Blog

How to Find Prior Art for a Software or AI Invention

· Published November 16, 2026

How to Find Prior Art for a Software or AI Invention
On this page
  1. Start with the right kind of search
  2. Describe the software or AI invention as a feature map
  3. Search in layers instead of relying on one query
  4. Useful search platforms
  5. Worked example: search an AI invention by mechanism
  6. Review each finalist in full
  7. Carry the search into claim strategy
  8. Choose the next action

How to Find Prior Art for a Software or AI Invention

To find prior art for a software or AI invention, first establish the relevant date and break the proposed invention into technical features. Then search patents and non-patent literature using multiple term sets, classifications, citations, patent families, and adjacent technical fields. Review the strongest documents in full, verify their public-availability dates, and map each reference to the proposed claim features.

The result is not proof that no prior art exists. It is a documented basis for deciding what claim scope and fallback positions may be worth pursuing with patent counsel.

A patentability-oriented prior-art search asks what was disclosed before the relevant date. Under U.S. law that includes public disclosures and, under 35 U.S.C. § 102(a)(2), certain U.S. patents and published applications that were effectively filed earlier, even if they were published later. MPEP § 2152 It is different from a freedom-to-operate search.

A prior-art search focuses on technical disclosures that may affect patentability. A freedom-to-operate analysis concerns potentially enforceable patent claims, territorial scope, and legal status. The same patent document can provide technical information in its description and drawings, legal information in its claims and status, and bibliographic information about its filing and priority history. The WIPO Guide to Using Patent Information explains these different types of information.

Before searching, establish three boundaries:

  1. The invention disclosure: What system, process, architecture, or technical method is actually proposed?
  2. The relevant date: What filing, priority, or other date should counsel use for the jurisdiction and claim under review?
  3. The claim directions: Which features may be essential to an independent claim, and which could serve as fallback positions?

For a U.S. application subject to the America Invents Act, 35 U.S.C. § 102(a) includes patents, printed publications, public use, on-sale activity, and other public availability before the claimed invention’s effective filing date. Exceptions can apply. The USPTO’s MPEP § 2152 describes these categories and the relevant exceptions. This is U.S.-specific guidance, not a universal rule for India or every other jurisdiction.

For an India-based reader, do not carry the U.S. date analysis across unchanged. Use the search to identify the invention’s technical mechanism and technical effect, then have Indian patent counsel apply the Patents Act and current Indian practice to the subject-matter and prior-art questions. Section 3(k) of the Patents Act, 1970 excludes a computer programme per se and algorithms from patentability, so frame the search around the technical mechanism and effect rather than the software as such. Patents Act, 1970, Section 3 A product label such as “AI assistant” is not a useful search boundary in either jurisdiction.

Describe the software or AI invention as a feature map

“AI fraud detection platform” is a product description. It is difficult to search and too broad to guide claim analysis.

A better starting point is a feature map that describes:

  • The technical problem
  • The system components
  • The data entering the system
  • The transformations performed on that data
  • The relationships between components
  • The process steps
  • The output or technical result
  • The constraint or technical effect
  • Alternative implementations

For example, an AI invention might involve:

  • Extracting prompt features from an incoming request
  • Predicting which model shards will be needed
  • Pre-warming selected shards
  • Routing the request to an inference server
  • Updating shard rankings from cache-hit and cache-miss telemetry

Those details produce search concepts that are more useful than “AI model optimization.”

Use the Claim-Direction Search Canvas

The following Claim-Direction Search Canvas is an editorial search framework, not a legal test.

Canvas field What to record
Technical problem What system limitation is being addressed?
Required mechanism What components, data transformations, or control steps must occur?
Inputs and outputs What enters the system, and what technical result leaves it?
Technical effect What changes in latency, accuracy, bandwidth, memory, security, or system operation?
Variants Synonyms, older terms, implementation alternatives, and adjacent-domain names
Claim priority Essential independent-claim feature, fallback position, or background context

The USPTO identifies text searching, patent-classification searching, and patent-citation searching as the three parts of a complete patent search. Its examiner guidance also describes reading the specification and claims, developing search terms and classification fields, refining the results, and searching non-patent literature. See the USPTO’s Patent Searching and Search Resources presentation and its Prior Art FAQ.

Search in layers instead of relying on one query

A single keyword query often fails because software and AI inventions can be described through different technical vocabularies. Search in layers so that each strong document improves the next search.

1. Find seed documents with focused keywords

Start with combinations that express a technical relationship, not merely a product category.

