DeepIP vs Solve Intelligence: 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.

Founded in 2023 to build this and nothing else, with domain expertise assembled deliberately: the team is described as patent attorneys, AI PhDs and software developers, and the product is an AI editor rather than an editor with AI in it. Every capability is model output: application drafting, claim drafting and amendments, office action responses, continuations and divisionals, invention harvesting from disclosure, claim charting, freedom to operate analysis, infringement and validity work. Style matching is model behaviour rather than templating, configuring output to an individual attorney's drafting voice and adapting for field, client and region. Remove the models and the product is a blank browser editor. Second consecutive A on this axis in ip-and-patents, and the category is tracking the plaintiff pattern: these vendors were built model first because the work was never automatable any other way.

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 colour coded confidence indicators.

A PUBLISHED EVALUATION METHODOLOGY, which is rare enough on this index to name, and no published result. The vendor states that large language model response correctness is measured through continuous human evaluation against real patent drafting tasks, assessing factual grounding. Three elements of that matter: the evaluation is continuous rather than a one time benchmark, it is conducted by humans against real drafting work rather than synthetic prompts, and factual grounding is named as the criterion. Almost every record on this index either publishes a number with no methodology or publishes neither; this vendor publishes the methodology and not the number. Grounding is also structural: output carries citations to source documents and integrated case law, so an attorney can trace an assertion. Held at B because nothing measured is published. No accuracy rate, no error rate on claim drafting or office action responses, no sample size, no evaluator identity, and no result of any kind from the continuous evaluation it describes. A vendor running continuous human evaluation has the numbers and has chosen not to publish them, which is a different and more interesting gap than not measuring at all.

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 behaviour, 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.

Oversight is built into the product model and into the company structure. The product is framed throughout as a Copilot operating inside a document editor where the attorney drafts, which places the human at the point of authorship rather than at the end of a pipeline, and style matching exists to make output conform to the attorney's own drafting rather than replace their judgement. Continuous human evaluation of model correctness is an oversight mechanism operating on the system itself. A Customer Advisory Board of senior patent professionals is stated to bring real prosecution and litigation experience into the product roadmap, which is governance of product direction by practitioners. Held at B because no boundary is published: nothing states what the system does unattended, whether any drafting or response can be filed without review, what confidence signal accompanies generated text, or what the continuous evaluation triggers when correctness falls. A copilot framing implies a pilot and never specifies what the pilot must do.

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.

The strongest named customer roster in the pull, published by the vendor rather than relayed by a directory. Named IP practices and corporate teams: DLA Piper, Siemens, Finnegan, BCLP, Troutman Pepper Locke, Haynes Boone, HGF, Bookoff McAndrews, Khurana and Khurana, Altacit Global, MKS and HG Law, against a stated base of more than 700 IP teams across six continents. Attributed comment with name and title: Richard Hodkinson, Chief Technology Officer at HGF, describing a competitive evaluation before committing. A Customer Advisory Board of senior patent professionals is named as a structure rather than a testimonial. Funding is dated and specific: a $12m Series A in April 2025 with Microsoft participating, a $40m Series B in December 2025, backers including Y Combinator and Thomson Reuters, and the acquisition of Palito.ai in March 2026. An outsider can identify the firms, the executive and the dates without contacting the vendor. Outcome claims are the weak part and the note records it: 60 to 90 percent efficiency and quality improvements, more than 50 percent productivity, and one firm exceeding billing targets by $1 million after reallocating AI saved time, all without methodology, baseline or sample. The grade rests on the named and checkable elements.

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 acknowledgement 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.

The most specific confidentiality architecture on the index, published as a structural claim rather than an assurance. Stated: data is sandboxed to individual users, the vendor has no access to or control over customer data, a zero trust and least privilege access model operates with regular vulnerability scans, penetration tests and access audits, and the vendor processes customer data solely on the customer's instructions without determining the purposes or means of processing, which is processor language drawn from data protection law rather than marketing. Encryption is named to the algorithm and protocol version, AES-256 and TLS 1.3. Customer selectable jurisdiction means confidential material can be kept where the client requires. The vendor publishes practitioner facing material specifically on how patent practitioners should evaluate AI tools for data security and confidentiality, which engages the professional question rather than only the technical one. Graded A because sandboxing to the individual user with a stated absence of vendor access is a stronger structural position than the tenant level segregation the B and A records elsewhere describe. Held short of perfection because the architecture is asserted rather than documented and no privilege specific attestation exists.

