A raw patent office action PDF costs more tokens because your AI assistant has to read the whole USPTO document as issued: the Office Action Summary cover sheet, the Notice of References Cited, the 35 U.S.C. 102, 103, and 112 boilerplate repeated under every rejection, the signature block, and whatever the scan did to the layout. The part a response actually turns on, which claims were rejected on which grounds over which art and why, is a fraction of the page count. The Juristat Data Layer (MCP) puts that fraction, and the examiner's full prosecution record, inside the AI tool you already use as data the model requests rather than a document it reads, at roughly 85 percent fewer tokens than raw OA PDFs.
Key findings
- Most of a patent office action PDF is form, not argument. The rejections are the payload; the rest is packaging the model pays to read.
- A scanned PDF adds cost on top of content: page headers repeated per page, flattened tables, broken line wraps, and text-searchable artifacts all become tokens.
- The document is charged on every turn of the conversation, not once. A 30-page PDF is 30 pages again when you ask the follow-up question.
- A raw PDF carries no examiner history. It costs more and tells the model less than the prosecution record does.
What is actually on a USPTO office action PDF?
A USPTO office action PDF is a form packet wrapped around the examiner's reasoning. A typical non-final rejection carries an Office Action Summary cover sheet (form PTOL-326) with checkbox dispositions of the claims, a Notice of References Cited (form PTO-892), the examiner's rejections written claim by claim, response-period and interview language, and a conclusion with the examiner's contact details and signature.
Inside the rejections, the examiner relies heavily on form paragraphs. A 103 rejection opens by reciting the statute, then states the standard, and only then gets to the mapping of the cited art onto your claims. That mapping and the examiner's reasoning are what your response will engage with. The recitations are identical across thousands of office actions, and your model reads them fresh every time.
Add the attachments. When cited references travel with the action, the model gets every page of them too, whether or not the rejection relies on more than one paragraph of each.
Why does an LLM pay tokens for the whole office action?
LLM usage is metered in tokens, and a PDF has to be turned into text before a model can read it. Everything on the page becomes input: the form checkboxes, the barcodes, the repeated headers, the page numbers, the mailing-date stamps. The model doesn't know a header from a holding until it has read both.
Scanned documents are worse. Text-searchable conversion is good enough for a human reader skimming to the rejection, but it leaves line breaks mid-sentence, flattens two-column layouts into interleaved lines, and turns a reference table into a run of fragments. Each fragment is still tokens, and each one is noise the model has to sort through before it can reason.
The result is a document where the examiner's actual argument sits inside a much larger block of material no response will ever cite. You're paying full price for the packaging.
Why does the office action cost tokens again on every follow-up?
A chat with an office action in it is not one read. The PDF stays in the conversation, and each time you ask a follow-up ("what if we amend claim 1 to add the limitation from claim 4?"), the model takes the whole context as input again. Input tokens are billed per call, so a 30-page PDF is 30 pages on the first question and 30 pages on the fifth.
Length has a second cost. The longer the context, the more places there are for a model to weight the wrong passage, and the form paragraphs give it plenty of plausible-sounding text to weight. A tighter context doesn't just cost less to read. It leaves less room for guesswork.
How to cut the token cost of a patent office action response
- Upload the rejections, not the packet. If you're starting from the PDF, remove the cover forms, the references list, and the signature pages before the model sees it.
- Leave the cited references out of the context until you need one. Bring in the paragraph the examiner relied on, not the whole reference.
- Start a new conversation per office action. A chat that still holds last week's PDF bills for it on every turn.
- Stop uploading, and let the model ask. Connect your AI tool to the prosecution record so it requests the office action, the examiner, and the art unit as data, and gets back only what it asked for.
The first three are housekeeping. The fourth is where the real saving is.
What changes when your AI tool asks for the office action instead of reading it?
The Juristat Data Layer (MCP) is a remote, read-only connection that puts more than 20 years of USPTO prosecution data, across 10M+ applications, inside Claude, ChatGPT, Microsoft Copilot, and the other AI tools your firm already runs. Nothing to install, and it uses the same login and permissions as the Juristat web app.
The mechanics are the point. Instead of swallowing a PDF, the model makes a request: give me this office action, give me this examiner's report, give me the win rate by response type in this art unit. Each request returns a scoped answer, as text, drawn from a record Juristat has already parsed and structured. No cover forms. No form paragraphs. No layout artifacts. Only the fields the question needs, which is what the roughly 85 percent fewer tokens than raw OA PDFs comes down to.