For each important feature, list:

  • Synonyms
  • Acronyms
  • Older terminology
  • Function-based descriptions
  • Component names
  • Process names
  • Broader and narrower terms
  • Terms used in adjacent industries

For the model-serving example, possible query families include:

("model prewarming" OR "predictive model loading") AND inference
("shard selection" OR "model partition") AND prompt
("cache miss" OR telemetry) AND model routing

Do not assume the inventor’s preferred terminology is the terminology used in earlier documents. A paper may describe “selective parameter loading” where the invention disclosure says “model shard pre-warming.” A patent may describe a “prediction-driven memory allocation process” instead of “preloading.”

The goal of the first round is to find strong seed documents. A useful seed can reveal better terminology, classifications, earlier family members, citations, and related non-patent literature. A large result count is not evidence of a strong search.

2. Expand through CPC and IPC classifications

Patent classifications can surface documents that use different words for similar subject matter. Extract CPC or IPC classifications from the strongest seed documents, then search those classifications with new technical terms.

Google Patents supports searches across title, abstract, claims, and the full document. It also supports CPC searches, field restrictions, Boolean syntax, and child classifications. Its Advanced Search page describes operators such as TI=, AB=, CL=, and CPC=.

For example, once a close document reveals a relevant classification, search that classification with terms such as:

  • Pre-warm
  • Load prediction
  • Shard selection
  • Model routing
  • Cache telemetry
  • Adaptive inference

Classification searching should not replace text searching. It is a way to widen the search after the initial terminology has been tested.

3. Follow citations and patent families

For each strong patent document, review:

  • Backward citations
  • Forward citations
  • Foreign counterparts
  • Earlier family publications
  • Related applications
  • Cited non-patent literature
  • Examiner-identified references where available

Patent families are especially important when the first document you find is a later publication. An earlier family member may contain an earlier public disclosure, or a foreign counterpart may use clearer technical language.

Record the publication date and priority information for each family member. Do not treat a family relationship alone as proof that every family member contains the same disclosure or has the same legal effect.

4. Search adjacent technical fields

The same mechanism may appear in another field under another name.

For software and AI inventions, adjacent fields can include:

  • Computer networking
  • Databases
  • Distributed systems
  • Information security
  • Computer vision
  • Robotics
  • Industrial control
  • Hardware acceleration
  • Cloud infrastructure
  • Telecommunications

An invention described as “adaptive model routing” might have relevant disclosures in load balancing, distributed cache management, or network traffic control. The search should follow the technical function and relationship between components, not only the industry in which the product will be sold.

5. Search non-patent literature deliberately

Patent databases do not contain every important software or AI disclosure. Search non-patent literature alongside patents, including:

  • Research papers
  • Preprints
  • Conference proceedings
  • Theses and dissertations
  • Standards and standards contributions
  • Technical documentation
  • Product manuals
  • Datasheets
  • Model cards
  • Public code repositories
  • Release notes
  • Public implementation guides
  • Archived web material

The USPTO’s software-related search resources include IEEE Xplore, ACM, Inspec, Ei Compendex, Dissertation Abstracts, MathSciNet, and ScienceDirect. Its Prior Art FAQ explains that examiners use non-patent literature databases as part of software and computer-related searches.

For AI inventions, search the places where technical teams actually publish:

  • A preprint server for an early model or method
  • A conference paper for an architecture or training technique
  • A model card for intended use, limitations, and technical configuration
  • A code repository for an implementation
  • A standards contribution for a proposed protocol feature
  • Product documentation for an exposed system behavior
  • A release note for a dated implementation change

Treat timestamps as evidence to verify, not as automatic proof of prior-art status. A repository commit, archived page, conference presentation, or model card must still be assessed for public accessibility, date, technical content, and legal effect in the relevant jurisdiction.

The Esgenix article What a Patent Search Misses: Finding Non-Patent Literature Before You Draft provides a more detailed source-by-source treatment of software repositories, preprints, standards, documentation, and other non-patent literature.

Useful search platforms

Use more than one source when the matter is important.

Potential starting points include:

  • USPTO Patent Public Search
  • Google Patents
  • WIPO PATENTSCOPE
  • Espacenet
  • IEEE Xplore
  • ACM Digital Library
  • Relevant standards repositories
  • Public preprint servers
  • Dated technical repositories

Google Patents is useful for field-specific and classification searches. WIPO PATENTSCOPE provides access to international and national patent collections and supports patent and non-patent literature searching. Its page states that the service searches more than 130 million patent documents, including published PCT applications.

These tools are starting points, not substitutes for reviewing the source documents. Search interfaces differ in coverage, indexing, family treatment, translation, and date presentation. For how far to take any search and when to stop, see what a thorough prior art search includes.

Worked example: search an AI invention by mechanism

Consider this hypothetical invention:

An AI inference-serving system predicts which model shards will be needed from prompt features, pre-warms only those shards before routing, and updates future shard rankings from cache-hit and cache-miss telemetry.

Step 1: Identify the technical problem

The problem is reducing GPU cold starts and inference latency without preloading every model shard.

Step 2: Extract searchable features

The feature map includes:

  • Prompt-feature classification
  • Prediction of a likely shard set
  • Selective model pre-warming
  • Inference routing
  • Telemetry-based ranking updates

Step 3: Search several query families

Possible searches include:

("model prewarming" OR "predictive model loading") AND inference
("shard selection" OR "model partition") AND prompt
("cache miss" OR telemetry) AND model routing

Then search the classifications revealed by close documents with terms such as:

  • Pre-warm
  • Load prediction
  • Model partition
  • Routing
  • Cache hit
  • Cache miss
  • Inference latency

Step 4: Search non-patent sources

Search conference papers, preprints, model-serving documentation, model cards, public implementations, and dated product documentation for the same mechanisms.

A repository may not use the phrase “model shard ranking.” It may describe a scheduler that predicts model partitions from request metadata. That may be a lead requiring technical review.

Step 5: Map the references

Proposed claim feature Reference A Reference B Search decision
Predict likely model shard set Disclosed Not disclosed Assess Reference A closely
Pre-warm selected shards Partial Disclosed Search variants and combinations
Update rankings from telemetry Not disclosed Partial Preserve as a potential fallback

“Disclosed” in this table means that the reference appears to teach the feature and requires further analysis. It does not mean that a claim is invalid. Counsel must assess the reference’s content, date, enablement, and the applicable novelty or inventive-step standard.

The table also shows how a search can shape claim strategy. If selective pre-warming is widely disclosed but telemetry-based ranking updates are not, the latter may deserve closer attention as a possible fallback, assuming the invention disclosure supports it.

Review each finalist in full

Abstracts and search snippets are useful for triage. They are not substitutes for the underlying document.

For each serious reference, review:

  • The claims
  • The detailed description
  • Relevant embodiments
  • Definitions and alternatives
  • Drawings and system diagrams
  • Cited references
  • Priority and publication information
  • The exact passages supporting each mapped feature

Stanford’s Office of Technology Licensing recommends skimming the first pages of patent documents to eliminate clearly irrelevant records, then reading the documents that appear relevant in detail. Its basic prior-art search guidance describes this progression from keyword searching to first-page review and detailed reading.

Create a search record for every finalist:

  • Full citation and stable source location
  • Publication or public-availability evidence
  • Priority and family data for patent documents
  • Exact passages, claims, figures, code, or documentation sections relied on
  • Feature status: disclosed, partial, or absent
  • Search strings and databases used
  • Classifications searched
  • Date of each search
  • Reason for including or excluding the reference

Do not rely on a later summary, machine-generated snippet, or citation to another paper when the primary document can be reviewed.

A stopping decision should be tied to the drafting or investigation decision the team needs to make. Pause when defined claim directions have been searched across patents and relevant non-patent literature; strong references have been expanded through classifications, citations, and families; finalists have feature maps and date checks; and the team can state the next drafting or investigation decision. Reopen the search if claim scope, terminology, jurisdiction, or the invention disclosure materially changes.

Carry the search into claim strategy

The useful output is not a stack of PDFs. It is a feature map that shows:

  • Which features appear in individual references
  • Which features appear across different references
  • Which terminology the field uses
  • Which claim directions remain supported
  • Which fallback positions deserve development
  • Which technical gaps require clarification
  • Which parts of the disclosure need stronger specification support

That record should reach the drafting team while the claims are still editable. If the search is completed only after the independent claim has been fixed, the strongest references may force avoidable amendments.

Product evidence: a search-to-claim workflow

The Esgenix Prior Art Search product workflow describes a process that begins by extracting searchable features from an invention disclosure. The user can review jurisdictions, classifications, and search features before search rounds run. References are graded as X, Y, or A, and top references are mapped against invention features cell by cell with supporting evidence quotes. The user directs the search rounds and approves the report. This is product evidence about the workflow, not an independent performance benchmark.

Esgenix positions prior-art intelligence as something that can inform strategy and drafting within the same matter workflow. Its prior-art search offering describes agent-planned search rounds, user-directed triage, X, Y, and A grading, feature matrices with supporting evidence quotes, and a report that can continue into drafting. See Esgenix Prior Art Search.

Within a decision-gated workflow, the professional still makes the substantive decisions. At Intake, the professional confirms the invention and resolves technical gaps. At Strategy, the professional sets claim direction, breadth, and fallback positions. The professional then reviews the claim set at Claims, checks the specification and figures at Specification & Figures, and completes professional review at Sign-off.

In Esgenix, graded references can continue into drafting in the same matter. It does not provide prosecution or office-action response.

Choose the next action

If you are a founder, inventor, or smaller patent team preparing an invention for counsel review, Create your Esgenix account and start with the invention disclosure and feature map.

For a corporate IP team or law firm that wants to inspect the search-to-claim workflow on a real matter, request a demo.