professional conduct rules. The terms state expressly that no legal advice is intended or offered by ipQuants in making the platform, its content or any tools, features or support interactions available, and disclaim liability for decisions taken in reliance on any of it. That is paired with substantive restrictions on how output may be used rather than left as a single sentence: output may not be relied on as a sole source of truth or as a substitute for professional advice, the customer must evaluate it including through human review, and output relating to an individual must not be used to make decisions with a legal or material effect on that individual. The public material is consistent with it, the company stating that the human is in charge at each step and declining to claim the tool replaces skilled drafting. What is absent is the professional layer: no bar or patent-office conduct rules are named, no ethics guidance is referenced, no jurisdiction limit is drawn, and nothing addresses a representative's supervision or competence duties when a machine drafts a claim set or a set of figures for filing.

Nothing published about how the models are governed, evaluated or monitored. No AI policy, no model card, no bias or fairness testing, no evaluation methodology or result, no accuracy monitoring, no drift statement, no named governance body, no ISO 42001 and no EU AI Act positioning were located, the last being notable for a company with substantial operations in Paris and a stated European customer base. The vendor's own published guidance on evaluating AI patent drafting software tells buyers to look for versioning and audit trails for internal review and compliance, which engages traceability without addressing model governance. This is the weakest governance position of the three ip-and-patents records built: Patlytics holds an ISO 42001 certificate and Solve Intelligence publishes a continuous evaluation process and an EU AI Act self classification. Checked the home page, the solution pages, the blog library and the funding announcements on 29 Aug 2026.

No governance position of any kind was located. There is no responsible AI page, no principles statement, no accountable owner or function named, no pre-release testing regime described, no management system and no certification such as ISO 42001. Nothing addresses uneven output at all, which on this product would bear on performance across technical fields, drawing types and the languages of the patent corpus. The company does publish a quality philosophy, stating that it withholds a skill until it is genuinely useful and that it is provider-agnostic about models, and that is product judgment rather than a governance framework: it names no reviewer, no criteria and no results. Its abuse monitoring configuration is a confidentiality control and is graded on the privilege row rather than counted twice here. The site navigation and footer were inventoried on 4 September 2026 and carry no governance surface.

A strong and repeatedly stated retention position, with the training question left to inference. Published consistently across the funding announcements, the law firm solution page and independent coverage: a strict Zero Data Retention policy, complete data segregation, encrypted storage, end to end encryption and GDPR compliance, hosted on Microsoft Azure with a customer quote referencing United States based Azure servers. Zero data retention is the strongest retention posture available and it is stated plainly rather than hedged. What is not stated plainly is model training. The vendor's own comparison content lists zero data retention policies that guarantee client data is never used to retrain models among the criteria it prioritized when assessing tools, which is criteria language applied to a field rather than a commitment made about itself, and the distinction is recorded on the signal row. Held at B rather than A on that gap and because no retention default, deletion right or scope boundary is published for data held in the platform as distinct from the model layer.

Substantive published policy covering most of the ground, thinner where the documents could not be read. From the terms: customer content stays under customer control and remains accessible until the customer deletes it; on expiry or termination ipQuants commits to permanently delete all customer accounts, data and information held in the platform; data transmitted to LLM providers is anonymized, sent over encrypted paid enterprise APIs, retained on a minimized basis aligned to those providers' policies, and not used to train them; each party must inform the other promptly of any unauthorised access to or disclosure of confidential information; and availability carries a 99 percent uptime commitment with fourteen days' notice of scheduled maintenance and prompt notice of unscheduled downtime. The gaps are specific rather than general. Retention on the model provider side is described as minimized and aligned with provider policies, which states no period. No encryption specification, access control model or audit logging is described in the material that could be read. A full Information Security Policy and a data processing addendum are published as ungated links in the site footer; both resolve to a document viewer that returned only a title and a thumbnail, so their contents are unread and are neither credited nor held against the vendor.

No published position located. Nothing was found on liability for AI output, warranty, service levels or remedy. The exposure profile is the category's own and is sharpened by one product feature: office action deadline management. A missed statutory deadline in prosecution can result in abandonment of an application, which is a consequence no review step recovers, and no published service level or liability position addresses it. The drafting exposure is equally severe, since claim scope lost at grant is permanent. Checked the home page, the law firm and corporate solution pages, the product pages and the site navigation on 29 Aug 2026. Enterprise agreements govern this and are not public.

A real published position with a vendor-side indemnity, and a cap a buyer should read carefully. Section 11.2 commits ipQuants to defend and indemnify the customer against third-party claims that authorized use of the platform infringes intellectual property rights, with damages and costs finally awarded by a US or European court covered, and sets out remedies if infringement is alleged, running to procuring the right to continue, modifying the platform, or ceasing use with a prorated refund. Named carve-outs cover customer alteration, unauthorised combination, use in an unauthorised region and claims arising from customer content. That indemnity is stated to be the customer's exclusive remedy for infringement. Section 12.1.1 commits to 99 percent uptime and section 12.2 offers dedicated SLAs to enterprise customers. Against that, section 13 provides the platform as is with all implied warranties disclaimed including accuracy, and section 14 caps cumulative liability at the lesser of five thousand euros or twelve months of fees, expressly including claims relating to generative AI output. There is no warranty on output, no insurance position, and no indemnity touching the accuracy of what the models produce, which is the exposure this product creates.

