Patent prosecution AI stack — drafting under the certification you sign

The USPTO answered the central question about AI drafting in April 2024, and most prosecution workflows still do not reflect the answer. Its guidance on AI-based tools in practice before the Office (89 FR 25609, April 11, 2024) makes the point plainly: relying on an AI tool’s accuracy is not, by itself, the inquiry reasonable under the circumstances that 37 CFR 11.18(b)(2) requires. Every paper you file carries that certification. The drafting tool does not sign it. You do.

That reorders the stack. The AI drafting layer is the part you can replace in a quarter — three well-capitalized vendors with overlapping feature sets are competing for the same seat. What decides whether AI-assisted prosecution survives an OED inquiry, an export audit, or a later inequitable-conduct fight is everything the drafting vendor does not own: where an unfiled disclosure is allowed to travel, who performed the human review before the signature went on, and whether the docket still owns the dates.

The shape: a patent-specific drafting tool (Patlytics, Solve Intelligence, or DeepIP) generates the text, Anaqua holds the portfolio record and the docket, iManage or NetDocuments holds the matter file, Intapp governs who may see which unfiled disclosure, and Claude for Legal covers the reading and reporting work the patent tools were never built for.

How the pieces fit

• The drafting layer is the swappable slot, and that is a feature. Patlytics closed a $40M Series B led by SignalFire on April 8, 2026, taking it to roughly $65M raised, and states use by more than 40% of Am Law 100 firms with Quinn Emanuel, Foley & Lardner, Canon, and Rivian named as customers. Solve Intelligence raised a $40M Series B to $55M total and shipped a claim-charting product alongside drafting. DeepIP raised a $25M Series B on March 3, 2026 to $40M total, then acquired Munich-based PatentMaker in June 2026 to take the German market. The functional split that actually matters at purchase is surface, not model quality: DeepIP runs inside Microsoft Word, where most patent attorneys already draft; Patlytics and Solve Intelligence are their own workspaces, which buys richer analysis and costs you the habit change.

• Anaqua is the system of record, and increasingly it is also your docket. Anaqua acquired Patrix on April 28, 2026 — the thirty-year-old Patricia platform with nearly 400 law-firm customers — after buying RightHub in 2025, and it added an IP-risk intelligence provider on August 4, 2026. Three acquisitions in eighteen months: if you run a mid-market IP management system, check who owns it now, because your docketing vendor may already be Anaqua. This layer earns its place because the drafting tool produces text while the system of record produces dates, and only one of those is a malpractice exposure.

• iManage or NetDocuments is where the file wrapper actually lives. The rule that makes or breaks this stack: anything the AI layer generates writes back to the DMS as a versioned document under the matter number, not as an artifact parked in the vendor’s workspace. A draft claim set that exists only inside a drafting tool is a record you cannot produce, cannot conflict-check, and cannot hand to the next attorney when the associate leaves. Pick between the two on what your firm already runs — on the AI-integration side the gap is thin, and switching a DMS to improve a drafting workflow is the wrong order of operations.

• Intapp owns intake, conflicts, and the wall. Unfiled applications are the most secrecy-sensitive documents a firm holds, and two competing clients’ disclosures inside one practice group is a live conflict that no written policy resolves. Ethical walls scoped at the matter level are what stop an AI search surface from surfacing one client’s unfiled disclosure to a lawyer working the other side of the same technology. Intapp is also where the intake record starts, which is what lets step one of the handoff below happen before the disclosure touches any model.

• Claude for Legal is the general layer the patent tools do not cover. Reading an office action against the references it cites, summarizing a file wrapper an attorney inherited mid-prosecution, drafting the reporting letter to the client, comparing claim language across a family. MCP connectors let it read from the DMS rather than from pasted text, which keeps the matter record intact. It is also the cheapest seat in the stack by an order of magnitude, which is the argument for not buying a second patent-specific tool to do general reading work.

Named handoffs

1. Disclosure received → matter opened → wall set. Invention disclosure arrives, matter opened in Intapp, conflicts cleared and the ethical wall applied before the disclosure reaches any AI surface. Reversing these two steps is the single most common way this stack fails.
2. Matter open → draft generated → written back. Disclosure into the drafting tool, generated spec and claims written back to iManage or NetDocuments as a versioned draft under the matter number, with the model output preserved as-generated so a later reviewer can see what the human changed.
3. Draft reviewed → signed → filed. The practitioner’s review happens against the as-generated version, and the diff between as-generated and as-filed is the reasonable-inquiry artifact.
4. Filing receipt → docket. Application number, filing date, and every subsequent Office communication flow into Anaqua. They do not live in the drafting tool, which has no obligation to still exist in three years.
5. Office action issued → docketed → analyzed → responded. Deadline set in Anaqua, cited references read in Claude for Legal against the claims as amended, response drafted in the patent tool, filed, and the response date closed in Anaqua rather than in someone’s calendar.

The export question nobody budgets for

This is the part that separates a patent AI stack from a general legal AI stack, and it is missing from most firm AI policies.

Under 37 CFR 5.11, a license from the Commissioner for Patents is required before filing abroad or exporting technical data for purposes related to a foreign application, where the invention was made in the United States and no US application has been on file for six months. Then comes the line practitioners consistently miss, which the USPTO restated in the 2024 guidance: a foreign filing license from the USPTO does not authorize exporting subject matter abroad for the preparation of patent applications to be filed in the United States.