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.

Not located, and the absence is conspicuous against what this vendor does publish. The company is built by patent attorneys, maintains a Customer Advisory Board of senior patent professionals, sponsored and presented at the FICPI ABC Meeting 2026 on what is required for AI to be ready for patent practice, and publishes detailed guidance on evaluating AI tools for confidentiality. None of that engages professional responsibility: no statement on the practitioner's duty of competence over machine drafted claims, no positioning on who is responsible for a generated office action response filed under a registered practitioner's signature, no reference to USPTO Rules of Professional Conduct or 37 CFR, and no bar guidance. Compare Patlytics at B in this same category, which names Model Rule 1.6, 37 CFR, USPTO Rules and 35 U.S.C. section 102 explicitly. Checked the home page, the trust centre summary, the security and confidentiality guidance, the blog and the funding announcements on 29 Aug 2026.

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.

The most operationally specific AI governance disclosure in the pull, without the certification Patlytics holds. Published: policies aligned with ISO 42001 and ISO 27001 and reviewed at least twice yearly, a stated compliance roadmap toward ISO 42001, an undertaking to comply with the European Union Artificial Intelligence Act as both AI Provider and Deployer, and continuous human evaluation of model correctness against real drafting tasks assessing factual grounding. Naming its own role under the EU AI Act as both provider and deployer is a legally consequential self classification that determines which obligations attach, and no other record on this index makes it. Continuous evaluation is a live governance process rather than a policy document. Held at B rather than A because none of it is certified or evidenced: ISO 42001 is a roadmap rather than a certification, no evaluation result is published, no bias or fairness testing appears, no model card exists, and no named governance body was located. Compare Patlytics at B with the ISO 42001 certificate and no published output. Two routes to the same grade: one holds the certificate and publishes nothing from it, the other publishes the process and holds no certificate.

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 prioritised 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.

The most complete stewardship position on the index, on all three limbs at once. Training: no data uploaded to or output from the product is ever used to train any AI model of any kind, which is the broadest formulation encountered, covering output as well as input and admitting no carve out for third party, public, shared or internal models. Retention: customers control zero data retention settings for the large language model providers and models used across the platform, which is customer exercisable rather than a policy to trust and reaches the downstream model layer. Access: data is sandboxed to individual users, neither the vendor nor any third party monitors it, and the vendor states it has no access to or control over customer data. Supporting controls are specific to the algorithm and protocol version. Graded A because it answers what the vendor does with data, what the downstream providers do with it, and who can see it, with a customer control on the retention limb, and no other record on this index answers all three. Held short of perfection because it is policy and product configuration rather than contractual terms located in this pass, and no retention default is stated for customers who do not configure it.

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.

No published position located. Nothing was found on liability for AI output, warranty, service levels or remedy. The exposure in this category is severe and specific: a generated claim set that is too narrow loses scope permanently once granted, an office action response that mischaracterises prior art can create prosecution history estoppel, and a missed reference in a freedom to operate analysis can expose a client to infringement liability. This vendor publishes more about how it protects data than any peer and nothing about what happens when its output is wrong. Checked the trust centre summary, the security and confidentiality guidance, the home page, the blog and the site navigation on 29 Aug 2026. Enterprise agreements govern this and are not public.

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.

Nothing located, and the product architecture explains why without excusing it. Solve is an in browser document editor, and independent comparison material draws the contrast explicitly, describing a competitor as appearing in a Microsoft Word sidebar so attorneys need not leave their existing workflow, and positioning Solve as the browser based alternative. No Word or Office integration, no IP docketing system connector, no patent management platform integration, no document management system and no API were located. Docketing integration is the material gap for a prosecution product, since deadlines, annuities and family relationships live in the docketing system. One March 2026 event is recorded and not credited as an integration: the acquisition of Palito.ai, described as unifying patent litigation and prosecution in one platform, which is consolidation of the vendor's own products rather than connection to a customer's estate. Checked the home page, the product pages, the acquisition announcement and independent comparison material on 29 Aug 2026.

