1Patsnap Eureka: disclosure to filed application in one agent workflow

Typically used by: patent agencies and in-house IP teams that want drafting, search and prosecution running off one evidence base rather than three tools.

Claims quality

Eureka’s drafting agents are Invention Disclosure, Patent Drafting and Office Action Response.² The claims work is not a single generate button. The agent first analyses the disclosure, then runs an embedded prior art search, then plans a drafting outline and confirms the essential technical features with you before any claim text is written, so the claim set is shaped against what the search actually found rather than against the inventor’s summary alone. It proposes product, method and use claims in layers, and you can steer whether independent claims are drafted broadly for enforcement or narrowly for a faster grant.

Three control layers sit on top: Templates for structure, Instructions for rules, and Styles for preferences, which the product describes as adapting to “patent-office rules, industry conventions, enterprise standards, law-firm requirements, and personal writing habits.”² Upload a reference patent and the AI extracts its structure and voice into a reusable template. Multimodal understanding covers mechanical drawings, circuit diagrams, chemical structures, LaTeX formulas and experimental data tables, which matters when the claim language has to track a reaction scheme or a parameter range.

Jurisdictions

Eureka states support for CNIPA, USPTO and EPO standards, with jurisdiction-specific rule libraries and templates behind them.² The Office Action Response agent works across the same three offices, auto-detecting the patent number, parsing the action, extracting every rejection into an analysis table and generating ranked response strategies before drafting. For teams filing into China and prosecuting in the US and Europe, that is one workflow rather than three.

Review

Compliance review is a required final step, not an optional check, and it runs on two tracks: a semantic review and a deterministic review. This is the part most generation tools leave out. The search side is measured too: Patsnap publishes a benchmark methodology, PatentBench, in which the novelty search agent records an 85% X Hit Rate and a 37% X Recall Rate over 340 cross-jurisdiction patent family samples, with ground truth taken from “X references cited by examiners across different patent offices,” deduplicated by family, tested July 2026.¹ The sample size, the ground-truth definition and the metric definitions are all published alongside the result, so the number can be checked rather than taken on trust.

AI, security and access

The platform runs across 200M+ patents in 174 jurisdictions and publishes SOC 2, ISO 27001, GDPR and CCPA compliance, with no AI training on user data.³ The Patsnap Open Platform publishes 31 MCP servers described as connecting “Patsnap’s domain-specific AI agents to LLM platforms such as Claude or custom models using the Model Context Protocol,” with a free Starter tier of 10,000 credits valid for 90 days.^{4, 5} New users start with 10,000 free credits, and paid tiers are published on the product page.

2DeepIP: an AI copilot that lives inside Microsoft Word

DeepIP was founded in 2024, has offices in New York and Paris and has raised $40M.^{6, 7} Its AI Reviewer will “automatically check for antecedent basis errors and dependency gaps” and flags §101 language that mirrors judicial exceptions.⁸ It names its offices explicitly: “USPTO, EPO, and every other major office, CIPO, CNIPA, DPMA, INPI, IPO, JPO, KIPO, UKIPO.”⁹ On data it states “we do not retain any client data nor use it for model training,” with ISO 42001, ISO 27001 and SOC 2 Type II.¹⁰ Pricing is not published.

3ClaimMaster: claims checking first, drafting bolted on

ClaimMaster was “originally developed by a practicing patent attorney and a former USPTO examiner,” and claims checking is its original purpose.¹¹ It will “find missing antecedents in claims,” “find claim terms without support in patent specification,” print claim trees and generate claim charts.¹² The drafting layer is bring-your-own-model, connecting to private Azure and OpenAI models or “locally running LLMs (i.e., no cloud processing).”¹³ It is a Word add-in that “executes locally on your computer (on-premise),” and it is USPTO-centric.¹⁴ Three tiers are named with a 30-day trial; no dollar figures appear.¹⁵

