Patent application drafting & office-action response

Writing a patent application and fighting it through examination is expensive, expert-only work: an attorney turns a few pages of engineering notes into 20–60 pages of claims and description with precise, consistent terminology, then spends hours mapping examiner objections against cited prior art. Patent-specific LLM copilots now draft first versions of claims, specifications and figure text from the disclosure, run support and consistency checks that patent offices care about (Art. 84 and Art. 123(2) EPC, antecedent basis), and break office actions down into a feature-by-feature comparison with suggested amendments and arguments. The attorney stays the author and signatory; the AI removes the drudgery and catches errors before filing.

Problem & value

• Who: patent law firms (Patentanwaltskanzleien), corporate IP departments in manufacturing, automotive, pharma, chemicals and software, university technology transfer offices, and inventors at start-ups and SMBs who pay outside counsel by the hour.
• Today: the attorney interviews the inventor, writes claims by hand, expands them into a description, briefs a draftsperson on figures and proofreads reference numerals and antecedent basis manually. Office actions are read, cited documents are compared feature by feature, and a response is written in the style of the particular office (EPO problem-solution approach, DPMA practice, USPTO §102/§103 arguments).
• Why AI helps: most of this is long-form, structured, highly formulaic text that has to stay consistent with a small set of core definitions. That suits LLMs combined with deterministic checkers. Value lies in faster first drafts, broader fallback positions (more dependent claims and embodiments), fewer formal errors and faster responses.
• Reported benefits (vendor-reported, not independently verified): Patlytics cites an Am Law 100 practice group that finished a 120-patent project in 20 instead of 100 hours, and a biotech estimating USD 5,000–7,500 saved per drafted application [3]. DeepIP reports more than 76k office actions handled and 343k drafts created on its platform, an average of 25+ hours saved per user per month, and a customer testimonial on its site (attributed to Philips' head of IP) of about 20 % efficiency gains seen in a trial of several drafting tools [2]. Solve Intelligence reports 700+ in-house IP teams and law firms as customers, including Siemens and DLA Piper [1]. Treat these figures as marketing claims and validate them with your own pilot.

Variants

• New application drafting: from disclosure to claims, then spec and figures, or from attorney-written claims to spec ("claims-first"; e.g. PatentPal's claims-to-spec and figure generation [4]).
• Office-action / EPO communication response: rejection mapping, examiner-error spotting, amendment proposals with a basis in the application as filed, and response-letter shells (USPTO, EPO, DPMA, plus national offices). This covers the "EPO/DPMA practice adaptation" variant from the research sweep: jurisdiction-specific templates for the problem-solution approach, two-part claim form (Rule 43 EPC) and German-language DPMA filings.
• Quality control / proofreading only: antecedent basis, unsupported claim terms, reference numeral consistency, claim-dependency errors, Art. 84/123(2) EPC support checks (Patent Bots, ClaimMaster [5][6]).
• Invention-disclosure intake: converting slides, transcripts and lab notes into a structured disclosure form for the IP committee (Patlytics, Solve [1][3]); upstream of drafting.
• Domain variants: chemistry and biotech (Markush structures, sequence listings, examples tables), mechanical (drawings from CAD), software and computer-implemented inventions (EPO technical-character framing, USPTO §101 eligibility arguments).
• Foreign filing: translating the priority application or validating an EP patent (with LANG-009); terminology lists carry over.
• Private inventors / micro-entities: consumer-grade tools exist (PatentPal free trial), but self-drafting without an attorney carries a high risk of worthless claims. Recommend AI only as preparation for counsel.

How it works

1. disclosure intake (docs, slides, CAD, transcripts)
2. parse (OCR/VLM)
3. structured invention record
4. prior-art context (KNOW-007 search: closest art, known features)
5. claim drafting (reasoning LLM: independent claim, fallbacks, claim tree) ← ATTORNEY REVIEW #1
6. spec expansion (sectioned generation per claim feature; templates per office/technology)
7. figures (flowcharts/block diagrams generated from claims; CAD
8. line drawing; numerals)
9. deterministic checks (antecedent basis, numeral consistency, claim support, term glossary)
10. attorney edit in Word add-in ← ATTORNEY REVIEW #2
11. filing (attorney signs) office action / communication PDF
12. parse
13. rejection items
14. fetch cited art (Espacenet/USPTO)
15. feature mapping table (claim element ↔ citation passage, with page/line)
16. strategy options (argue / amend / both)
17. amendment with basis in application as filed
18. Art. 123(2) / new-matter check
19. response letter draft ← ATTORNEY DECIDES

