Up: AI Office Action response software

An Office Action response workflow with the attorney in control at every gate

Responding to an Office Action has a natural order: understand what the examiner said, decide what to do about it, write it, check it, file it. Most of the ways a response goes wrong are failures of order rather than of skill: an argument drafted before the cited passage was read, a claim narrowed before the history was checked, a listing typed from memory of the last amendment. This guide explains the seven-step workflow PatentReply keeps with a matter, and why each step waits for the one before it. Step-by-step instructions are in the documentation: The response workflow.

Scope: The workflow as PatentReply runs it inside Claude, ChatGPT, Grok, or another MCP client, for a U.S. application with a pending Office Action. Nothing here is legal advice, and the response is the practitioner’s to decide, sign, and file.

1. Compile the matter

The workflow begins by giving PatentReply the application number, not by uploading documents. It fetches the file history from the USPTO’s public record, indexes every paper, reconstructs every claim set ever presented so the operative text of a claim can be stated as of any legal date, identifies the outstanding action and its grounds, retrieves the cited references in full, and flags what it could not resolve: disputed claim text, a paper the parser could not read, a reference it does not hold. The first compile takes a few minutes.

Why first: everything downstream reads from this record. An analysis built on whichever PDFs were uploaded inherits their gaps; a matter compiled from the wrapper does not depend on anyone remembering what to include. The matter then persists across conversations, clients, and prosecution events, so this step happens once per application.

2. Review the Office Action

The review accounts for every objection and rejection with its statute, the claims it reaches, and the examiner’s reasoning as written; resolves each citation to the stored passage it points at and sets the examiner’s assertion beside it; lists the evidence gaps; and reports the history constraints, what was argued before and what scope was given up. It ends with a readiness note: whether the matter is ready for traversal drafting and amendment drafting, and what blocks either.

Why verification precedes strategy: a plan built on an unverified characterization of the art is a plan built on the examiner’s summary, or the model’s. Reading the cited passage first is what tells the attorney whether the right move is to argue, to amend, or both. The guide on AI Office Action analysis covers this step in depth.

3. Develop the strategy

For each issue, the attorney and the AI develop options, traverse, amend, both, concede and cancel, request an interview, preserve for appeal, each with a theory, its advantages and risks, and the record passages it rests on; arguments framed and anchored to the record; and amendment proposals as full text per claim, pinned to the version they amend, with a scope-cost note and an estoppel note. The plan is stored with the matter, per issue and revisioned, so it survives the conversation and can be read back in any client.

Why stored rather than remembered: a strategy that lives only in a chat window is lost when the window closes and cannot be approved, audited, or handed to a colleague. Storing it per issue also makes completeness checkable, so that every ground has a decision before anything is drafted.

4. Attorney approval

When the plan is settled, it is confirmed as a frozen revision and the practitioner is handed a link to an approval page on patentreply.ai. Approving there, signed in, is the legal sign-off, recorded with the matter and the revision it approved. A yes typed into the chat is not approval, and the model cannot approve on the attorney’s behalf. Pending approvals wait on the site’s Approvals page.

Why a hard gate, and why on the site: the decision to amend or to argue is the practitioner’s, with consequences the practitioner owns. A gate the model could pass by agreeing with itself would be no gate. Putting it on a page the attorney signs into, with the plan as stored, makes the decision explicit and attributable, which is also what a firm standardizing AI use needs to show. The trust page on approvals and legal judgment sets out what requires approval and what stays with the practitioner.

5. Prepare amendments and Remarks

From the approved plan, PatentReply generates the claim listing under 37 CFR 1.121 deterministically from the proposed claim text, with the status identifiers and markings the rule requires. Your AI drafts the Remarks from the approved strategy, one issue at a time, citing the claim language, the examiner’s assertions, and the resolved passages; quotations are checked against the stored text, and every anchor the argument relies on has to resolve. PatentReply then assembles the response document (.docx): caption, the 1.111 or 1.116 title, the listing, the Remarks, conclusion, and signature block.

Why the listing is generated and not drafted: the rule’s format is mechanical and unforgiving, and a listing assembled by a model from claim text it read is exactly where a transcription error would enter the filing. The guide on claim amendments in an Office Action response covers the decision to amend, the support, the history, and the listing in depth.

6. QC against the record

Before anything is signed, an independent pre-filing check reads the draft against the prosecution record: the caption and posture, the listing, the amendments against the operative claims, the evidence behind each assertion about a reference, and whether every live issue is addressed. Every finding is anchored to a passage or marked unverified; a negative assertion such as “the reference does not disclose X” stays unverified until the reference has been read in full, because a search miss is never proof of absence. The draft can come from anywhere, including the generated document after the attorney’s edits or a document that never touched PatentReply.

Why a separate check: the drafter is the wrong party to find the drafter’s errors, and the check should measure the draft against the record as it stands on the day of filing, not as it stood when the plan was written. QC never advances the workflow; it is a reading of the draft, not a step toward filing it.

7. Filing by the practitioner

The attorney reviews the assembled response, edits it, signs it after the reasonable inquiry the rules require, and files it with the USPTO. PatentReply does not file, does not compute the response period, and does not give legal advice; it reports the action and its mail date and tells you to confirm the period against your docket. The signed download is the end of PatentReply’s part.

Why filing stays with the attorney: the practitioner of record is the one who certifies the paper and answers for it. A workflow that kept that act with the practitioner by design is one a firm can adopt without rewriting who is responsible for what.

Returning for the next Office Action

The response you filed becomes part of the record. When the next action arrives, whether a final rejection under MPEP § 706.07, an allowance, or a decision on appeal, refresh the matter and the new papers join the same record. The workflow starts again at review, with what you argued and what you amended available as history constraints, and the compiled matter you return to is the one you left, in whichever AI client you open it. Asking where you are on the response returns the live stage from stored state, not from anyone’s memory of the conversation.

Up: AI Office Action response software