Paximal vs PowerPatent: how they compare in 2026
Paximal and PowerPatent both draft full patent applications from invention materials for patent attorneys and agents, with a practitioner expected to finish every draft. PowerPatent sits in the top two bands on ten of fifteen axes and Paximal on nine of fifteen, identical on eight. PowerPatent's lead is in its agreement. Its terms discard patent documents immediately after results are returned, keep processing on US computers at Google Cloud or AWS, and bar OpenAI from training on the content. It publishes a price of $199 for 200 tokens, which it says covers most applications. Paximal's lead is evidence from the patent office. It publishes a cohort study of applications it drafted, reporting no enablement rejections and about 70 percent allowed after one office action. It reaches its models only through Microsoft Azure and names users, including Wood IP and Xsensus. Paximal's commitments on training and confidentiality sit on its web pages, and its terms require users not to assert patents derived from its output against it.
At a glance
All 15 axes, side by side
The same grid applied to every vendor in the index, graded from public sources. Hover a grade to see what the letter means on that axis.
AI Centrality
How much of the product is actually AI. Whether the machine learning is the mechanism the buyer is paying for or a feature layered onto conventional software, and whether the vendor is specific about which is which.
The models are the product and there is nothing underneath them. What a customer buys is generation: orchestrated agents scaffold the sections of a patent application, harmonize terminology across the document, map claimed features to the specification and embed fallback support, assembling a complete draft from invention materials rather than filling a template. The vendor's own framing is agentic end-to-end drafting, prompt-free, with the practitioner setting claim focus, terminology and embodiments up front and the agents doing the writing. Remove the models and nothing remains to sell: there is no document management layer, no docketing, no portfolio system and no database product beside it. The only non-generated component named on the estate is a preliminary prior art search report, and that content comes from a third party under that party's own terms rather than from the vendor's own models. Verified 20 September 2026.
The generative drafting is what a buyer pays for, on top of a drafting workspace that would still work without it. From a summary, a claim set and annotated drawings, the software writes the background, a preliminary set of claims and a detailed description tied figure by figure to the part lists, and the vendor's own framing is a first-draft co-pilot: 'we write the story' from the data gathered. Underneath sit invention disclosure capture, flowchart and drawing management, graphical claim drafting, collaboration with inventors and Private PAIR integration, which are workflow tools in their own right. Verified 20 September 2026.
Citation Accuracy and Hallucination Disclosure
Whether the vendor publishes measured accuracy on citations and assertions, grounds output to primary sources, and says plainly what its system does when it does not know. Legal has a documented public record of fabricated citations reaching filed briefs, so an untested claim of accuracy is not evidence.
Outcome is measured and the generation risk is never addressed. This product writes specification text rather than citing authority, so the failure that matters is unsupported description, a claim the specification does not enable, or terminology that drifts between claim and body. Nothing published speaks to any of that: there is no hallucination disclosure, no statement about how the agents are constrained to the materials supplied, and no error rate for generated content. What exists instead is prosecution evidence, a published cohort of the vendor's own drafted applications reporting no section 112(a) enablement rejections and roughly 20 percent with minor section 112(b) clarifications promptly cured, which is a downstream signal rather than a measurement of the generation. The published terms disclaim accuracy entirely and disclaim liability for actions taken on the contents. The search report carries its own warning that it is informational and not a legal opinion. Verified 20 September 2026.
Accuracy is claimed repeatedly and measured nowhere. The site promises unparalleled accuracy and quality, says the drafts adhere to patent office guidelines, and offers two figures in a demo pitch, five or more hours saved per application and Section 112 issues caught 30 percent more often, neither attributed, dated or explained. The one real check described is internal: diagnostics that flag Section 112 problems, enablement, written description and indefiniteness, and gaps in claim support between the specification and the claims. Against that, the terms state plainly that PowerPatent makes no representation that the service will be accurate or reliable and is not responsible for errors in processing. Verified 20 September 2026.
Autonomy and Oversight Model
What the system decides on its own, what a lawyer must approve, and whether the vendor documents where the review point sits. A tool that drafts under review and a tool that files without one are different products and different risks.
The supervising practitioner is named as the control point, repeatedly and in the vendor's own voice. The security page states that every draft stays under the supervising practitioner's control from input to filing-ready. The trust center describes the service itself as one through which patent application documents are auto-drafted under the supervision of a patent practitioner user, which places the supervision inside the definition of the product rather than in a disclaimer. The workflow supports it: attorneys set the target for protection, the primary independent claim, the initial dependent claims, terminology and embodiments before generation, and the solutions pages describe drafts arriving structured and ready for attorney review, attorney-led throughout. What is missing is the categorical form: the commitment is a blanket statement across the whole service rather than a constraint attached to a named mode or tier, no threshold or escalation is described, and nothing states what the agents may not do. Verified 20 September 2026.
The product is built as a first draft for a qualified person to finish, and the agreement says who that person must be. The eligibility clause in the terms requires a user to represent that they can practice before a national patent office in their jurisdiction, which is a written limit on who may operate the drafting at all. The workflow puts the human at the end of every step: preliminary claims are generated for user review, the description is generated for editing, and the diagnostics report issues for a person to resolve before filing. What is not published is any internal limit: no confidence threshold, no statement of what a generated draft may not be used for, and nothing on what happens when the model invents support that is not in the drawings. Verified 20 September 2026.
