AI Patent Drafting Software: Who It's Actually For
AI patent drafting software is built for the patent attorney or agent who writes applications and office action responses, and it earns its keep on the first draft: turning an invention disclosure into a structured specification, generating figure descriptions, checking antecedent basis and term consistency, and proposing claim sets for a practitioner to rework. It is a different purchase from an IP intelligence platform, which serves the analyst mapping prior art and competitor portfolios. Solve Intelligence and DeepIP sit on the drafting side, PatSnap and IPRally on the search and landscape side. Nobody in this market publishes seat prices, the productivity figures are vendor-reported, and the USPTO has said plainly that the practitioner, not the tool, answers for every claim that gets filed.
What can AI legitimately do in patent drafting?
A patent application is a long, heavily cross-referenced document with rigid structure. That is exactly the kind of text large language models are good at producing, and it is why drafting was one of the first corners of legal work to get dedicated AI products.
The work AI does well is the work a senior practitioner would happily hand off:
Specification from disclosure. Turning an inventor's notes, slides and a call transcript into background, summary and detailed description sections that follow the house structure.
Consistency and formalities. Every claim term needs support in the specification, every reference numeral needs to match a figure, and every dependent claim needs a valid antecedent. Checking that by hand across a 60-page draft is tedious and error-prone. Software does it in seconds.
Alternative embodiments and boilerplate. Expanding a described mechanism into variants, and dropping in the firm's standard language for definitions and scope.
Office action first passes. Mapping an examiner's rejection to the cited references and producing a structured response outline.
The work AI does less well is the part that decides whether the patent is worth anything. A claim has to be broad enough at the point of novelty to be hard to design around and narrow enough to survive the prior art, and it has to be written with an eye on how a court will construe each term years later. That judgment depends on the client's business, the competitive landscape and the practitioner's read of the art unit. Models can produce claims that are grammatical and properly structured. Whether those claims are the right claims is still a human call.
The law points the same way. 35 U.S.C. 112 requires the specification to conclude with claims "particularly pointing out and distinctly claiming" the invention, and the written description and enablement requirements in the same section are where an AI-expanded specification is most exposed, because generated text can describe variants nobody actually conceived or tested.
Drafting copilots vs IP intelligence platforms: which are you buying?
This is the split that most "best tools" lists blur, and it matters more than any feature comparison.
Drafting copilots are bought by the people who write and prosecute applications: outside patent counsel and in-house patent attorneys and agents. Their job is output. They are judged on how much of the first draft survives review and how many hours come off each application. Solve Intelligence, DeepIP and Rowan Patents live here.
IP intelligence platforms are bought by people who need to understand a technology space: in-house IP analysts, R&D strategy teams, freedom-to-operate and invalidity specialists. Their job is insight. They are judged on the size and freshness of the patent corpus, the quality of search and the analytics layered on top. PatSnap and IPRally live here.
The categories overlap at the edges, which is where the confusion starts. PatSnap now markets IP drafting agents alongside its novelty, FTO and design FTO search agents, and Solve Intelligence lists prior-art and non-patent literature search and freedom-to-operate analysis in its own feature list. But each product's center of gravity is obvious once you look at who it was built for.
The practical test: if the person who will use the tool every day bills time for writing claims, you are buying a drafting copilot. If that person produces landscape reports, search results and FTO memos for others to act on, you are buying an intelligence platform. A firm that needs both usually ends up with one of each, and that is fine. Buying a landscape platform because it has a drafting tab, or a drafting copilot because it has a search box, is the common mistake.
How do the leading tools differ?
The scorecard below uses what the directory lists and what each vendor states on its own site as of September 2026. Where a vendor does not state something clearly, the cell says so rather than guessing. That is itself a finding: the gaps are the questions to ask in the demo.
| Tool | Drafting depth | Prior art and landscape | Microsoft Word workflow | Jurisdiction coverage | Pricing transparency |
|---|---|---|---|---|---|
| Solve Intelligence | Full applications, office action responses, invention harvesting, claim charts | Lists prior-art, non-patent literature and FTO search | Drafts in its own web workspace; states Word compatibility for output | States review against MPEP and EPO Guidelines; lets customers choose where data is stored (e.g. US or Europe) | On request |
| DeepIP | Full application generation, office action responses with case law citations | Prior-art search and FTO clearance; states 150M+ patents indexed | Microsoft Word add-in | States formal-requirement checks for USPTO and EPO plus other major offices including JPO, CNIPA, KIPO and UKIPO | On request |
| Rowan Patents (Clarivate) | Integrated environment building spec, claims and figures together, plus a review module | Not its focus | Own drafting environment; confirm Word round-trip in a demo | States US and European technology and jurisdiction templates | On request |
| PatSnap | Drafting agents offered as one module of a larger platform | Core product: patent search, analytics, novelty and FTO search, landscapes | Not stated as a focus; ask | Global patent data is the product | On request |
| IPRally | None; search, review and classification | Core product: graph-based search that shows why a result matched | Not applicable | Search corpus rather than drafting rules | Free trial; paid on request |
Three things stand out.
Where you draft is a workflow decision, not a feature. DeepIP is a sidebar inside Word. Solve Intelligence and Rowan Patents are environments you move your drafting into. Neither approach is better in the abstract. A practice with deep Word macros, templates and document management integration will feel the switch; a practice that wants every draft, figure and claim chart in one system will value the environment.