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 licence 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.

THE BEST RESIDENCY DISCLOSURE ON THE INDEX, and the only one that is a customer control rather than a vendor statement of fact. Published: customers choose where their data is stored and processed, with the United States and Europe named as examples, and the vendor states the data never leaves the chosen jurisdiction. That is selectable residency with an exclusivity commitment attached. It extends down the chain: based on customer selection, only United States or European Union based subprocessors are engaged for data processing, which closes the gap that undermines most residency claims, where storage is regional and processing or inference quietly is not. The vendor states the purpose plainly, being to help enterprises meet internal and regulatory requirements. For patent work this is not a preference: unpublished applications engage foreign filing licence requirements and export controls, and a practitioner may be unable to let an application leave a jurisdiction at all. Graded A because it is checkable, actionable and complete across storage, processing and subprocessors. Held short of perfection because no hosting provider is named, no full jurisdiction list is published beyond examples, and no single tenant option is described.

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 centre, 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.

CHECK THE NOUN. A live trust centre operates at trust.solveintelligence.com and SOC 2 Type II certification is stated consistently, which alone would sit high on this axis. But the vendor's certification claims are not consistent between its own pages, and the difference is material. The home page states adherence to standards including SOC 2 Type II and ISO 27001 certification, as well as ISO 42001, GDPR and CCPA requirements, which a reader would take as three certifications. The vendor's own detailed security guidance says something weaker and more precise: SOC 2 Type II certified, with controls aligned with ISO 27001, the NIST Cybersecurity Framework and CSA CCM, and policies aligned with ISO 27001 and ISO 42001 standards, and it separately describes a compliance roadmap that builds on preparing for ISO 42001. Aligned with is not certified to, and preparing for is not holding. On the detailed page's own account this vendor holds one certification, not three. Only SOC 2 Type II is credited. Held at B rather than A on that basis and on the absence of certificate dates, examination periods or a named auditor. Contrast Patlytics in this same category, which states ISO 42001 certification without the qualifying language and is graded on that claim.

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.

The chain is disclosed structurally and no party in it is named, which is the inverse of the usual gap and is more useful than it sounds. Published: the platform uses large language model providers, plural, confirmed by the existence of customer controlled zero data retention settings for those providers and models; only United States or European Union based subprocessors are engaged for data processing, based on customer selection; and no uploaded data or output is used to train any model of any kind. A buyer therefore knows that third party models are involved, where they may be located, that the customer chooses, and that a zero retention posture can be enforced against them. That is more actionable for a procurement review than a provider name with no accompanying controls. Held at B rather than A because no provider, model family or version is named and no subprocessor list was located outside the trust centre, which was not entered in this pass. Compare Onspring at B for naming Anthropic and saying nothing about handling: two records at the same grade from opposite disclosures.

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 characterising 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.

No pricing located in vendor material at any level. No price, no range, no tier structure and no unit of charge, and no indication of whether the platform prices per seat, per application drafted, per office action or per chart. An independent legal software directory reports a starting price of $199 per month, which is a specific figure suggesting a published pricing page exists that was not reached in this pass, and it is recorded as third party context rather than credited as disclosure. Flagged as a correction candidate in the upward direction on that basis. The gap is notable against the rest of this record, which is among the most disclosure forward in the pull on every other dimension. Checked the home page, the product pages, the pricing navigation, the trust centre summary and independent directory material on 29 Aug 2026.

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.