Operational and Outcome Evidence
Named, dated evidence that the product works in production at real firms or legal departments. Case studies with figures and identified customers count. Unattributed testimonials and launch announcements do not.
A real cohort study against the public file wrapper, and the cohort is anonymous. Case Study 01 reports on accelerated applications drafted by the vendor and prosecuted at the USPTO: roughly 10 percent allowed on first action, about 70 percent allowed after one office action and the remainder after a second; roughly 65 percent drew section 102 or 103 rejections at first action, of which about 80 percent were traversed by targeted claim amendment; no section 112(a) enablement rejections, about 20 percent with minor section 112(b) clarifications; and average preparation time of around four hours from inventor materials to filing-ready. That is measured against records anyone can check. What is absent is the join: the cohort is not named, no sample size appears, the figures are approximations, and no named customer carries a figure. Named users exist separately, a testimonial from the founder of Wood IP LLC, a second case study with RapidSOS behind an email form, and an announced selection by Xsensus in June 2026. Buyers should note that the terms take a license to use a customer's name and logo on registration. Verified 20 September 2026.
Logos and praise stand in for deployment evidence. The homepage carries a Trusted By strip and recognition images, and states that many law firms, corporations and inventors rely on the product, without naming one of them in text; the testimonials that appear are unattributed. The concrete items are awards rather than deployments: finalist for World IP Review's LSPN Awards USA 2025 in Technology Solution of the Year, and completion of SOC 2 Type 2 certification announced in a press release on the site. The two performance figures quoted in the demo pitch carry no customer, date or method. Verified 20 September 2026.
Privilege and Confidentiality Posture
How client confidences are handled: attorney client privilege and work product treatment, segregation of one client matter from another, whether client data trains any model, and what the vendor commits to in writing rather than in marketing.
The commitments are specific and readable in advance, and none of them is contractual. The trust center states that the vendor does not use a customer's confidential information except to provide the service, and does not directly or indirectly, without authorization, reveal, report, publish, disclose or transfer it to any third party. Inputs, meaning invention materials, and outputs, meaning the application documents, are stated not to be available to other customers, not to train any AI model, and not to be used to improve any Microsoft or third-party model. All data is stored and processed in the United States, with physical, electronic and procedural safeguards maintained. For unfiled invention material, which loses novelty on disclosure, those are the right commitments. The gap is where they live: the only published agreement is a set of website terms carrying no confidentiality clause at all, no obligation of non-disclosure, no return or destruction on termination and no notice before compelled disclosure. No data processing agreement or subprocessor list was located. Verified 20 September 2026.
For a product that handles unfiled applications, the published commitments are specific and they sit in the agreement rather than on a marketing page. The terms state that once results are provided the patent documents are immediately discarded, with no storing, caching or logging of their content; that processing runs on Google Cloud Platform or Amazon Web Services exclusively on computers in the United States; that portions of a document are transmitted to OpenAI only with the user's authorization; that no data is shared with OpenAI for model training or improvement; and that PowerPatent will not disclose document content or results to any third party other than those processors. The privacy policy of 11 July 2026 adds that some services involve privileged and confidential patent material and describes safeguards for it. What is absent is separation between customers and any treatment of privilege or work product in what the account retains, and the same policy tells users that they remain responsible for judging whether a workflow suits the sensitivity of what they submit. Verified 20 September 2026.
UPL and Professional Responsibility Posture
Whether the vendor is clear that it supplies a tool rather than legal advice, who its audience is, and how it addresses unauthorized practice of law, competence and supervision duties, and jurisdiction limits. ABA Formal Opinion 512 is the reference point. Where the advice line is not the duty a product raises, the axis is read through the nearest professional duty it does raise: judicial conduct rules and the reviewing duty for products sold only to courts, and the duty to bill for time actually spent for products that draft time entries.
The advice line is drawn in the published terms and the supervising practitioner is named as the person who holds the work. The terms state that communication through the site does not create or constitute an attorney-client relationship, is not intended to convey or constitute legal advice, and is not a substitute for obtaining legal advice from a qualified attorney. The preliminary search report clause repeats it in product terms: those reports are informational only, do not constitute legal advice or a legal opinion, and the user should consult qualified counsel before acting on search results. The audience is unambiguous throughout, patent attorneys and agents, in-house IP teams and solo practitioners, with no consumer-facing surface, and the service is defined as auto-drafting under the supervision of a patent practitioner user. What holds it below the top band: nothing names a practice jurisdiction limit, and the competence duty is engaged only through buyer-facing guidance rather than as the vendor's own commitment. Verified 20 September 2026.
The agreement states who may use the product, which few vendors do. Under the eligibility clause a user represents that they are at least 18 and can practice before a national patent office in their jurisdiction, and a separate clause states that nothing communicated through the service creates an attorney-client relationship, is a solicitation, or conveys legal advice or a substitute for it. The marketing points the other way: the contact form asks whether the visitor is a patent attorney or an inventor or founder, and the pitch to founders is first-to-file protection in days. Nothing published addresses how the two fit together, what an unrepresented inventor should understand about filing a machine-drafted application, or how a supervising attorney should check one. Verified 20 September 2026.