4Patent Bots: proofreading, drafting and examiner analytics

Founded in 2018 by a patent attorney, Patent Bots is a Word add-in running on Windows and Mac.¹⁶ It checks claim numbering and dependency errors, identifies antecedent basis errors, runs separate word and phrase support checks, and verifies that claim markup matches the previous version.¹⁷ It is USPTO-focused throughout, down to an Art Unit Predictor trained on USPTO data.¹⁸ Customer data is “not used for any AI model training” and it holds SOC 2 Type II.¹⁹ Individual practitioners at law firms pay $702 per person per year.²⁰

5XLSCOUT Drafting LLM: drafting inside a wider AI patent suite

XLSCOUT is a Toronto-headquartered platform describing itself as “an SOC 2 Type II certified, AI super intelligence platform for innovation and IP.”²¹ Its Drafting LLM produces claims, abstract, background, detailed description and drawings, alongside Novelty Checker and Invalidator modules.²² It states it “does not use your data to train its LLMs.”²³ Two things to check: no patent office is named as a supported drafting standard, and its own FAQ says Drafting LLM “acts as an aiding tool” to be used “with expert judgment.”²⁴ Pricing is quote-only.

6PowerPatent: first drafts with §112 diagnostics

PowerPatent, based in Santa Clara, is software “Created By Patent Lawyers, For Patent Lawyers. Enhanced by Generative AI.”²⁵ It drafts a background, “a series of claims for you to apply your patent skills” to, and a detailed description per figure “while flagging Section 112 issues and claim support.”²⁶ Those are the only claims-quality features it names, and it is USPTO-oriented.²⁷ Read the confidentiality statement closely: the local-processing guarantee is scoped to the Word add-in and the policy is silent on training.²⁸ Pricing is public and token-based: 200 tokens for $199.²⁹

7PatentPal: generation only, from claims to specification

PatentPal is a browser tool with a narrow job: “Generative AI for Intellectual Property… Automate mechanical writing in your patent applications.”³⁰ Drop in your claims and it generates flowcharts, block diagrams, figure descriptions and an “abstract and summary to support all claims.” Claims are the input, not the output: no claim drafting, no antecedent basis checking, no review function, and no patent office named. It processes documents “in memory” but says nothing about training.³¹ Pricing is not published, and every footer still reads “Copyright © 2022,” with PatentPal Invent marked ended.³²

Questions worth asking any vendor

Most of the differences that matter later are answerable in a single email before a demo.

• Which offices are supported, and what does support mean? A rule library and jurisdiction-specific templates are a different thing from an LLM that has read some EPO documents.
• Does it check claims, or only write them? Ask specifically about antecedent basis, dependency chains, claim support in the description and reference numeral consistency.
• Is the prior art search part of the drafting flow, or a separate purchase? Claims drafted without a search in front of them get rewritten after the first office action.
• Is customer data used for model training, and is that written down? Several vendors state a retention policy but say nothing about training. Absence of a statement is not a guarantee.
• Does the confidentiality promise cover the whole product? At least one vendor here scopes its local-processing guarantee to the Word add-in only.
• What does the contract say about accuracy? Every vendor in this list disclaims accuracy in its terms. Read that language next to the marketing percentages.

What no drafting tool settles

• None of them is practising law. Every vendor in this list says so in its own terms. DeepIP’s are typical: outputs “are not intended to be used as a substitute for legal advice,” and the company “is not a law firm, professional lawyer or qualified attorney.”³³
• A clean proofreading report is not a valid claim set. Antecedent basis and dependency checks catch mechanical defects. They say nothing about whether the claim is novel, enabled, or the right scope for the business.
• Generation speed is not the bottleneck it looks like. The expensive part of drafting is deciding what to claim, and that decision depends on what the prior art search found. Tools that separate the two move the cost rather than removing it.
• Published accuracy figures are self-reported. Where a vendor publishes a percentage, look for the sample size, the ground-truth definition and the metric behind it, and read it next to that vendor’s own accuracy disclaimer.