The most completely characterised coverage statement in the pull, across three dimensions rather than one. Technical domains are enumerated with the specific artifacts each requires: life sciences with biological sequences and chemical structures, telecoms and standard essential patents at volume, software and electronics from machine learning architectures to chip designs, and mechanical work with native CAD file support, auto generated figures and labelling. That is a claim a practitioner can test against their own docket, because a platform that cannot handle a sequence listing or a CAD figure is disqualified for their work regardless of drafting quality. Workflow coverage runs invention harvesting, drafting, continuations and divisionals, office action responses, claim charting, freedom to operate and clearance, infringement, validity, standard essential patent mapping and litigation. Geographic coverage is stated as more than 700 IP teams across six continents with US, Europe and Asia named and offices in New York and Munich, and drafting is configurable by region. Graded A because coverage is stated at the level a buyer checks rather than as a count. Held short of perfection because no patent office list, corpus date range or update lag is published for the search and charting side.

Silent, and the distinction being drawn here is deliberate. The quoted commitment is real, repeated across funding announcements and product pages, and it is a retention commitment: data is not kept. It is not a statement that data is not used to train models while it is present. Those are different questions and this index has separated them consistently. 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 prioritised when assessing tools in this field, which conflates the two concepts and is criteria language applied to a market rather than a commitment made about this product. Under the standing rule that an ambiguous claim earns nothing, and that this must hold when it costs a grade, the no training position is not credited. Recorded as silent rather than as a negative commitment: zero retention makes training on retained data substantially harder and does not exclude training on data in process. Flagged as a correction candidate in the upward direction, since a direct no training statement may exist on a surface not reached in this pass. Checked the home page, the law firm solution page, both funding announcements and the blog on 29 Aug 2026.

Policy never, and the broadest formulation encountered in the pull. The quoted commitment admits no carve out: it covers any AI model of any kind, which excludes the third party scoping used by LinkSquares, the shared model scoping used by Eve, the product layer scoping used by Exterro and the consent or necessity qualifier used by Patlytics. It also covers output as well as input, stating that no data uploaded to or output from the product is ever used, which closes the gap where generated work product is treated differently from source material. It is reinforced by two adjacent commitments rather than left standing alone: neither the vendor nor any third party monitors customer data, everything being sandboxed to the individual user, and customers control zero data retention settings for the underlying model providers, which pushes the position down the chain to parties the vendor does not control. Held at policy never rather than contractual never because the statement appears on product and security pages rather than in terms or a data processing agreement located in this pass. Checked the home page, the trust centre summary and the security and confidentiality guidance on 29 Aug 2026.

Disclosed fixed at zero. The vendor states a strict Zero Data Retention policy, published consistently across dated funding announcements, the law firm solution page and independent coverage, and describes it as ensuring complete data segregation and confidentiality. A stated zero is the floor on this signal and requires no configuration by the customer to achieve. Held at disclosed fixed rather than the customer configurable value because the policy is presented as the vendor's standing posture rather than a setting the customer controls, which distinguishes it from Solve Intelligence where the customer sets zero retention at the model provider layer. Two boundaries are not published and are recorded here: the scope of the policy is not defined as between the platform, the model layer and the Word add in, and nothing states how drafts in progress are held during a working session.

Customer configurable at zero, and the only record in the pull to reach this value. The vendor states that customers control zero data retention settings for large language model providers and models across the platform, described as giving full control of the data. Two things distinguish this from every other retention disclosure encountered. It is a customer exercisable control rather than a vendor policy the customer must accept, and it operates at the model provider layer rather than only within the vendor's own systems, which is where retention actually persists and where most vendors are silent. Compare Exterro at disclosed fixed with no storage stated for its AI layer, DigitalOwl at customer configurable through manual deletion, and Patlytics at disclosed fixed with a 90 day history window. Held at this value rather than higher because no default is published for customers who never configure the setting, and no retention period is stated for data held in the platform itself as distinct from the model providers.

Claimed and not documented. The vendor states complete data segregation and confidentiality alongside encrypted storage and its zero retention posture, and a published customer quote describes extensive due diligence on the security infrastructure before selection, so segregation is asserted and has been examined by at least one buyer. Nothing documents the mechanism: no statement of whether segregation operates between customers, between users within a firm, or at matter level, and no description of how the Word add in bounds access when an attorney has several clients' applications open in the same environment. The conflicts scenario in patent practice is concrete, since a firm may prosecute for competitors in the same technical field, and no published material addresses it. Compare Solve Intelligence, which states sandboxing to individual users, a finer claim at the same level of documentation.