AI Governance and Bias Disclosure
Published governance over model behavior: who owns it inside the vendor, what is tested before release, and what is disclosed about disparate output across matter types, parties, or populations.
One improvement mechanism is published and no governance sits around it. The platform page states that drafting patterns are continuously tightened from the real prosecution signals that filed applications generate, which is a described feedback loop and an unusual thing to disclose. A buyer should pin down what feeds it, because the estate also commits that customer inputs and outputs never train or improve any model; the two reconcile if the signals come from published file wrappers, which are public records, but nothing published says so. Beyond that there is nothing: nobody is named as accountable for model behavior, no pre-release testing or evaluation regime is described, no model card or change log exists, and no disclosure addresses whether output quality differs across technology areas, art units or drafting styles. On a product whose value claim is that its drafts survive examination, the absence of any published evaluation method behind that claim is the gap. Verified 20 September 2026.
Nothing published addresses how the drafting models are built, tested or controlled. There is no responsible AI page or policy, nobody is named as accountable for model behavior, nothing describes what is evaluated before a release ships, no results are published, and nothing addresses whether draft quality varies by technology area or by the language a specification is written in, which the vendor advertises as a strength. The published AI material is about data handling rather than governance. Checked the homepage, the How It Works and pricing pages, the terms, the privacy policy and the press page on 20 September 2026. Verified 20 September 2026.
AI Safety and Data Stewardship
Retention, deletion, access control, and what happens to prompts and documents after they are processed. Whether the vendor states its subprocessors and its incident practice, or leaves the buyer to assume.
The training and residency questions are answered well and most of the rest of the axis is untouched. Published: all data stored and processed within the United States; inputs and outputs never used to train or improve any model, the vendor's or Microsoft's; materials never available to other customers; no disclosure or transfer to third parties without authorization; physical, electronic and procedural safeguards; SOC 2 certification. What is absent is everything a security reviewer would ask next. No retention period is stated for invention materials or generated drafts, and nothing describes deletion on request or on termination. No subprocessor list is published beyond the Microsoft Azure route. No incident response or breach notification practice appears anywhere, in the terms or on the security page. No encryption standard, access control model or logging detail is given. The trust center that would carry the detail is an application that did not render, and one search recovered nothing from it. Verified 20 September 2026.
Most of what this axis asks for is published across two documents. The terms give the retention answer outright, immediate discard of patent documents with no caching or logging, and name the processors: Google Cloud Platform or Amazon Web Services, with all patent document processing on computers in the United States, and OpenAI for generation. The privacy policy of 11 July 2026 describes access controls, encryption, logging, monitoring, network protections and vendor management, sets out retention by service and setting, and covers international transfers. Missing are an incident notification commitment to customers, a named subprocessor list beyond the three providers in the terms, and any reconciliation of the immediate-discard promise with the policy's statement that user content is retained as long as reasonably necessary. Verified 20 September 2026.
AI Liability and Recourse
What the vendor stands behind contractually when its output is wrong. Indemnities, caps, carve outs, insurance, and whether any of it is published or only reachable through a negotiated agreement.
No position on liability is published, and the only allocation of risk in the agreement runs toward the vendor. The single published legal document is a set of website terms of use. They disclaim all warranties, provide the site as is without warranty of accuracy, and state that the vendor disclaims all liability for actions taken or not taken based on any or all of the contents. There is no indemnity, no liability cap, no warranty of any kind for the service, no service level and no remedy described for a defective draft, which matters because the output of this product becomes a filed patent application whose defects surface years later as a section 112 rejection or an unenforceable claim. No subscription agreement, order form or master agreement is published anywhere. The terms also extract a covenant running the other way: a user agrees not to assert, and not to permit others to assert, any patent rights based on or derived from outputs generated using the service against the vendor. Verified 20 September 2026.
The published position allocates the entire risk of a bad draft to the user. The terms include an assumption of risk clause stating that the user bears all risk, that PowerPatent holds no liability for outcomes, that it is not responsible for errors or inaccuracies in the processing of patent documents, and that the user is solely responsible for the content of their patent documents; warranties are disclaimed in full, liability is capped at the fees paid in the preceding twelve months, the user indemnifies PowerPatent, and disputes go to individual arbitration with a class waiver. No indemnity runs to the customer, no warranty attaches to the generated text, and no insurance position appears, which matters where a defective specification can cost a priority date. Verified 20 September 2026.
Practice Systems Integration Depth
How deeply the product reaches into the systems legal work already lives in: document management such as iManage and NetDocuments, Word and Outlook, contract lifecycle management, matter management, e-billing, and court filing systems.
The product stands alone and work has to be carried into it and out of it by hand. No integration with any practice system is published: no document management system, no docketing or IP management platform, no matter management, no e-filing path to the USPTO, no word processing round-trip and no e-signature. No API, developer documentation or connector directory was located. What the estate does describe is the input side, and only as uploads: invention disclosures, technical specifications, design documents, system diagrams, source code excerpts, white papers, slide decks, lab notes and example use cases, assembled in the platform. The output is described as a filing-ready draft without naming a format or a destination system. One external service is named, Project PQAI for preliminary prior art search, which is a content source rather than a practice system. The absence was checked against the navigation, the solutions pages, the pricing page and the footer. Verified 20 September 2026.
