AI Patent Practice Transformation: What Actually Changes for an In-House Team

Your filing target went up again this year and your headcount did not. You have three levers (work later, send more out, or file less) and you have already pulled all three as far as they go. Most writing about AI patent practice transformation is aimed at law firms and argues about billing models. That argument does not reach you, because you do not bill. For a corporate IP department the transformation is narrower and more useful: AI moves the line between the work you keep inside and the work you send out. Everything else follows from that. This article sets out where the line currently sits, what actually moves it, and what does not move it despite the marketing.

The constraint is headcount, not hours

A law firm's ceiling is billable capacity. Yours is a headcount plan approved by someone who does not distinguish between a continuation and a first filing, and a filing target set by an R&D organization that has no reason to slow down. Between those two numbers sits the outside counsel budget, and it absorbs the difference. That is what it is for. But every matter that goes out for capacity reasons rather than expertise reasons costs you three things beyond the invoice. It costs consistency, because a firm drafts to its own house style rather than yours. It costs institutional knowledge, because the reasoning behind a claim scope decision ends up in someone else's file. And it costs your attorneys the work they are best at, since the matters that go out are rarely the interesting ones. They are whatever was on the desk when the deadline arrived. So the useful question about any AI product is narrow: does it change how many matters you can hold inside without adding people? If it does not, it is a productivity tool, and productivity tools do not change a headcount plan.

Why assistance-shaped tools move the line less than expected

Patent AI is moving from assistance toward workflow execution, and products differ mainly in where they draw the line between machine and professional. It sits beside the attorney while the attorney drafts and proposes the next passage, the next limitation, the next amendment. The attorney remains the drafter; the software makes drafting faster. That helps. It also has a hard ceiling, and the ceiling is structural rather than technical. If the attorney must be present for every instruction, your capacity is still bounded by attorney presence. You have made each paragraph faster to produce without changing how many applications one attorney can be responsible for in a quarter. Ten per cent faster paragraph assembly does not become a headcount argument, and it does not bring a single matter back in from outside counsel. The number worth measuring is not minutes per page. It is how many separate instructions one complete application requires from a qualified person. That figure is what caps your throughput, and no amount of faster typing changes it.

What moves the line

Esgenix is an agentic patent workflow platform. It produces filing-ready specifications from an invention disclosure across US, European and Indian practice. An agent here means a piece of software given a defined job, the tools to do it, and a standard to meet: it completes the job and reports back rather than waiting for the next instruction. Several run in sequence, on different underlying models chosen for what each does well: one searches the prior art, one drafts the claims, one writes the description, one produces the figures, one grades the result. The difference from an assistance-shaped tool is not that Esgenix is faster. It is that the production happens between the attorney's decisions rather than beside them. Where a tool is shaped around assistance, the professional stays present through the drafting itself. With Esgenix, the attorney makes five decisions and the platform does the production between them. The right analogy is delegation to a senior associate with agreed check-ins. You set the scope, you review at the points that matter, and you sign. That is also, not coincidentally, the shape of the work you currently send outside, which is why this is the thing that moves the make-versus-buy line. The five gates are these. Intake. You confirm the invention as the platform understood it, answer the technical gap flags, and set jurisdictions. Gap detection means the system reads the disclosure and identifies what is missing (an unexplained step, an undefined term, a result asserted with no mechanism) and proposes how to close each one. The inventor email that normally arrives in week three is sent in hour one. Strategy. You set claim breadth on the scope slider, an on-screen control that widens or narrows the claim set as you drag it, with support re-checked against the disclosure at each position. Scope becomes something you explore and see the consequences of, rather than something you commit to and then test by hand. Claims. You approve or amend the claim set. Nothing proceeds without this. Specification and figures. You review the draft with provenance visible on every claim element. Provenance means each element carries an identifier tracing it to the passage of the invention disclosure it came from, so support is verified by following a pointer rather than reading two documents in parallel. Sign-off. A natural person has to approve of the draft and only then it is downloaded as Docx and PPtx. Human involvement is not reduced. It is relocated: off paragraph assembly and support-checking, onto the decisions that determine what the patent is worth.

One table: what happens when filings rise forty per cent

Lever What it costs What it does to quality How fast it works Add headcount New requisition, recruitment, ramp Improves, eventually Two to four quarters Send more to outside counsel Budget line grows with volume House style and reasoning move outside Immediate Assistance tooling Per-seat license per attorney Unchanged; attorney still drafts Immediate, bounded by attorney presence Agentic drafting with gates Per matter delivered House style enforced; provenance on every element Immediate, bounded by decision capacity The bottom row's commercial structure follows from the same logic. The market norm is a per-seat annual license, which prices your headcount. Esgenix is priced per matter delivered, across tiers, so cost tracks filings. In a department where filing volume and attorney count move independently (which is to say, in most departments) those two models diverge sharply over three years, and only one of them is measuring the thing you are actually buying. It also aligns the incentive: a per-seat vendor is paid whether or not the matter completes.

What depth in drafting buys you specifically

The prior art search covers 220 million documents from 108 jurisdictions and matches on meaning rather than keywords. Semantic search means the system compares what a passage says against what your invention does, so it surfaces the reference that never uses your vocabulary. What matters operationally is that the search output feeds the drafting agents in the same session: the claim set is positioned against the closest art from the first draft, the specification carries fallback positions where the art is dense, and the figures illustrate the distinguishing features. That is one vendor, one session, and one invoice instead of a search report that arrives separately and gets carried into drafting by hand. Freedom-to-operate and white space analysis run inside the same session for the same reason. Figures are generated with no baseline drawing required. The system reads the specification, decides what the figure set should be, draws it, and reconciles every reference numeral against the text. For most departments this removes a separate draughtsman line item and the reconciliation pass that follows it. Drafting profiles store a house style and enforce it across every agent. This is the answer to the consistency problem above: your style survives volume, and it survives the matter going through a workflow rather than through a particular person. An independent model grades the finished draft against the standards a senior patent attorney would apply and returns it to be rewritten when it fails. This is LLM as a judge: a quality gate. It catches the flat, generic texture that machine-written prose falls into and the substantive defects underneath it, before the draft reaches your reviewer. Esgenix has completed ISO 27001 Stage 1, covering the platform and the handling of invention and client data.

What does not change

Three things, and they should be said plainly because the category oversells them. The attorney remains responsible. Papers filed at the USPTO carry a natural person's signature, and that signature is a personal certification about the contents. No workflow touches that. Inventorship is unaffected. USPTO guidance is clear that an AI system cannot be named as an inventor and that a natural person must have made a significant contribution to conception. Drafting assistance does not bear on this; conception does. And there is no general duty to disclose that AI assisted with drafting. The duty of disclosure concerns information material to patentability, prior art, not the tools used to prepare a document. Any vendor selling you an audit trail as malpractice insurance or as a regulatory requirement is selling you something that does not exist. Provenance is worth having because it lets you verify support in minutes instead of hours, which is a practical benefit, not a legal shield.

A specific next step

Before evaluating any platform, run a make-versus-buy audit on last year's filings. For each matter that went to outside counsel, write down the reason in one word: expertise, jurisdiction, litigation-adjacent, or capacity. Then total the capacity column. That number is the only part of your budget an agentic drafting workflow can address, and it tells you immediately whether this is a strategic decision or a marginal one. If it is large, bring three of those matters (the actual disclosures, not a sample) to a working session and run them through the five gates. Count the instructions each one needed from a qualified person, and compare the output against what came back from the firm.