Claimed and not documented, at a granularity no other record claims. The vendor states that all data is sandboxed to individual users and that it has no access to or control over customer data, operating a zero trust and least privilege access model with regular access audits. User level sandboxing is finer than the tenant level isolation described elsewhere on the index and, if implemented as stated, would address the conflicts scenario directly, since two attorneys in one firm on opposing matters would not share a data space. What is not documented is any of the mechanism: no description of how sandboxing is enforced, no statement of what happens where a team needs shared access to a single matter, which the vendor elsewhere describes as a collaboration feature, and no account of how the boundary interacts with the Customer Advisory Board or with continuous human evaluation against real patent drafting tasks, which necessarily involves people examining work product. That last tension is unaddressed and is the question a firm should raise.

Not addressed. No government or law enforcement request clause, no commitment to notify a customer before producing their data, and no transparency report were located. The zero data retention posture is materially relevant here and is not a substitute: a vendor holding nothing has less to produce, which reduces exposure without constituting a notice commitment, and nothing states what would happen regarding data in process or held under an on premise arrangement. The vendor operates from New York and Paris, so requests could arrive under two legal regimes, and nothing addresses either. Checked the home page, the solution pages, the funding announcements and the site navigation on 29 Aug 2026.

Not addressed. No government or law enforcement request clause, no commitment to notify a customer before producing their data, and no transparency report were located. The vendor's statement that it processes customer data solely on the customer's instructions and does not determine the purposes or means of processing is processor positioning under data protection law and is not a notice commitment. The stakes are the category's own: the platform holds unpublished applications and pre filing disclosures, and disclosure of an unpublished application to a third party can bear on novelty and prior disclosure. Checked the trust centre summary, the security and confidentiality guidance, the home page and the site navigation on 29 Aug 2026.

Named at the level of category and nothing beyond it. The vendor states that the platform combines proprietary models, third party tools and exclusive data access, and separately that custom models are trained on patent data and that relevant prior art is surfaced automatically. Exclusive data access is an unusual thing to claim and it is a provenance statement of a kind, since it asserts privileged rights to something. Nothing identifies what: no source is named, no patent office or database is credited, no licensing basis is stated, no jurisdictional scope or date range is given for the prior art corpus, and no update lag is published despite multi office filing support across the USPTO, EPO, CNIPA, PCT and KIPO implying substantial data coverage behind it. A practitioner relying on an automatic prior art surface cannot determine what it searched. Compare Patlytics, which states more than 50 million global public patents with stated verification and update processes.

Not addressed. The product integrates case law into office action responses and supports prior art work, freedom to operate analysis and validity challenges, all of which require a patent and legal corpus, and nothing published names a source, a licensing basis, a jurisdictional scope, a date range or an update lag for any of it. That is a wider gap than for the drafting side, where the corpus is the customer's own disclosure material. Contrast Patlytics in this same category, which states models fine tuned on more than 50 million global public patents with regular verification and update processes. This vendor describes its security posture in far greater detail than its data sources. Checked the home page, the product pages, the Charts announcement and the blog on 29 Aug 2026.

Not addressed, and applicable on both limbs this category presents. The product generates office action response suggestions, which in prosecution involve characterising cited references and often citing legal authority, and nothing states whether authority is checked for current treatment. The patent specific analogue also applies through automatic prior art surfacing and patentability analysis, where the equivalent question is legal status: whether a surfaced reference is a granted patent still in force, an abandoned application, or a document whose claims were amended or invalidated in post grant proceedings. Nothing published addresses either. Third consecutive record in this category to leave this signal unanswered, after Patlytics and Solve Intelligence. Checked the product pages, the patentability page, the blog and the home page on 29 Aug 2026.

Not addressed, and applicable on two counts rather than one. The product integrates case law into generated office action responses, so the ordinary currency question arises: nothing states whether cited authority is checked for current treatment or whether a superseded decision would be flagged before it reaches a filing. The patent specific analogue also applies through the Charts product covering freedom to operate, infringement and validity work, where the equivalent check is legal status, being whether a patent is in force, lapsed, amended in reexamination or invalidated in post grant proceedings. Neither is addressed. Second record in this category to leave this signal unanswered, after Patlytics. Checked the product pages, the Charts announcement, the home page and the blog on 29 Aug 2026.