One integration is named and none is documented. Private PAIR integration appears in the published list of what the platform does, and the privacy policy refers in passing to Word add-ins, plugins and APIs among the services it covers, without a page, listing or document describing any of them. Nothing states what syncs with a docketing or IP management system, what a firm must configure, or how drafts move into the systems a prosecution practice runs on. Checked the homepage, the How It Works and pricing pages, the terms, the privacy policy and the free tools page links on 20 September 2026, and searched the Microsoft marketplace for a PowerPatent Word add-in without finding one. Verified 20 September 2026.
Deployment Model and Data Residency
Where the software runs and where the data sits. Multi tenant cloud, single tenant, private deployment, on premises, and whether region of residence is a published option or an enterprise conversation.
Both halves are answered and neither is offered as a choice. Residency is stated plainly and in three places, on the home page, the security page and the trust center: all data associated with the service is stored and processed within the United States. The model path is part of that answer rather than an exception, because models are reached only through Microsoft Azure and the service does not interact with any system operated by a third-party model provider. Separation between customers is asserted as an outcome: a customer's invention materials and drafts are never available to any other customer. What is not published is the architecture behind either statement. No cloud region, tenancy model, single-tenant or private option is described, no on-premise deployment exists, and a buyer outside the United States has no region to choose. For a product holding unfiled inventions the residency commitment is the material one, and it is unambiguous. Verified 20 September 2026.
Where processing happens is stated, and the customer gets a say in it. The terms provide that patent documents are processed on Google Cloud Platform or Amazon Web Services, that the choice between them rests with the user, and that all processing of patent documents on those platforms is done exclusively on computers in the United States, with each provider's own data processing terms named and linked. The privacy policy adds that some services may process data locally on the device while others run in PowerPatent-controlled or third-party cloud environments, and that international transfers may occur for other information. No tenancy model is described, and no region outside the United States is offered. Verified 20 September 2026.
Security Certifications and Trust Center
Independent attestation a buyer can pull without a sales call: SOC 2, ISO 27001, penetration test summaries, a trust center with current reports and named scope rather than a badge image.
A standard is named, a trust center exists, and the certificate is never qualified. The vendor states it is SOC 2 certified with independently audited security controls, says so on the home page, the security page and in a frequently asked question, and links a trust center at its own subdomain. That is above the floor reserved for a site claiming certification without naming a standard. What is missing is every particular that would let a buyer weigh it. No type is stated anywhere, neither Type I nor Type II, which is unusual because that distinction is the difference between a design opinion and one covering operating effectiveness over a period. No audit period, no scope or trust services criteria, and no auditor are published. No subprocessor list or penetration testing statement appears. The trust center itself is an application that did not render on reading, and one search recovered no document list from it, so whether a report is obtainable could not be established. Verified 20 September 2026.
The certification is real, stated and dated by a press release on the vendor's own site: PowerPatent, Inc. has completed SOC 2 Type 2 certification, with the AICPA SOC mark shown on the homepage. What the top band asks for beyond the name is missing: no auditor is identified, no scope statement or report period appears, there is no trust portal, and no route to obtain the report is published. A firm's security reviewer would be asking for the report and would have to ask the sales team. Verified 20 September 2026.
Model Supply Chain Disclosure
Which models sit underneath, whose they are, where they run, and whether the vendor commits to telling customers when that changes. A legal buyer inherits every dependency it cannot see.
The route to the models is named and described more carefully than most, and the models themselves are not. The vendor states the service is model-agnostic and reaches leading large language models only through Microsoft Azure over secure API calls, never interacting with any service operated by a third-party model provider directly. The trust center spells out the arrangement: Microsoft fully controls the third-party models provided through Azure and hosts them in its own environment, and customer inputs and outputs are not used to improve any Microsoft or third-party model, product or service. For a buyer whose material is an unfiled invention, knowing the data never leaves that path is the point of the disclosure. What is absent is identification: no model, family or version is named, nothing says which model performs which drafting step, and nothing commits to notifying customers when the underlying model changes beneath a drafting standard they have tuned. Verified 20 September 2026.
The provider is named in the agreement, which is rare in this market. The terms state that PowerPatent offers generation using OpenAI's GPT technology, that using the feature authorizes transmission of portions of the document to OpenAI via its API subject to OpenAI's data usage policies, and that no data is shared with OpenAI for training or improvement; Google Cloud Platform and Amazon Web Services are named as the processing platforms with their own terms linked, and the OpenAI and AWS marks appear on the homepage. What is not published is the model or version, or any commitment to notify customers when the model behind the drafting changes. Verified 20 September 2026.
Commercial Transparency
Whether a buyer can learn what this costs without entering a sales process: published rates, the unit being charged, what sits behind an enterprise tier, and what implementation adds.
Three tiers are named with what each contains, and no figure appears anywhere. The Pilot is genuinely free and genuinely self-serve, described as no-cost pilot access with no procurement and no sales call, reachable from a sign-up link, and giving access to standard agents. Professional is described as predictable per-user pricing, which names the unit while withholding the rate and directs the buyer to a call for a current quote. Enterprise covers unlimited users, custom drafting agents, prosecution-aware drafting and priority support and onboarding. Custom drafting agents that encode a firm's templates and standards are stated as available on the two paid tiers. Movement between tiers is addressed, with pilots scaling up as volume grows. What is missing is the number: no rate, band, minimum, term or implementation cost is published for either paid plan, and no currency appears on the estate. Verified 20 September 2026.
