DeepIP Review

Patent drafting is one of the few kinds of legal writing where the form is almost as constrained as the content. Claims have a grammar, specifications have required parts, and jurisdictions have rules that are checkable rather than arguable.

That structure is exactly what makes the work suitable for automation, and exactly what makes automation dangerous. A specification that is merely competent is fine; a claim drafted slightly too narrowly is a defect that surfaces years later, in litigation, when it cannot be fixed.

DeepIP's approach is to cover the whole lifecycle rather than a single step, from the invention disclosure through to freedom-to-operate analysis.

DeepIP: Patents From Idea To Enforcement

• Five lifecycle stages: ideation, prior art, drafting, prosecution and risk assessment, in one platform.
• Full application generation from invention disclosures, with USPTO and EPO compliance checks and Word integration.
• Prior art search with semantic understanding, across a corpus the site describes with two different figures.
• Series B, $40M total raised, per the homepage.
• US-based Microsoft Azure deployment with SOC 2 Type 2 and ISO 27001, cited in a customer's own words.

Can AI Draft A Patent Application?

DeepIP states that it generates complete, jurisdiction-compliant patent applications from invention disclosures — claims, specifications and abstracts, structured and ready for review — with USPTO and EPO compliance checks and Microsoft Word integration.

The honest framing is that different parts of an application have very different risk profiles, and treating them identically is the mistake.

The specification is largely descriptive. It explains the invention, its embodiments and its context, at length and in a conventional structure, and it is exactly the kind of long structured prose where generation from a disclosure genuinely saves hours.

Claims are a different craft. Their scope determines what the patent is worth, and a claim that is slightly too narrow surrenders coverage a competitor will use, while one slightly too broad invites invalidation. Both errors are invisible on the page and expensive later.

The site's own phrasing — structured and ready for review — is the correct positioning and worth taking literally. The reasonable use is generation of the bulk with attorney judgement concentrated on claim scope, which is where the value of the professional actually sits, in the same way CoCounsel positions drafting as work product requiring verification.

A claim slightly too narrow surrenders coverage a competitor will use. One slightly too broad invites invalidation. Both errors are invisible on the page and expensive later.

Prior Art Search At Scale

DeepIP describes semantic prior art search, surfacing relevant results in minutes rather than hours, and patentability assessment producing structured novelty and non-obviousness evaluations with confidence scores.

Semantic search is a real improvement over keyword search here, for a reason specific to patents: the same invention is routinely described in completely different vocabulary across jurisdictions, decades and drafting styles.

A keyword search misses prior art that used different words for the same mechanism, which is the most dangerous kind to miss, because it is the kind an examiner or an opponent will eventually find.

Confidence scores on patentability assessment are useful and require careful reading. A confidence score reflects how certain the system is given what it found — not how certain you should be, which also depends on what it did not find.

The practical discipline is to treat a clean prior art result as a reason to continue rather than as clearance. Absence of evidence in a search is weaker than it feels, particularly across languages and older filings.

Two Different Patent Counts On One Site

The site displays 150M+ patents indexed and searchable in one place, while the prior art search module describes searching across 120M+ patents.

We are noting this as an observation rather than a criticism, because there are ordinary explanations. The two numbers may count different things — total indexed records against patents available to one specific module, or documents against families — and pages are updated at different times.

It is worth raising only because corpus size is one of the few objectively comparable specifications in this category, and it is the kind of number a buyer will quote internally.

The useful question in a demo is not which figure is correct but what the corpus actually contains: which jurisdictions, whether non-English filings are included and how, how far back coverage extends, and whether applications as well as grants are searchable.

Those details matter far more than the headline. A 150 million document index heavy on one jurisdiction is less useful for freedom-to-operate work than a smaller one with genuine international depth.

• Generate the specification freely; concentrate attorney time on claim scope.
• Treat a clean prior art result as a reason to continue, not as clearance.
• Ask what the corpus contains, not how large it is.
• Read confidence scores as certainty given what was found, not overall.
• Confirm where drafts are processed and stored before any disclosure goes in.

Security Claims In A Confidential Field

A customer quotation on the site describes extensive due diligence on DeepIP's security infrastructure, citing deployment on US-based Microsoft Azure servers together with SOC 2 Type 2 and ISO 27001.

Security matters more in patent work than in almost any other legal application, for a reason specific to the field: an unfiled invention disclosure is the most confidential document a technology company produces.

Public disclosure before filing can destroy patentability outright in many jurisdictions. That makes the handling of pre-filing material a substantive legal question rather than an IT procurement one.

Naming the hosting arrangement and specific certifications is the right level of detail, and citing it through a customer who performed diligence is more credible than asserting it directly.

What certifications do not tell you is retention and training use. Whether disclosures are retained after drafting, and whether any client material contributes to model improvement, are separate questions to get answered in writing — the same questions that matter for contract platforms such as LegalOn handling privileged material.

Reading The Efficiency Numbers

DeepIP publishes four figures: 150M+ patents indexed, 25k+ applications drafted with the platform, up to 70% drafting time reduction on a full application, and 2h+ saved per attorney per day.

The two efficiency claims are the ones that would drive a business case, and they are the ones with the least stated basis on the page.

Up to 70% is a ceiling rather than an average, and the phrasing is doing work — up to permits a distribution where most cases save considerably less. That is standard marketing language rather than anything improper, but it should be read as such.

2h+ saved per attorney per day is the more striking claim, since it implies a substantial fraction of a working day. Whether that reflects measured practice, projection, or best-case users is not stated.

The 25k+ applications drafted figure is the most useful of the four, because it describes something that happened rather than something saved. Volume of real use is weak evidence of quality but strong evidence that the product functions at scale.

This is a common pattern in legal AI, and worth recognising. Scale figures — documents indexed, signals processed, matters connected, as platforms like Darrow also publish — describe how much is happening rather than how often it was right, because outcome data in legal work is genuinely difficult to collect.

Trialling Patent AI Responsibly

A patent workflow is the wrong place to learn a tool's limits in production, because the errors are latent and surface at the worst possible moment.

This sequence tests the system against outcomes you already know before it touches anything unfiled.

1. Settle data handling in writing first — retention, training use and processing location.
2. Run it against a granted patent you drafted and compare claim scope carefully.
3. Run prior art search on a matter where you know what the examiner eventually cited.
4. Use generation for specifications first, keeping claims fully attorney-drafted.
5. Record which sections were machine-generated, for your own file.
6. Expand into prosecution responses only once drafting output has proved reliable.

DeepIP: Modules And Funding