• Core models: a frontier reasoning LLM for claim construction and argument. Long context (100k+ tokens) is needed because spec plus cited art routinely exceeds 50k tokens.
• Retrieval: embeddings + hybrid search over the firm's own past filings and responses (precedent bank), the application as filed (for amendment basis) and the cited documents.
• Deterministic checkers (rule-based NLP) matter as much as the LLM: antecedent basis, numeral lists, dependency loops and words that are risky in claims are checked by code, not generated.
• Agentic workflow: planner, section writers and an "examiner" critic agent. This is the architecture of research systems such as AutoPatent [8] and of most commercial tools.
• Human-in-the-loop: claim scope and every argument to the office are legal judgements. The practitioner reviews, edits and signs; USPTO rules explicitly place that responsibility on the signer [10].

Data preparation

Input data is small but highly confidential: invention disclosures, lab notebooks, slides, CAD exports and meeting notes, plus office actions and cited prior art. In an SMB or law firm these sit in mail, file shares, a docketing system or the DMS, often as scans or PDFs. Typical gaps: scanned office actions, inconsistent terminology between disclosure and claims, scattered earlier filings of the same family, and unclear access rights between clients or projects.

Steps: - Convert scans and PDFs to searchable text with OCR (PREP-006, DOC-014). - Collect past filings and successful responses as style examples (PREP-016). - Build a terminology list per technical field (PREP-015). - Separate matters and clients with strict permissions (PREP-023). - Define which data may go into which AI tool (PREP-024).

Tools: Paperless-ngx or the DMS, OCR, a vector database per matter. Preparation is a small share, around 10 to 20 percent, of the effort; the main work is clean access control and a curated set of templates.

Implementation paths

Which path for whom

• Enterprise: B, because in-house IP departments handle unfiled inventions and want integration with invention-disclosure systems and docketing, with API cost under control.
• SMB: A, because patent-specific copilots with EU hosting and confidentiality terms are mature and small firms or IP teams have no capacity to build.
• Developer: C, because self-hosted open models allow building and testing drafting and analysis tools without sending unfiled inventions to a third party.

Path A — Buy (recommended for most law firms and SMB IP teams)

• Patent-specific copilots (mostly Word add-ins plus web app): Solve Intelligence (UK-founded; drafting, office-action analysis and responses for any office, EU or US data residency, ISO 27001, SOC 2 Type II) [1]; DeepIP (French-founded; drafting, prosecution, Word/Outlook integration, lists ISO 42001/27001, SOC 2, GDPR and German §203 StGB compliance, but hosting is stated as US-based Azure, so check EU residency) [2]; Patlytics (US; drafting, office actions, disclosures, infringement) [3]; PatentPal (claims → spec and figures) [4].
• Proofreading / prosecution utilities: Patent Bots (proofreading, Sprocket drafting assistant in Word, OA shells, examiner statistics) [5]; ClaimMaster (Word add-in; antecedent basis, spec support, OA analysis; runs locally, GenAI features via OpenAI GPT) [6].
• Search platforms adding drafting: IPRally (Finnish, EU; drafting with closest prior art in view, office actions "coming soon", dedicated EU Google Cloud environment, MCP/API access) [7].
• Selection criteria: EU data residency and no training on client data, a confidentiality agreement covering professional secrecy (§203 StGB, §39a PAO for German patent attorneys), EPO/DPMA templates (not only USPTO style), German-language quality, Word integration, audit trail of AI-generated passages.

Path B — Build on cloud APIs

• Models: Claude (via AWS Bedrock Frankfurt or Google Vertex EU), Gemini (Vertex AI EU regions), GPT via Azure OpenAI with EU Data Zone, Mistral Large / Magistral (EU provider). Use long-context reasoning models for claims and arguments, and smaller models for section drafting.
• Integration: Office.js Word add-in (where attorneys work), DMS (iManage, NetDocuments), docketing (Anaqua, CPA Global/Clarivate IPMS, PATRIX), patent data from EPO OPS (Espacenet) and USPTO APIs for cited art.
• Orchestration: LangGraph / agent SDKs with MCP tools (fetch citation, check antecedent basis, render flowchart via Mermaid/Graphviz). Use structured outputs for claim trees and mapping tables.
• When: large IP departments that want their own playbooks and precedent retrieval, or firms that need a specific office's practice which tools cover poorly.