THE BEST INTEGRATION POSITION IN THIS CATEGORY, and it closes the gap that held both peers at D. Native Microsoft Word integration is the core architectural decision rather than a connector: the product runs as a Word add-in so attorneys draft, edit and validate inside the environment they already work in, and independent comparison material treats this as the distinguishing feature against browser based competitors. Beyond Word: stated integration with leading IP management platforms, which is the docketing and portfolio layer a prosecution practice actually runs on and which neither Patlytics nor Solve Intelligence names at all, and a documented API for custom integrations, with the Series B announcement stating an intention to extend API capabilities for partner ecosystem developers. The vendor also frames integration as a traceability argument, stating that workflow integrations preserve audit trails by design where legacy tools require manual exports that erode traceability. Held at B rather than A because no IP management platform is named individually, so a firm cannot confirm its own docketing system is supported, and no API documentation was reached in this pass.

One documented connection into a system a buyer already runs, and no documentation an implementer could use. The substantial fact is the Questel relationship: Questel integrates Qthena into its own IP management platform and both companies have published the arrangement, so a firm already on Questel can reach the product there. Beyond that the platform is reached at its own application domain, and the terms contemplate authorized distributors marketing and providing access. What is absent is everything a practice would need to plan a deployment: no document management integration is named, so nothing addresses iManage or NetDocuments, nothing addresses Word, Outlook or an IP docketing system, no API or developer documentation was located, and no statement describes what data moves, in which direction or on what trigger. The terms in fact restrict automated access, prohibiting robots, AI agents and crawlers from accessing the platform or its content, which bears on programmatic integration. Feature pages for workflow automation and collaboration exist in the navigation and were not opened in this pass; they are named here as the limit and the grade rests on what was read.

ON PREMISE DEPLOYMENT, which no other record in this category offers and which is the only complete answer to the confidentiality problem this category has. The vendor states it is the sole patent drafting solution offering both cloud based and on premise deployment. For a firm or corporate IP department that cannot let an unpublished application leave its own infrastructure, whether because of a client mandate, a foreign filing license constraint or an export control regime, cloud residency selection is a mitigation and on premise is an answer. Cloud deployment is also specified rather than gestured at: Microsoft Azure is named as the host, with end to end encryption, and a published customer quote references United States based Azure servers assessed during security due diligence. Graded A because naming the host and offering an on premise option together exceed the customer selectable cloud residency that earned Solve Intelligence an A in this category, on the dimension that matters most for unpublished material. Held short of perfection because the only solution claim is a competitive assertion by the vendor about its rivals, no on premise architecture or support detail is published, and no region list is given for cloud customers.

Cloud delivery is unambiguous and neither of the co-equal limbs is stated in the material that could be read. The terms define the platform as a suite of online, web-based applications reached at a named application domain and its subdomains, so delivery is not in doubt and no on-premises or self-hosted option is offered anywhere. Nothing read states whether the platform is single or multi-tenant, and no dedicated or isolated option appears. No region is named for storage or for processing of customer content, and no residency commitment was located, which is a live question for European IP work given that the named LLM providers operate across multiple regions. The only geographic facts published are corporate and contractual rather than architectural: a Schaffhausen address, Swiss governing law, mediation seated in Schaffhausen and fees denominated in euros, none of which locates the data. The Information Security Policy and data processing addendum published in the footer are the surfaces where tenancy and residency would normally sit; both returned a viewer shell rather than their contents, and that is recorded as a limit on the reading rather than as a gap in the disclosure.

Two certifications stated consistently and nothing evidencing them. SOC 2 Type II and ISO 27001 are claimed across the home page, the law firm solution page, both funding announcements and independent coverage, with GDPR compliance alongside, and the consistency of the claim across dated press releases is itself worth something since a false certification claim in a funding announcement carries more consequence than one in marketing copy. A published customer quote states that the certifications and the Azure deployment were assessed in extensive security due diligence, which is third party corroboration that the claims survived a buyer's review. Held at B rather than A on the familiar absences: no auditing firm is named, no examination period, scope or certificate date is published so currency cannot be established, and no trust center, security page or self serve documentation request route was located. Under the three tier test the artifact is absent rather than gated. Calibration in this category: Patlytics reaches A on three certifications with named penetration testing partners published open, Solve Intelligence sits at B after its certification language was read against its own detailed page, and this record sits at B on two consistently stated certifications with no route to either.

No independent security attestation was located. No SOC 2 of either type, no ISO 27001 or ISO 42001, no penetration test summary, no named auditor, no examination period and no certification claim of any kind appears on the pages that were read, and there is no trust center or trust portal on the estate. The company does publish an Information Security Policy as an ungated link in its site footer, and that is a real disclosure artifact; it is a self-published policy rather than an independent examination, so it does not answer what this axis asks, and its existence is credited on the stewardship row instead of twice. Its contents could not be read: the link resolves to a hosted document viewer that returned the file title and a thumbnail only, which is a retrieval limit on this side and is not treated as an absence. The distinction that matters to a buyer is undrawn either way, since nothing on the readable estate claims an audit, so there is no unsupported badge here and no attestation either.