Put that next to how AI tools work. Uploading an unfiled, US-made invention disclosure to a model hosted outside the United States is an export question, and the license you would obtain for the PCT filing does not answer it. The Export Administration Regulations add the second edge: releasing technology to a foreign person inside the United States can be a deemed export under 15 CFR 734.13(b), which reaches a vendor’s offshore support engineer with production access as readily as it reaches a foreign server.

Guard: get the inference region, the support-access geography, and the no-training term written into the contract, not asserted in a sales call. Run the vendor through a documented diligence pass — the vendor DD questionnaire skill covers the data-flow questions — and keep the answer in the matter file, because “the vendor said it was US-hosted” is not a record.

The reasonable-inquiry gate

The 2024 guidance ties three existing rules together and adds no new ones, which is exactly why it binds. Under 37 CFR 11.18(a) the practitioner must personally insert their own signature; under 11.18(b) that signature certifies the statements are true to the signer’s knowledge and that a reasonable inquiry was performed; and 37 CFR 11.106(d) requires reasonable efforts to prevent inadvertent or unauthorized disclosure of client information, which is where third-party AI systems enter the analysis.

The operational consequence is narrow and concrete: someone must read the generated specification against the disclosure and the generated claims against the art, correct what is wrong, and be able to show that they did. Guard: make the as-generated version a required artifact in the DMS. A firm that overwrites the model output with the reviewed version has destroyed the only evidence that a review occurred.

What you disclose, and what you do not

You do not disclose that you used AI. There is no such requirement. What you owe is the ordinary duty of candor under 37 CFR 1.56 — and the guidance is explicit that this extends to information about the use of AI tools by inventors, parties, and practitioners where that information is material to patentability.

The place this actually bites is inventorship. If a model contributed to conception rather than to drafting, that is a materiality question, not a tooling question. Guard: the invention disclosure form should capture what the named inventors conceived before the drafting tool touched the file, and that form should be dated and stored in the DMS.

Cost baseline

None of the three drafting vendors publishes list pricing, and that absence is itself the planning problem. A third-party comparison — published by a competing patent tool, so treat it as an anchor rather than a quote — puts DeepIP at roughly $350 to $420 per user per month and estimates Solve Intelligence near $775 per user per month; the same source notes Solve does not publish pricing directly. For a ten-drafter practice that is a drafting layer somewhere in the $42,000 to $93,000 per year range before negotiation.

Anaqua, iManage, NetDocuments, and Intapp are all quote-only, and for a mid-size IP practice the system-of-record and DMS lines together typically exceed the drafting layer. Claude for Legal is the outlier at seat pricing starting near $20 per user per month.

Then the fees the stack does not reduce. The non-DOCX filing surcharge on new utility applications runs up to $430 per application and has since January 17, 2024. The continuing-application surcharges effective January 19, 2025 add $2,700 for a continuation, divisional, or CIP filed more than six years after the earliest benefit date, and $4,000 past nine years. Faster drafting produces more applications; it does not move a benefit date. Model the filing fees before you model the seat savings.

Variations, and when to swap

• In-house IP group rather than a firm. Drop Intapp — the conflicts problem is different when there is one client. Keep the system of record, and expect the AI vendor diligence to run through your own security review rather than a firm’s.
• Solo or two-attorney practice. Skip Anaqua and the enterprise drafting seat entirely. The same third-party comparison lists ClaimMaster around $75 to $90 per user per month and Patent Bots around $25 to $34 per user per month, which is the honest starting point below roughly twenty filings a year.
• EPO-weighted portfolio. DeepIP is the defensible pick after the PatentMaker acquisition, which brought a German practitioner base and European-jurisdiction workflows with it.
• The rule for swapping the drafting layer. Swap when drafters stop opening it, not when a competitor ships a demo. Seat utilization is the only honest signal here, and every one of these vendors will show you a productivity figure derived from their own customers.

What this stack does not replace

• The signature and the review behind it. Nothing in this stack performs a reasonable inquiry. It only makes the evidence of one easier to keep.
• Docketing discipline. Anaqua automates data extraction from office documents; it does not absorb the consequence of a missed response date. A second human check on entered dates stays in the process.
• A prior art search you can defend. AI-assisted search changes recall, not the duty to disclose what you found.
• Inventorship determination. That is a legal judgment about conception, and it is now a materiality question too.
• Claim strategy. The tools draft claims well and choose scope badly, because scope is a business decision about a market you have to understand.

Match rules

Right pick when you file enough to amortize a seat — roughly fifty or more applications a year across the group — you already run a docketing system you trust, and you can obtain contractual answers on inference region and training. In that configuration the AI layer compounds against infrastructure that already works.

Wrong pick when any of those three are missing. Below roughly twenty filings a year the seat cost per application dominates any time saved. Without a docketing system, buy that first — an AI drafting tool bolted to a spreadsheet docket increases filing volume against a deadline process that was already the weakest link. And if a vendor will not put the inference region in the contract, the export analysis has no answer, which makes the tool unusable for US-made unfiled disclosures regardless of how well it drafts.