A buyer can see the entry price without talking to anyone. The pricing page publishes the unit and the rate: an initial purchase of 200 PowerPatent tokens for 199 dollars, additional tokens at one dollar each, a minimum balance of ten tokens to use the AI chat, and expiry six months from purchase, with the published guidance that most standard applications use fewer than 200 tokens and that complex workflows with searches or repeated checking need more. What is withheld is the rest of the range: enterprise licensing is a contact-us, and the terms describe per-user subscriptions that renew automatically and are non-refundable, a structure the pricing page does not price. Verified 20 September 2026.
Firm and Practice Coverage
Who the product is actually built for. AmLaw, midlaw, small firm and solo, in house departments, government and courts, and which practice areas are supported rather than merely claimed.
Three buyer segments carry their own pages and the named users span all three. Law firms are addressed on carrying more matters through the same team without thinning quality; in-house and corporate IP on scaling a filing program without scaling cost; solo and small firms on drafting like a big shop starting in minutes, which is the segment most often left out of enterprise legal AI. The named users match: a solo founder-managing partner at Wood IP LLC, a mid-sized IP firm in Xsensus, and an in-house team at RapidSOS. Practice coverage is patent application drafting for the USPTO, with the cohort study, the section 112 discussion and the prosecution-aware drafting claim all framed around US practice. What is absent is the boundary. Nothing states whether foreign filings, PCT practice, office action responses, prosecution beyond drafting or other rights are supported, so a buyer has to infer the edges of the product. Verified 20 September 2026.
The technical coverage is described in detail and the professional boundary is not. Twenty-three technology areas have their own pages, from semiconductors, software and communications to biopharmaceuticals, therapeutic antibodies, medical devices and agricultural biotech, and the drafting is said to handle chemical formulas and specifications in several languages. The buyers addressed are patent attorneys, agents and prosecution teams on one side and inventors and founders on the other, with a separate Office Action product for prosecution responses. Nothing states which jurisdictions the drafting is suited to beyond the United States, which technology areas the models handle less well, or where the tool stops being appropriate. Verified 20 September 2026.
The 12 legal signals, side by side
Recorded rather than graded. These are the questions a practitioner has to answer before a tool touches a client matter, and the answers are taken from public material only.
Client Data in Training
Can material a lawyer puts into this product be used to train a model?
The answer is absolute, stated in three places, and not in any agreement. The trust center sets it out as a list covering both sides of the transaction: a customer's inputs, meaning invention-related materials, and outputs, meaning the patent application documents, are not available to other customers, are not used to train any AI model, and are not used to improve any Microsoft or third-party model, product or service.
The security page repeats it and the home page carries it as a headline commitment. The only published legal document is a set of website terms of use, which contains no clause on data, training or confidentiality at all, so none of this is contractual and the vendor may revise those terms by updating the posting. One thing a buyer should pin down: the platform page states that drafting patterns are continuously tightened from the real prosecution signals that filed applications generate, and nothing published says whether that loop draws on published file wrappers, which are public records, or on anything else.
The commitment is in the agreement. The terms limit PowerPatent's license over user content to the sole purpose of providing the services, and state that data is not shared with OpenAI for model training or improvement. The privacy policy of 11 July 2026 goes wider, saying that customer-submitted confidential content is not used to train, retrain or fine-tune third-party foundation models or PowerPatent's own general-purpose models unless expressly agreed in writing for a particular customer.
Prompt and Output Retention
How long does the product keep what a lawyer typed, and can that be set to zero?
How long anything is kept is not addressed anywhere. The estate answers what the vendor will not do with invention materials and drafts, that they are never used for training and never available to other customers, and where they sit, within the United States, but says nothing about for how long. No retention period is published for uploaded disclosures, specifications, diagrams or source code excerpts, none for generated application drafts, and nothing describes deletion on request, on account closure or on termination.
The website terms carry no data clause to fill the gap and no subscription agreement is published. The question matters more here than on most records because the material held is unfiled invention content, where the customer's exposure is not confidentiality alone but novelty, and because the free self-serve pilot means a practitioner can upload a real disclosure before any agreement is negotiated. Checked the security page, the trust center, the terms, the privacy link and the pricing page on 20 September 2026.
The terms set a fixed position the customer does not control: once results are provided, patent documents are immediately discarded, with no storing, caching or logging of their content. The privacy policy pulls the other way, saying user content is kept as long as reasonably necessary with periods varying by service, setting and agreement; the policy itself provides that a signed agreement controls where the two conflict.
Ethical Walls and Matter Segregation
Does retrieval respect the firm’s ethical walls, or can the model read across them?
Separation between customers is asserted plainly and nothing describes how it is enforced. The security page lists isolation from other customers as a safeguard, stating that a customer's invention materials and drafts are never available to any other customer, and the trust center repeats it for both inputs and outputs. That is the right assurance and it is published rather than offered on a call. What is absent is the mechanism: no permission model is described, nothing says whether isolation is architectural or procedural, and no access controls, roles or audit logging appear.
Nothing at all addresses separation inside a customer, which is the sharper question for this buyer, because a patent firm routinely holds competing clients in the same technology area and a drafting workspace that surfaces one client's embodiments while drafting for another is a conflict rather than a bug. No matter-level wall, client screening or per-user access boundary is published.