Jurisdiction claims need testing, not reading. "EPO compliant" can mean formal checks on claim format or it can mean drafting that anticipates problem-solution analysis and the strict added-matter standard. Those are very different levels of help. Put a European-first application through the pilot if you file in Europe.
Nobody publishes a price. Every commercial tool in the table is demo-gated. That pattern runs across legal AI, which we lay out in the legal AI pricing guide. Any "price" you see quoted for these tools outside a signed order form is an estimate.
What do the productivity claims really mean?
The marketing numbers in this category are large and not directly comparable. Solve Intelligence's homepage reports "50%+ average workload saved with Solve products," alongside 700+ IP teams, 433K+ patent applications created and 103K+ office action responses created. DeepIP's homepage claims "up to 70%" drafting time reduction on a full patent application and "2h+" saved per attorney per day. On the same page, DeepIP quotes Philips' head of IP describing "approximately 20% improvement in efficiency for drafting and prosecution" observed during a trial.
That last figure is the most useful number on either site, precisely because it is the smallest. It comes from a named customer measuring during a trial rather than from a best case. "Up to" describes the ceiling. "Average workload saved" describes a mean across a customer base whose baseline nobody outside the vendor can see.
To verify a productivity claim for your own practice, you would need:
A real baseline. Hours per application from your own time records, split by technology area. A software application and a biotech application with sequence listings are different jobs.
Review time counted. If a first draft arrives in an hour but takes the supervising attorney four hours to fix, the saving is smaller than the drafting clock suggests. Rewriting weak claims can take longer than writing strong ones from scratch.
Downstream quality tracked. The real cost of a weak draft shows up months later as 112 rejections, narrower allowed claims or an extra round of prosecution. A pilot measured only at filing will overstate the benefit.
Enough volume. Ten applications is an anecdote. Aim for a quarter's worth of comparable matters before you believe a number, including one your team has never seen used in a vendor demo.
None of this means the vendor numbers are wrong. It means they are claims about someone else's matters, measured someone else's way.
How should an IP practice pilot one?
The general method in how to evaluate legal AI applies. Five points are specific to patent work.
Start with confidentiality, before any demo data. The USPTO's guidance on AI-based tools in practice before the Office warns that invention details entered into AI systems can be retained, used for training or exposed in a breach, and that servers outside the United States can raise export control and foreign filing license issues. Read the tool's terms, data retention and hosting location before you paste a single unfiled disclosure. Some vendors now let you choose the storage region, and DeepIP cites SOC 2 Type II and ISO 27001 certification. Ask for the documents, not the badges.
Pilot on closed matters first. Run applications you have already filed through the tool and compare its draft to what you actually filed. You know the right answer, so you can measure what it missed.
Test the claims separately from the specification. Score the generated specification for accuracy and support. Score the generated claims on whether you would file them without rewriting. Expect very different results.
Watch for invented embodiments. The same USPTO guidance notes that AI can introduce alternative embodiments the inventors did not conceive, and that the practitioner must then check inventorship and 112 support before filing, because fixing it afterward may be new matter. The inventorship framework itself has since changed: the Office's revised inventorship guidance for AI-assisted inventions, published November 2025, rescinded the February 2024 approach and applies the same conception-based standard to every invention, whether or not AI was used.
Check the IDS path. If the tool also gathers prior art, remember that a person signing an Information Disclosure Statement certifies a reasonable inquiry into every reference on it. An AI-assembled reference list still has to be read.
A fair pilot takes a quarter, costs real attorney hours, and should end with a written go or no-go against numbers you set at the start. Our full IP bench, with what each tool covers, is at the IP and trademark category. If trademark clearance is also on your list, the split in that market is different, and we cover it in AI trademark search tools. We do not sell any of these tools.
Common questions about AI patent drafting software
Can AI write patent claims?
It can write claims that are correctly structured, grammatical and supported by the specification it drafted. What it cannot reliably do is choose the right scope: broad enough at the point of novelty to matter commercially, narrow enough to survive the art, and worded with later claim construction in mind. Treat AI-generated claims as a first proposal from a capable junior, useful for coverage of variants and dependent claims, and expect the independent claims to be rewritten by the practitioner who signs the filing.
What is the difference between Solve Intelligence and PatSnap?
They are built for different people. Solve Intelligence is patent drafting and prosecution software for attorneys and agents, covering applications, office action responses, invention harvesting and claim charts. PatSnap is an IP intelligence platform whose core is patent search, analytics, landscapes and freedom-to-operate work, with drafting agents added as one module. If your team writes applications, start with the drafting copilots. If it researches technology spaces, start with the intelligence platforms. Both are listed with details in the IP and trademark category.
Do patent AI tools work in Microsoft Word?
Some live inside it and some export to it. DeepIP runs as a Microsoft Word add-in, so drafting stays in the document you already use. Solve Intelligence drafts in its own web workspace and states that its output is Word-compatible for editing and review. Rowan Patents is its own integrated drafting environment. If your firm depends on Word templates, macros or a document management system, test the full round trip, including tracked changes and numbering, during the pilot.
Is AI-drafted patent text safe to file?
Only after a practitioner has verified it. The USPTO's position is that existing duties apply unchanged: the signer certifies a reasonable inquiry, claims known to be unpatentable must not be filed, the specification must be technically accurate and satisfy 35 U.S.C. 112, and inventorship must reflect human conception. The tool changes how the draft is produced. It does not change who is responsible for what gets filed, and it does not remove the confidentiality questions that come with sending unfiled inventions to a third-party system.
This article is software evaluation and industry analysis. It is general information, not legal advice.