Path C — Self-host (open source)

• Models: strong open-weight reasoning LLMs, e.g. Qwen3 (235B-A22B MoE or 32B dense), gpt-oss-120b, DeepSeek-V3.x/R1-class or Mistral Medium/Small open weights, served with vLLM or SGLang. The German-language patent register needs a model with good German output, so evaluate on your own filings.
• Stack: Docling/MinerU for parsing office actions and cited art; bge-m3 / multilingual-e5 embeddings
• Qdrant/OpenSearch hybrid search over the precedent bank; LangGraph agents; open-source antecedent-basis checkers written in spaCy rules.
• Hardware: HW3 (one 80–141 GB GPU) for a ~30–120B model at a few concurrent users; HW4 for the largest MoE models. Ops effort is moderate: one ML engineer part-time plus IT.
• When: firms that must not let unpublished inventions leave their premises (defence suppliers, export-controlled technology, strict client mandates). The USPTO guidance explicitly warns about export-control and confidentiality issues with third-party AI [10].

Path D — Custom training

Worth it only for large firms or vendors. Public corpora such as HUPD (4.5M+ US applications) [9] and EPO full-text data allow continued pre-training or fine-tuning on patent register. The firm's own filed applications paired with the original disclosures, and office actions paired with granted responses, are the valuable proprietary signal. AutoPatent reports that a fine-tuned 7B model in a multi-agent framework outperformed much larger general models on its benchmark [8], so small specialised models are a credible cost lever. Expect thousands of document pairs and expert evaluation.

State of the art (2026)

• Category maturity: dedicated patent-drafting copilots went from niche to mainstream in large firms in 2024–2026. Vendors report hundreds of firm customers each and large funding rounds (DeepIP and Patlytics each report USD 40M raised) [2][3]. Most mid-size European firms are still piloting.
• Techniques: multi-agent drafting (planner → writers → examiner-critic) [8]; claim-first generation with structured claim trees; CAD/file-to-figure conversion and automatic reference labelling [1]; office-action "error spotting" that checks whether the examiner's mapping really discloses each feature; precedent retrieval of the firm's successful arguments.
• Benchmarks: research benchmarks are young. D2P (Draft2Patent, ~17k-token patents) [8] and PatentEval's error typology for claim and abstract generation [11] show that general LLMs still produce characteristic errors: unsupported features, inconsistent terms, overly narrow or vague claims. No standard benchmark yet covers office-action responses or EPO-specific practice.
• Limits and failure modes: hallucinated citation passages or case law; added subject matter during amendments (fatal under Art. 123(2) EPC [13]); claims that read on the prior art; boilerplate that narrows scope; weak German legal-technical register; US-centric defaults (means-plus-function habits, §101 framing) that are inappropriate before the EPO. Strategy decisions (claim breadth, divisional, appeal) remain human.

Shared resources & combinations

• Upstream: KNOW-007 (patent search & prior-art analysis) uses the same patent corpus, embeddings and OPS/USPTO connectors. Drafting with the closest art in view and responding to cited art both rely on it. KNOW-008 (scientific literature) covers non-patent literature, and KNOW-022 covers trademark/design clearance in the same IP department.
• Same document stack: DOC-014 (PDF parsing for RAG) for office actions and citations; DOC-013 (technical drawing understanding) for CAD → patent figures; KNOW-002 (RAG over firm precedents).
• Same "expert long-form drafting with legal accountability" pattern: DOC-010 (contract drafting & redlining), LANG-035 (litigation pleadings/Schriftsätze), LANG-022 (technical documentation). One EU-hosted reasoning-LLM endpoint, Word add-in framework and precedent vector store can serve all four.
• Translation: LANG-009 for priority translations and EP validations, with shared terminology bases.
• Hardware: a self-hosted HW3 node running a ~30–120B reasoning model for legal drafting is shareable across the legal and R&D use cases above.

Effort, cost & KPIs

• Path A: pilot in 2–6 weeks; per-seat SaaS licences (prices on request with most vendors). The main effort goes into templates, house style and training attorneys.
• Path B: 3–6 months for a Word add-in with precedent retrieval and checkers. Token cost per application is small compared with attorney hourly rates; the cost drivers are engineering and evaluation.
• Path C: add GPU hardware (HW3) and ops; 4–8 months to production quality.
• KPIs: attorney hours per application and per OA response; formal objections and clarity (Art. 84) objections per examination; number of fallback positions per application; first-action allowance / grant rate trends (long lag); proofreading defects caught pre-filing; attorney adoption rate.

People & change

Involve management and the head of IP or the senior attorney first, since professional responsibility stays with the attorney. Add IT and IT security early because unfiled inventions are trade secrets and must not leak or be used for model training; legal-compliance checks confidentiality duties to clients (including professional secrecy for law firms). The DPO is needed when inventor data or personal data in correspondence is processed; sign an AVV (Art. 28 GDPR) with every provider.