Not addressed, and an independent review reports the failure mode the vendor does not describe. Nothing published states whether a low confidence generation is flagged, whether the system declines where a disclosure is insufficient to support a claim, or what signal accompanies an office action response suggestion the model is unsure of. An independent three month review of this product states that an AI hallucination issue exists and is manageable with proper review processes, which confirms the behaviour occurs while placing the entire burden of detecting it on the attorney. That is a reasonable expectation of a patent practitioner and it is not a substitute for the product indicating where it is uncertain. Compare Patlytics at documented in this category, on colour coded confidence indicators surfaced to the reader. Checked the home page, the product pages, the blog and independent review material on 29 Aug 2026.

Not addressed, with a measurement process disclosed and no behaviour described. The vendor states that model correctness is continuously evaluated by humans against real patent drafting tasks assessing factual grounding, which establishes that the vendor knows where its output is weak. Nothing published describes what the user sees at the point of use: no confidence indication accompanying generated claims or office action responses, no flag on a low grounding output, no statement of whether the system will decline to draft where the disclosure is insufficient, and no description of what the continuous evaluation triggers when correctness falls. Contrast Patlytics at documented in this same category, which surfaces colour coded confidence indicators to the reader. Measuring correctness internally and showing the user nothing are different things, and only the first is disclosed here.

None located, with the instrument named. General web searches combining the vendor and product names with court, order, sanction, fabricated citation and patent terms returned nothing on 29 Aug 2026, and no named docket database, USPTO record system or court record tracker was searched. Recorded as a statement about what this search found, not as a clearance. The exposure shape is a mischaracterised prior art reference or a defective citation inside an office action response filed at a patent office, which would surface in a prosecution file wrapper rather than in a published court opinion, and the independent report of a hallucination issue makes the search worth repeating with a proper instrument on a later pass.

None located, with the instrument named. General web searches combining the vendor and product names with court, order, sanction, fabricated citation and patent terms returned nothing on 29 Aug 2026, and no named docket database, USPTO record system or court record tracker was searched. Recorded as a statement about what this search found, not as a clearance. The exposure shape in this category is distinctive: the analogous failure is a fabricated or mischaracterised case citation inside an office action response filed at a patent office, or prior art mischaracterised in an invalidity contention, and both would surface in a prosecution file wrapper or in litigation rather than in a court opinion. Flagged as worth a proper instrument on a later pass given stated adoption across more than 700 IP teams.

Not addressed. No named ethics opinion, no USPTO Rules of Professional Conduct reference, no 37 CFR citation, no duty of competence discussion and no bar guidance was located. The vendor publishes a substantial buyer education library covering evaluation criteria, drafting workflows and best practices, and that material addresses security, output quality and workflow fit without reaching the professional rules that govern the practitioner signing the filing. Second of three records in this category at this value, with Patlytics the exception at generic reference. Checked the blog library, the home page, the solution pages and the site navigation on 29 Aug 2026.

Not addressed. No named ethics opinion, no USPTO Rules of Professional Conduct reference, no 37 CFR citation, no ABA Formal Opinion 512 and no state bar guidance was located. The absence stands out because the vendor engages the adjacent question seriously, publishing detailed practitioner guidance on evaluating AI tools for data security and confidentiality and presenting at a professional congress on what AI readiness for patent practice requires. That material addresses the technical and contractual dimension thoroughly and the professional rules dimension not at all. Contrast Patlytics at generic reference in this same category, which names Model Rule 1.6, USPTO Rules, 37 CFR and 35 U.S.C. section 102 directly. Checked the security and confidentiality guidance, the blog library, the home page and the trust centre summary on 29 Aug 2026.

Savings claims only, corroborated independently at a lower figure than the vendor states. Vendor published: up to 50 percent of drafting time saved. Customer quoted in vendor material: approximately 20 percent improvement in drafting and prosecution efficiency during a trial. Independent three month review: 40 to 60 percent reduction in initial drafting time across patent types. The spread between the vendor's own customer quote at 20 percent and its headline at 50 percent is worth recording, since both appear on vendor surfaces and neither carries methodology, baseline or scope. Nothing appears on the client's side of the equation: no position on how AI assisted drafting time should be recorded on an invoice, and no exportable record showing what portion of an application or office action response was machine generated, which matters for a product sold to firms billing prosecution hourly to corporate clients who increasingly ask.

Savings claims only, with one published figure that reaches further than efficiency. Standard claims: 60 to 90 percent efficiency and quality improvements, and more than 50 percent productivity. The unusual one is a stated case where a firm adopting the platform mid fiscal year exceeded its billing targets by $1 million by reallocating AI saved time to higher value work, which is a claim about firm revenue rather than firm effort and is the only one of its kind in the pull. It carries no methodology, baseline or firm identity. Nothing appears on the client's side of the equation: no position on how AI assisted drafting or prosecution time should be recorded on an invoice, and no exportable record showing what portion of an application or office action response was machine generated. That omission is material for a product sold into firms that bill patent prosecution by the hour and whose corporate clients increasingly ask.

Not addressed. Certifications are published openly and are citable, being SOC 2 Type II, ISO 27001 and GDPR compliance alongside a stated Zero Data Retention policy and named Azure hosting, so a firm has real content for a client questionnaire. No route to anything underneath was located: no trust centre, no security page, no request path for the SOC 2 report, no subprocessor list, no named model provider despite the vendor confirming third party tools are in use, and no data processing agreement. A firm can repeat the claims and cannot obtain a document. Same position as Patlytics in this category, which publishes more and also offers no route, and behind Solve Intelligence which operates a live trust centre. Checked the home page, the solution pages, the funding announcements and the site navigation on 29 Aug 2026.

On request, through a live trust centre, and the pack behind it is the most substantive in the pull. The trust centre at trust.solveintelligence.com is the destination, and the disclosures a firm could put in front of a client are unusually complete: SOC 2 Type II certification, encryption named to algorithm and protocol version, zero trust and least privilege access with regular penetration testing and access audits, an unqualified no model training commitment covering input and output, customer controlled zero data retention at the model provider layer, customer selectable storage and processing jurisdiction with data stated never to leave it, a commitment to engage only United States or European Union based subprocessors, and a stated undertaking to comply with the EU AI Act as provider and deployer. Held at on request rather than higher because the trust centre gate was not entered in this pass, no subprocessor list was located outside it, no model provider is named, and no certificate dates or auditor are published. A firm has an unusually strong set of claims and still cannot obtain the underlying documents from public surfaces.

Partial record, and this vendor makes traceability an explicit argument rather than an incidental feature. The stated position is that workflow integrations with Microsoft Word and leading IP management systems preserve audit trails by design, where legacy software requiring manual exports erodes traceability, and the vendor's own buyer guidance names versioning and audit trails for internal review and compliance as a criterion to demand. Drafting inside Word means the document's own revision history captures the work, which is a real and unusually practical answer to the process limb. The gaps are the familiar two, and the first is sharpened by the architecture: nothing indicates that output is marked or recorded as machine generated, so a Word revision history shows edits without distinguishing which passages the model produced, and no human verification record is captured showing that a practitioner reviewed and adopted generated claim language before filing. The forum here is a patent office file wrapper rather than a court, and the same question applies.

Partial record. The source limb is answered: output carries citations to source documents and integrated case law, so an examiner, an opposing party or a tribunal can be shown what a generated passage rests on, and in the Charts product a claim chart maps evidence to limitations element by element, which is inherently inspectable. The other two limbs are absent. Nothing indicates that output records which model produced it or when, which matters more here than in most categories because the vendor discloses that multiple large language model providers are in use and that customers can select jurisdictions and retention settings, so the processing path for a given document is configurable and therefore variable. And no human verification record is captured, so a practitioner who reviewed and adopted a machine drafted claim set or office action response cannot evidence that they did. For prosecution work the audience is a patent office rather than a court, and the same gap applies.