The terms prohibit sharing an account or using it on behalf of another person, and restrict subscription content such as diagnostic results and examiner statistics to the user's own company, while allowing attorneys and agents to share output with their clients. That is licensing and confidentiality, not segregation: nothing describes separation between matters inside a firm or between customers.
Third Party Request and Subpoena Notice
If someone subpoenas the vendor for a firm’s data, does the firm hear about it first?
Nothing published says what happens when a demand for a customer's material arrives. The trust center states that the vendor does not directly or indirectly, without authorization, reveal, report, publish, disclose or transfer a customer's confidential information to any third party, which is a commitment about voluntary disclosure and leaves compelled disclosure untouched. No clause anywhere commits to telling the customer that a subpoena, court order or government demand has been received, to waiting before producing anything, to narrowing the response, or to assisting with a protective order.
The published website terms contain no confidentiality section at all, so there is no document in which such a commitment could sit, and no subscription agreement is published. No transparency report exists. For a platform holding unfiled invention disclosures, where disclosure destroys novelty, the absence is material. Checked the terms, the security page, the trust center and the privacy link on 20 September 2026.
The privacy policy of 11 July 2026 allows disclosure to governmental, regulatory, judicial or law enforcement authorities where required by law, subpoena, court order or legal process, with no commitment to notify the customer. The terms cut across it for one category, promising that the content of patent documents and processing results is not disclosed to any third party other than the named processors, but neither document says what happens when the demand is legal process.
Primary Law Corpus Provenance
Where does the law in this product come from, and does the vendor have the right to use it?
Two external sources are named and the rights basis is stated for only one. The drafting itself works from what the customer supplies, invention disclosures, technical specifications, design documents, diagrams, source code excerpts and related materials, so the corpus behind a draft is largely the customer's own. Beyond that the estate names Project PQAI as the source of prior-art search outputs and metadata in preliminary search reports, and it does state the basis: use of that material is subject to PQAI's posted terms and API terms, which control over the vendor's own terms for that content.
The second source is named without a basis, the real prosecution signals that filed applications generate, which the platform page says continuously tighten its drafting patterns; nothing identifies which records those are, how they were obtained or how current they are. No licensing is stated for anything the models themselves were built on, since those are reached through Microsoft Azure.
Checked the homepage, the How It Works and pricing pages, the terms and the privacy policy on 20 September 2026. The drafting works from the user's own claims, summary and drawing annotations; the marketing says the models analyze vast amounts of patent data, and the terms note that some products rely on third-party data such as patent office data, without identifying any corpus, license or update cadence.
Good Law Verification
Does the product tell you when the authority it just cited has been overruled?
Nothing here cites authority, so there is nothing for a status or treatment signal to attach to. The product generates specification text from invention materials; it does not quote a statute, rule or decision, and it produces no citations a reader would need to verify. The one place external documents enter is the preliminary search report, which returns prior art from a third-party source, and even there nothing published addresses whether a surfaced patent or application is still in force, has been abandoned, has expired or has been narrowed, which is the patent equivalent of the question.
The vendor's own cohort study discusses prosecution outcomes on its drafted applications rather than the status of anything cited to a user. Checked the home page, the solutions pages, the platform section, the terms and the security page on 20 September 2026.
For a patent product this asks whether the tool says a patent is still in force. Checked the homepage, the How It Works and pricing pages, the terms and the privacy policy on 20 September 2026. The product drafts applications and checks them for Section 112 problems rather than reporting legal status, and nothing addresses whether any patent surfaced through the examiner statistics or search features remains in force.
Refusal and Uncertainty Behavior
What does the product do when the answer is not in the corpus?
Nothing published describes what the agents do when the inputs will not support a draft. The workflow is described as confident throughout: materials go in, an alignment step captures claim focus, terminology and embodiments, and a complete fully-enabled application comes out in minutes. No confidence signal accompanies a generated section, no threshold is described below which the system declines to draft, and nothing says what a practitioner sees when a disclosure is too thin to support the claim scope requested, which is the ordinary situation the product must handle and the one where silent confident generation is most costly.
The nearest published material runs the other way, promising fallback support embedded in the draft and specifications built to survive examination. The terms disclaim accuracy without describing any behavior. Checked the home page, the platform section, the solutions pages, the security page and the terms on 20 September 2026.
Checked the homepage, the How It Works page, the terms and the privacy policy on 20 September 2026. Nothing describes the drafting declining to generate, flagging that it lacks support in the drawings or claims, or attaching a confidence level to generated text. The published safeguards work the other way, by checking a finished draft for Section 112 problems and gaps in claim support.
Fabricated Citation Record
Does a public court record exist addressing fabricated or hallucinated legal citations in output from this product?
No record was located of this product's output being found fabricated or inaccurate in a proceeding, a regulatory action or a published account. Searches on 20 September 2026 across the vendor's estate, press and directory profiles returned nothing of the kind. The exposure here is not an invented citation, since the product asserts no authority: it would be a generated specification that does not support the claims it accompanies, surfacing as a section 112 rejection during prosecution or as an enforceability problem later.