Works council: depends. Drafting support itself rarely monitors staff, but usage logs and productivity dashboards can fall under §87(1) no. 6 BetrVG; inform the council in good time under §90(1) no. 3. In a small firm without a council this does not apply.

AI literacy: support attorneys and assistants with training on hallucinated citations, claim support checks and confidentiality (AI Act Art. 4; GUIDE-006). Pilot with one technical field and two attorneys on already filed cases first. Typical concerns are quality, liability and secrecy; address them with mandatory attorney review of every claim, source citations and an approved-tools rule (GUIDE-007). See also GUIDE-005.

Risks, regulation & compliance

• EU AI Act: typically minimal risk. Drafting support for private applicants is not listed in Annex III (point 8 covers AI used by judicial authorities, not by parties or counsel). Art. 4 AI literacy applies. The output is professional text submitted under the practitioner's signature, so Art. 50 labelling is generally not triggered. Patent offices' own use of AI for examination is a separate topic.
• Confidentiality and novelty: unpublished inventions are trade secrets. Disclosure to a vendor under confidentiality is not a publication, but use only contractually bound, no-training, access-controlled services. German patent attorneys are bound by professional secrecy (§39a PAO, §203 StGB). The USPTO 2024 guidance stresses confidentiality, national-security and export-control risks of third-party AI systems [10].
• Responsibility: the USPTO guidance confirms that duty of candor (37 CFR 1.56) and signature certification (37 CFR 11.18) apply in full. Practitioners must verify every AI-generated statement and citation [10]. The same applies to EPO and DPMA representatives under their professional codes.
• Inventorship: AI cannot be named as inventor (EPO DABUS decisions J 8/20, J 9/20; German BGH X ZB 5/22). Using AI as a drafting tool does not affect this, but keep a record that humans conceived the invention.
• Formal law: Art. 84 EPC requires claims to be "clear and concise and be supported by the description" [12]; Art. 123(2) EPC forbids amendments extending beyond the application as filed [13]. AI-proposed amendments need a verified basis citation.
• GDPR: limited personal data (inventor names and addresses, examiner names); a DPA with the vendor is still required. Prefer EU data residency.
• Mitigations: source-linked suggestions only, deterministic support checks, a log of AI vs. final text, red-team the tool on your own past cases, and keep strategy decisions with the attorney.

Getting started

1. Pick 10–20 recently filed applications and answered office actions (EPO and DPMA, plus USPTO if relevant) as a gold set, including the original disclosures.
2. Pilot 2–3 Path A tools (e.g. Solve Intelligence, DeepIP, IPRally drafting) under NDA with EU hosting. Compare AI drafts and responses against what your attorneys actually filed.
3. Turn on proofreading checks (antecedent basis, numerals, support) first. They are low-risk and show value immediately.
4. Build firm templates for EPO problem-solution responses and German DPMA style. Feed in successful past arguments.
5. Write an internal AI policy covering client consent, permitted tools, export-controlled cases (on-prem only) and verification duties. Measure hours per case and objections per examination for two quarters.

Sources

1. Solve Intelligence — product, security and customer page — https://www.solveintelligence.com/
2. DeepIP — AI patent drafting & prosecution platform — https://www.deepip.ai/
3. Patlytics — AI patent platform, case studies — https://www.patlytics.ai/
4. PatentPal — claims-to-specification and figure generation — https://patentpal.com/
5. Patent Bots — prosecution & proofreading software — https://www.patentbots.com/
6. ClaimMaster — patent drafting & proofreading Word add-in — https://www.patentclaimmaster.com/
7. IPRally — search, drafting and prosecution platform — https://www.iprally.com/
8. Wang et al., "AutoPatent: A Multi-Agent Framework for Automatic Patent Generation", arXiv:2412.09796 — https://arxiv.org/abs/2412.09796
9. Suzgun et al., "The Harvard USPTO Patent Dataset (HUPD)", arXiv:2207.04043 — https://arxiv.org/abs/2207.04043
10. USPTO — Guidance on Use of Artificial Intelligence-Based Tools in Practice Before the USPTO, 89 FR (11 Apr 2024) — https://www.govinfo.gov/content/pkg/FR-2024-04-11/pdf/2024-07629.pdf
11. Zuo et al., "PatentEval: Understanding Errors in Patent Generation", arXiv:2406.06589 — https://arxiv.org/abs/2406.06589
12. EPO — Article 84 EPC (Claims) — https://www.epo.org/en/legal/epc/2020/a84.html
13. EPO — Article 123 EPC (Amendments) — https://www.epo.org/en/legal/epc/2020/a123.html
14. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj