How Do You Build an IDS from Unstructured Documents?

Almost never from a clean list. In a typical US application with foreign counterparts, the references arrive in whatever form the sender had:

• Foreign office actions and search reports, often as scanned PDFs, sometimes not in English
• International search reports and written opinions from the PCT stage
• Emails from foreign associates, with citations typed into the body or attached
• Papers and product literature the inventor sends over
• Citations from related US applications in the same family

The duty of disclosure under 37 CFR 1.56 applies to all of it. None of it arrives in the format the IDS form asks for.

37 CFR 1.98(b) sets the identification fields for each type of reference. Every field has to be read off a source document and typed correctly.

On top of the fields, the form needs a legible copy of each foreign document and each NPL item, and a concise explanation of relevance for anything not in English. Those come from the same scattered sources.

Because each source formats the same reference differently, and the form accepts only one format. The common slips:

• Number formats. Foreign offices write document numbers with and without country codes, kind codes and separators. A mistyped digit cites a different document.
• Dates. Search reports give dates in local formats. A day and month swapped is a wrong citation.
• Duplicates. The same reference is cited by two or three offices, sometimes as different family members. Listed twice, it inflates the IDS; dropped by mistake, it is never disclosed.
• Incomplete NPL. An associate's email gives an author and a title, but no publisher, pages or date.
• The date of the foreign communication. Whether a 1.97(e) statement is available depends on when a foreign office first cited the reference. If that date is not captured at intake, it has to be reconstructed later, under deadline.

None of these are hard. They are just easy to get wrong when done by hand, one reference at a time.

Yes, and the USPTO has said so. Its April 2024 guidance on AI-based tools names automatically populating the PTO/SB/08 with citations as one permitted use. The same guidance sets the limits:

• A person signs, and the signature means review. Signing an IDS certifies a reasonable inquiry under 37 CFR 11.18(b). That includes reviewing each piece of prior art listed, not just the form. Filing an IDS without reviewing its contents may violate the rule.
• More is not better. The Office flagged the risk that AI-populated IDSs grow larger and fill up with cumulative or irrelevant references. Practitioners are expected to remove what is clearly irrelevant.
• Materiality stays with the practitioner. Deciding what is material to patentability under 1.56 is a judgment the guidance leaves to people.

There is a second reason to keep the list tight. In Ecto World v. RAI Strategic Holdings (IPR2024-01280, 2025), the Director made it harder to institute review on art already listed in an IDS, but not where that art sat in a voluminous one.

So the useful job for AI is narrow and mechanical: read the documents, extract the fields, and show where each one came from, so the attorney's review is fast and actually happens.

ClaimxArk turns the unstructured material references arrive in into a structured IDS:

1. You give it what you already have: emails, PDFs of foreign office actions and search reports, NPL attachments.
2. It reads the documents and extracts the citation data for each reference.
3. It fills a PTO/SB/08-format IDS.
4. Each entry is shown next to the source passage it came from, so you can check it against the original in one look.
5. The attorney reviews, edits and signs.

It does not decide what is material, and it does not file anything on its own. It removes the retyping, which is where the errors were.

ClaimxArk was built by a patent attorney with over ten years of US prosecution experience. See the two-minute demo.