It also returns what the PDF never had. Examiner allowance rate. Art unit trends. How often an interview, an amendment, an RCE, or an appeal has actually worked with this examiner. Sibling applications the same examiner already allowed. The same connection, the same conversation, no second tool. [Link: win rate by response type]
Here's what the model receives in each case.
|
What the model receives |
Raw office action PDF, uploaded |
Juristat Data Layer (MCP), requested |
|---|---|---|
|
Cover form, checkboxes, headers, signature block |
Yes, on every page |
No |
|
Statute recitations and form paragraphs |
Repeated under each rejection |
No |
|
The office action text |
Whole document, every turn |
On request, as parsed text |
|
Text-searchable layout artifacts |
Yes |
No |
|
Examiner allowance rate and report |
No |
On request |
|
Win rate by response type for this examiner or art unit |
No |
On request |
|
Similar applications and the amendments that were allowed |
No |
On request |
|
Private client data transmitted |
Everything in the PDF |
None; read-only, public record |
|
Re-read on every follow-up turn |
Full document |
Only what was requested |
Source: Juristat product documentation, Juristat Data Layer (MCP), as of November 2026. Token comparison per the cleared Juristat claim: roughly 85 percent fewer tokens than raw OA PDFs.
Why a cheaper office action is also a better-grounded one
Fewer tokens is the easy win. The bigger one is what the PDF can't carry. An office action, raw, says nothing about the examiner who wrote it: not the allowance rate, not how often an interview or an amendment or an RCE has worked in that art unit, not whether the same examiner already allowed a sibling application on claims that look like yours.
A general-purpose LLM working from the PDF alone will produce a competent, generic response strategy, because the law is the only thing it has. The same model connected to the prosecution record can tell you what has actually worked on this examiner, quantified, before you write a word. Accurate answers cost less than guesswork, and they cost less in a second way: fewer drafts, fewer re-reads, fewer rounds.
Dan Block, Director at Sterne Kessler, put the mechanism plainly after running Juristat data inside Claude: "Your MCP server returns the full data. That context is what lets you build the multi-step workflows like the white space analysis and do the things other vendors can't."
What token cost means for a flat-fee office action response
More corporate clients are moving outside counsel to flat fees that include office actions, which puts the margin on OA work under pressure. Tokens are now a line item in that margin, and every one spent on a cover form is one not spent on the argument. A model that pulls the rejection and the examiner's record in one request, inside the tool the firm already licenses, is the cheapest hour in the matter. [Link: flat fees that include office actions]
For in-house teams enabling AI tools across their outside counsel, the same connection does something else: it lets you control the context the model works from without dictating the conclusions.
The takeaway: a raw office action PDF is the most expensive way to put the least useful version of the office action in front of your AI assistant. Connect it to the record instead.
FAQ
How do you respond to a patent office action with an AI assistant without paying for the whole PDF?
Don't upload the packet. Connect your AI tool to the prosecution record so it can request the rejection, the cited art, and the examiner's reasoning as text, and get back only what it asked for. The Juristat Data Layer (MCP) does this at roughly 85 percent fewer tokens than raw OA PDFs, and adds the examiner's record the PDF never carried.
Does plain text or markdown use fewer tokens than a PDF in Claude, ChatGPT, or Copilot?
For the same content, yes, because a scanned PDF adds page headers, broken line wraps, and flattened tables that all become input tokens. The bigger saving comes from leaving out what the response won't use, which is most of an office action.
Which AI tools does the Juristat Data Layer work with?
Claude, ChatGPT, and Microsoft Copilot, and it also works with Harvey, Legora, and Perplexity. Setup for Claude takes a few minutes; Copilot is set up by your IT team. There is nothing to install.
Does the Juristat Data Layer draft the response?
No. It is the data and intelligence layer, not a drafting tool. It gives whichever model you use the office action, the examiner's record, and what has worked by response type. Your tool, and your judgment, do the drafting.
Does connecting it send our client data anywhere?
No. The Juristat Data Layer is read-only. It returns public prosecution data to your AI tool, does not modify data or take actions, and does not retrieve your private files.
What about OA Strategy Briefs?
Same data, packaged. An OA Strategy Brief is a compact, LLM-ready file Juristat generates automatically for each office action and delivers with Juristat Office Action Response (OAR). The Data Layer is the live version of that record, available on request inside your AI tool.