The composition of the stack is described in unusual terms and no party is named. The vendor states that the platform uses a combination of proprietary models, third party tools and exclusive data access, which is a three part disclosure acknowledging that external components and privileged data sources both exist. Naming the existence of exclusive data access is uncommon and is a real disclosure about the supply chain rather than about the models alone. What is missing is identity on every limb: no foundation model provider, model family or version is named, no third party tool is identified, no subprocessor list was located, and the exclusive data access is not attributed to any source. Recorded and deliberately not credited as a supply chain fact: an investor in the Series A is associated with a foundation model company, and an investment relationship is not a processing relationship. Compare Onspring at B for naming its provider outright, and Solve Intelligence at B for disclosing the structure with customer controls attached.

Three of the four limbs are answered, and the fourth is expressly reserved rather than merely absent, which is a better disclosure than most and stops short of the top band. The providers are named in the agreement itself rather than in marketing: the terms identify the LLM API providers as OpenAI, Microsoft Azure OpenAI and Google Gemini, and describe the hosting arrangement as paid, enterprise-grade API access with anonymized transmission and no training on customer content. Specific models are named in current public material, the company stating that it defaulted to GPT-4o and has defaulted to Gemini 2.5 since January 2025, with an expectation of moving again as newer models arrive. What fails is change notification, and it fails deliberately: the same statement declares the product LLM-agnostic and says the company can switch at any time, so a customer has no committed notice before the model reading its patent documents changes. Two limits belong on the record. The model names sit in a blog post dated November 2025 rather than on a maintained model page, so a buyer in September 2026 cannot tell whether they are current. And no provider is identified for hosting or infrastructure outside the model layer.

No pricing published at any level. No price, no range, no tier structure and no unit of charge, and no indication of whether cloud and on premise deployment price differently, which is a material question given the vendor presents on premise as a differentiator. An independent three month review states that the pricing concern is real while advising that productivity gains typically justify the investment, which is a reviewer characterizing cost without publishing a figure and is recorded as context rather than credited. Every route is a contact or demo request. Checked the home page, the solution pages, the pricing navigation and independent review material on 29 Aug 2026.

The commercial structure is published in real detail and there is no figure anywhere. There is no pricing page in the site navigation and every route to a number is a demo request, but the terms set out the shape of the deal more fully than most published price pages do: subscriptions are per user, one subscription entitles one named user and may not be shared, billing is annual in advance in euros, invoices are due net fourteen days, fees are non-refundable and non-cancellable, the subscription auto-renews unless canceled with sixty days written notice, the number of user subscriptions cannot be reduced during a term, overdue amounts accrue interest at one percent per month, and taxes are excluded. Section 6 adds that the number of accessible reports and the volume of downloads are limited by subscription tier and subject to fair use, with limits amendable without notice. A buyer can therefore describe the unit, the term, the currency and the commitment before contacting the company, and cannot estimate the cost. Tier names are not published and no enterprise or entry rate appears at any level.

THE FIRST RECORD IN THIS CATEGORY TO NAME PATENT OFFICES, which is the coverage question a prosecution practice actually asks. Multi office filing support is stated across the USPTO, EPO, CNIPA, PCT and KIPO among others, with the vendor stating that language and format are adapted to each jurisdiction's standards, and drafting output is matched to the user's jurisdiction as well as their style. Naming the offices lets a practitioner check whether their own filing route is supported, which no other record in this category permits. Technical domain coverage is enumerated across life sciences, chemistry, biotech, pharma, materials and software intensive systems, with the specific artifacts named that determine whether a tool is usable in those fields: Markush structures, sequences, drawings and experimental data. Buyer coverage spans law firms and corporate IP departments with published material addressing invention harvesting, patentability, portfolio pruning, competitive intelligence, prosecution and freedom to operate. Graded A because coverage is stated at the level a buyer verifies against their own docket. Held short of perfection because no prior art corpus scope, date range or update lag is published for the search side.

Coverage is described with real substance across two dimensions, and the boundary is left open. The vendor publishes five named use cases rather than a generic audience statement: patent workflows, research and science workflows, technology and IP strategy, legal marketing and business development, and remote work, and it names both corporate IP teams and law firms as users. Practice depth within patents is specific and unusual for its jurisdictional precision, with skills built around EPO and USPTO examination and analytics covering patent offices in Europe, the United States and Germany, plus examiner and opposition and appeal member insights that only exist where the underlying decision data does. That jurisdictional shape is itself the strongest coverage statement on the record. What is absent is the limit: nothing states which jurisdictions are not covered, no firm size is addressed, nothing says whether trademark, design or litigation work is in or out, and the marketing and business development use case sits oddly beside the patent depth without any statement of how far it extends.