The vendor publishes cohort figures addressing exactly that risk, reporting no section 112(a) enablement rejections in the applications studied, which is the vendor's own account rather than an independent finding.
Searched the AI Hallucination Cases database maintained by Damien Charlotin on 20 September 2026 on the name PowerPatent. No court order, opinion or disciplinary record naming the product was located. This is a statement about the public record rather than a finding about the product, which drafts applications for a patent office rather than briefs for a court.
Bar Guidance Alignment
Has the vendor engaged in public with the ethics opinions its buyers are bound by?
Professional duty is engaged in general terms and no guidance is named as binding the product. The terms state that communication through the site does not create or constitute an attorney-client relationship, is not intended to convey or constitute legal advice, and is not a substitute for obtaining legal advice from a qualified attorney, and the search report clause tells users to consult qualified counsel before acting.
The service is defined as drafting under the supervision of a patent practitioner, which engages the supervision duty without citing it. The estate does name real authority, in an article advising firms how to build AI policies that cites California Rule 1.1, the USPTO Rules of Professional Conduct at 37 CFR 11.101, California Bar practical guidance and the Virginia State Bar model AI policy. That is the vendor telling buyers how to govern themselves rather than a statement of what the product does, and the article itself says it is informational only, so it does not lift this row.
The terms engage a practitioner duty in general terms: a user must represent that they can practice before a national patent office in their jurisdiction, and a separate clause states that nothing communicated creates an attorney-client relationship, is a solicitation, or conveys legal advice. No ethics opinion, rule of professional conduct or patent office guidance on the use of AI in drafting is named anywhere on the estate.
Billing and Fee Posture
Does the vendor address what happens to the bill when the work takes an hour instead of six?
This product sits squarely inside the lawyer-client fee relationship and addresses only one side of it. The buyer is frequently a firm that bills clients for patent drafting, and the commercial case is explicitly about the economics of that work: more matters through the same team without thinning quality, scaling a filing program without scaling cost, and a published average of roughly four hours from inventor materials to filing-ready against what the vendor describes as days.
The estate goes further than most in discussing what that does to patent strategy and spend, including a piece arguing that the economics which shaped filing decisions for forty years no longer hold. What is absent is the client-facing half. Nothing addresses whether a practitioner should tell a client that an application was machine-drafted, how time saved should be reflected in a bill, or how the tool's cost is treated on a flat fee.
The published case is that the same work takes less time: first drafts in minutes, five or more hours saved per application, and a feature headed increase billable efficiency, work less and earn more. Where the buyer is a firm drafting for clients, those hours are billed hours, and nothing published addresses how a machine-drafted specification is described on the bill or whether the token cost is passed to the client.
Outside Counsel Guideline Readiness
Can a firm get this vendor through a client’s AI clause without a bespoke negotiation?
The model provider is published in detail, which is the disclosure a client guideline most often asks for, and the rest of the pack is missing. The trust center and security page name Microsoft Azure as the sole route to the models, state that Microsoft controls and hosts those models in its own environment, and state that the service never interacts with any system operated by a third-party model provider directly. Alongside that sit US-only storage and processing, a no-training commitment covering Microsoft and third-party models, and SOC 2 certification.
A reviewer could answer several questionnaire lines from public pages alone. What is absent is the rest: no subprocessor list beyond the Azure route, no data processing agreement, no audit report type, period, scope or auditor, no penetration testing statement, no breach notification commitment, and no client-facing disclosure material. The trust center itself did not render and one search recovered no document list.
The three providers that touch a patent document are named in the agreement itself: Google Cloud Platform or Amazon Web Services for processing, with their data processing terms linked, and OpenAI for generation, with a statement that no data is shared for training. The privacy policy describes further categories of subprocessor, including AI infrastructure, without naming them, and no client-facing disclosure pack was located.
Court Disclosure Support
If a judge’s standing order requires an AI disclosure, can the product produce one?
Nothing published addresses disclosing the machine's involvement to anyone outside the firm. No court sits in this product's path, and the forum that does, the USPTO, is not addressed either: nothing raises whether a practitioner filing a machine-drafted application should disclose that, nothing engages the duties owed to the office, and no exportable record of how a draft was produced is described. The platform keeps the practitioner's alignment inputs, the claim focus, terminology and embodiments set before generation, but nothing says those are retrievable afterwards or distinguishable from what the agents added.
The vendor does publish an article advising firms to build AI policies, citing USPTO and state bar guidance, which is guidance for the buyer rather than a capability of the product. Checked the home page, the solutions pages, the terms, the security page and the trust center on 20 September 2026.
Checked the homepage, the How It Works page, the terms and the privacy policy on 20 September 2026. The product drafts and checks applications for filing at a patent office rather than producing court filings, and nothing records which parts of a specification were machine generated, which model produced them or who reviewed them, so no disclosure record could be exported if one were required.
The questions both sides leave open
Derived from the records above rather than written, so it cannot favor either vendor. Take these into both conversations and ask each side the same question.
- Good Law Verification
- Refusal and Uncertainty Behavior
- Court Disclosure Support
Which one fits
Choose Paximal if
- You want evidence of how the drafts fare at examination. Paximal publishes a cohort study of its drafted applications at the USPTO reporting about 70 percent allowed after one office action, no section 112(a) enablement rejections, and around four hours from inventor materials to a filing ready draft.
- You want the model path kept inside one cloud. Paximal reaches its models only through Microsoft Azure, never directly through a model provider, states that inputs and drafts never train any model or reach other customers, and stores and processes everything in the United States.
- You want to try it on a real matter without procurement. Paximal offers a free self serve pilot with standard agents and no sales call, and custom drafting agents that encode a firm's templates on its paid tiers.
Choose PowerPatent if
- You want the discard and training terms in the contract. PowerPatent's terms state that patent documents are discarded immediately after results are provided, with no storing, caching or logging, and that no data is shared with OpenAI for training.
- You want to price one application before you commit. PowerPatent publishes 200 tokens for $199, which it says covers most standard applications, with further tokens at $1 each and expiry six months from purchase.
- You want diagnostics before filing. PowerPatent flags section 112 problems in enablement, written description and indefiniteness and gaps in claim support, writes descriptions figure by figure from part lists, and offers a separate Office Action product for responses.
In summary
Paximal
Paximal, from Paximal, Inc., a Delaware company operating from California and founded by the founder of Specifio, is an agentic patent drafting platform. A practitioner uploads invention materials and sets claim focus, terminology and embodiments, and orchestrated agents produce a complete application with terminology harmonized and claims mapped to the specification. The AI Legal Index grades it in the top two bands on nine of fifteen capability axes, with an A on AI centrality. It publishes a cohort study of its drafted applications at the USPTO, reaches its models only through Microsoft Azure, and states SOC 2. Named users include Wood IP and Xsensus. As of 20 September 2026 the index located no customer agreement, retention period or price figure.
PowerPatent
PowerPatent, from PowerPatent, Inc. of Santa Clara, California, is generative AI drafting software for patent applications, sold to patent attorneys, agents and inventors. From an invention summary, claims and annotated drawings it writes a background, preliminary claims and a figure by figure description, runs section 112 diagnostics, and offers collaboration, Private PAIR integration and an Office Action product. The AI Legal Index grades it in the top two bands on ten of fifteen capability axes. Its terms name OpenAI, Google Cloud and AWS, discard patent documents after processing and bar training on them, and it publishes a price of $199 for 200 tokens. It states SOC 2 Type 2. As of 20 September 2026 the index located no named customer or accuracy measure.
Questions buyers ask
Paximal vs PowerPatent: which is better for drafting patent applications?
The grid barely separates them: PowerPatent sits in the top two bands on ten of fifteen AI Legal Index capability axes and Paximal on nine of fifteen, identical on eight. PowerPatent puts its discard, training and processing terms in its agreement and publishes a price. Paximal publishes prosecution outcomes for its drafts and names users. Firms that want contract terms first have more to read from PowerPatent.
How do Paximal's drafted applications perform at the USPTO?
Paximal's own cohort study of its drafted applications reports roughly 10 percent allowed on first action and about 70 percent after one office action, no section 112(a) enablement rejections, about 20 percent with minor section 112(b) clarifications, and around four hours of preparation. The cohort is not named, no sample size is given, and the figures are approximations. Graded by AI Legal Index against 15 capability axes and 12 legal signals, including privilege handling and citation accuracy, from each vendor's own published materials, verified September 27, 2026. No vendor pays for placement.
Which AI models do Paximal and PowerPatent use?
PowerPatent's terms name OpenAI's GPT technology, reached through OpenAI's API, with processing on Google Cloud or AWS in the United States at the user's choice. Paximal states it is model agnostic and reaches leading models only through Microsoft Azure, never a model provider directly. Neither names a model version or commits to notice when the model changes. Graded by AI Legal Index against 15 capability axes and 12 legal signals, including privilege handling and citation accuracy, from each vendor's own published materials, verified September 27, 2026. No vendor pays for placement.
How much do Paximal and PowerPatent cost?
PowerPatent publishes 200 tokens for $199, which it says covers most standard applications, with extra tokens at $1 and expiry after six months; enterprise licensing is quoted. Paximal offers a free self serve pilot and per user Professional and Enterprise tiers whose prices are not published. Graded by AI Legal Index against 15 capability axes and 12 legal signals, including privilege handling and citation accuracy, from each vendor's own published materials, verified September 27, 2026. No vendor pays for placement.
What do Paximal and PowerPatent both leave unpublished?
An accuracy measure for generated text and a record of what the machine wrote. Neither publishes a hallucination or error rate for its drafts, neither describes what its drafting does when the materials cannot support a claim, and neither records which parts of a specification a model produced. Neither addresses whether a client should be told an application was machine drafted. Graded by AI Legal Index against 15 capability axes and 12 legal signals, including privilege handling and citation accuracy, from each vendor's own published materials, verified September 27, 2026. No vendor pays for placement.
Three readings to weigh. Paximal's terms contain no confidentiality clause and require users not to assert patents derived from its output against it; its training and confidentiality commitments sit on its trust and security pages. PowerPatent's privacy policy says user content is kept as long as reasonably necessary, against its terms' immediate discard, and its terms place all risk of a bad draft on the user. Paximal's cohort figures are approximate and unnamed. Paximal and PowerPatent were both verified on 20 September 2026. Neither vendor reviewed this page.
Neither vendor paid for inclusion, placement or a grade, and neither reviewed this page before it published. Everything above comes from public material on the dates shown. How the index grades.