Where AI Patent Drafting ROI Actually Comes From

The ROI model in the vendor deck counted the hours saved producing a first draft. That is the cheap side of your ledger, and it is not where the money in this decision sits. AI patent drafting ROI is decided almost entirely by verification cost: the time your most expensive people spend establishing that a draft is sound before anyone signs it. Production time is real but small, and it is billed to the wrong person. This article sets out the three cost pools that actually move, explains why fluent output can raise your costs rather than lower them, and gives you a way to build a model you can defend to finance. One thing up front: this article contains no benchmark percentages, no industry averages, and no "typical firms report" figures. Every number in this category is a vendor's number with a baseline you cannot inspect, and putting one into your business case is how the business case gets returned. The model below runs on your numbers.

Three cost pools, only one of which gets measured

Production. The hours between a complete disclosure and a draft that exists. This is what demonstrations measure and what most ROI models are built on. It is the smallest of the three and it is spent by whoever drafts: often the least senior person touching the matter, or an outside firm on a flat fee where the hours are not yours anyway. Verification. The hours between a generated patent draft and the time for someone to review the draft. Support checked against the disclosure. Antecedent basis. Vocabulary drift between claims and specification. Whether the claim set actually accounts for the closest art. This pool is spent by senior counsel, it is the largest of the three in most departments, and almost nobody models it. Downstream prosecution. What a weak draft costs you in years two through five: avoidable office actions, amendments that narrow scope because the specification did not carry a fallback, and, worst and least visible, claims you cannot get because the support was never in the priority document. A tool that halves the first pool and doubles the second has made your department slower and more expensive while showing you an excellent efficiency number.

The verification trap

Here is the mechanism that makes this counterintuitive, and it is the single most important thing to understand before signing anything in this category. Fluent prose is harder to check than clumsy prose. When a junior drafter produces a limitation with no support in the disclosure, it usually reads wrong. The phrasing is awkward, the term appears from nowhere, the sentence sits oddly against its neighbours. Your reviewer's eye catches it without looking for it. When a language model produces the same unsupported limitation, it reads beautifully. It uses your vocabulary, it sits in the correct paragraph, it is grammatically indistinguishable from the limitations on either side of it that are fully supported. Nothing in the text signals which is which. So the reviewer cannot skim. They have to go back to the disclosure and check every element, one at a time, because fluency has removed the surface cues they have relied on for twenty years. That is why some teams adopt an AI drafting tool, report enormous production savings, and find that their filing throughput has not changed. The work moved. It did not disappear, and it moved from the drafter to the reviewer, which is the wrong direction. The explainability dividend is the answer to exactly this, and it is an economic mechanism rather than a compliance feature. If every claim element carries a pointer back to the passage of the disclosure it came from, verification stops being a reading exercise and becomes a checking exercise. That is the difference between hours and minutes, applied to the most expensive time in the department.

Why per-hour models mislead an in-house team

Law firm ROI arguments run on realisation and billable capacity. Yours cannot, because you do not bill. Your currency is attorney attention and outside counsel spend, and your unit is the matter. So the model that works is cost per matter delivered, fully loaded: internal attorney hours at your loaded rate, plus outside counsel invoices, plus drawing costs, plus search costs, plus the prosecution queue. Then the question becomes answerable. Does this change my cost per matter, and does it change how many matters I can hold inside without adding headcount? That framing also explains why the commercial structure matters. The market norm in this category is a per-seat annual licence, which prices your headcount. Esgenix is priced per matter delivered, across tiers, so cost tracks filings. If your filing volume and your attorney count move independently, and in most corporate departments they do, those two models diverge sharply over three years, and only one of them is denominated in the unit your own model uses. It also aligns the incentive: per-seat revenue arrives whether or not the matter completes.

How Esgenix is built around the second cost pool

Esgenix is an agentic patent workflow platform. 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: one searches the prior art, one drafts the claims, one writes the description, one produces the figures, one grades the result. Almost every competing product is a copilot: it sits beside the attorney while the attorney writes, suggesting text section by section. Esgenix runs the drafting work itself and stops at five gates for the attorney's decision: intake, where you confirm the invention, answer the technical gap flags and set jurisdictions; strategy, where you set claim breadth and choose fallbacks; claims, where you approve or amend the set, and nothing proceeds without this; specification and figures, where you review with provenance visible on every element; and sign-off, a natural person's signature that cannot be delegated. Human involvement is not reduced. It moves from paragraph assembly to decisions worth an attorney's time, which is also the move that shifts hours out of the expensive pool. Four mechanisms map onto the three cost pools directly. Provenance, against verification. Every claim element carries an identifier tracing it to the passage of the invention disclosure it came from. Support is confirmed by following a pointer rather than by reading two documents in parallel. This is the mechanism the whole ROI argument rests on, and it is the one to time in the room: pick five limitations and time the trace. The judge model, against verification. 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 machine-written prose falls into and the substantive defects underneath it: before the draft reaches the person whose time costs the most. Search feeding the draft, against downstream prosecution. The prior art search covers 220 million documents from 108 jurisdictions and matches on meaning rather than keywords. What matters economically 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, and the specification carries fallback positions where the art is dense. Fallbacks that exist at filing are amendments you can make later. Fallbacks that do not exist are scope you lose. Gap detection, against all three. At intake the system identifies what the disclosure is missing (an unexplained step, an undefined term, a result asserted with no mechanism) and proposes how to close it. The inventor email that normally arrives in week three arrives in hour one, which is the cheapest point in the matter to fix anything. 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 that removes a line item and the reconciliation passes after it. Drafting profiles store a house style and enforce it across every agent, so style survives volume.

One table: build the model on this

Cost pool What vendors measure What you should measure How to get the number Production Time to first draft Instructions needed from a qualified person Count every prompt and accept/reject decision in a live run Verification Nothing Minutes for a signing attorney to reach filing-ready Have that attorney do it once, timed, on a real disclosure Verification Nothing Time to trace support for five limitations Pick five, time the trace, per product Downstream "Fewer office actions" Whether the claim set accounts for art you withheld Hold back one close reference you already know about Commercials Per-seat licence cost Cost per matter delivered, fully loaded Your loaded rate × hours, plus outside counsel, drawings, search Run this on three real matters. Three is enough to see differences that matter in this category, and it is small enough that you will actually do it.

What not to put in your business case

Do not claim a reduction in office actions you have not measured. It is probably true, the mechanism is sound, and you cannot evidence it inside a budget cycle. Claim the verification saving, which you can time in an afternoon, and let the prosecution saving arrive unclaimed. Do not present an audit trail as compliance protection. There is no general duty to disclose that AI assisted with drafting, and the duty of disclosure concerns information material to patentability rather than the tools used to prepare a document. Any vendor selling provenance as protection against inequitable conduct or as a regulatory requirement is describing an obligation that does not exist, and the claim invites a conversation you do not want in front of your general counsel. Provenance is worth buying because it converts reading time into checking time. That is a sufficient reason and it survives scrutiny. And do not model headcount reduction. The attorney remains responsible: a natural person signs, and that signature is a personal certification about the contents. What changes is what your attorneys spend the day doing, not how many of them you need.

A specific next step

Build the model with three numbers you already have: your loaded attorney rate, last year's outside counsel spend split by reason, and the number of matters you filed. Then get the fourth number, which is the only one you are missing: have the attorney who signs take one AI-drafted application to filing-ready and time it. Bring three real disclosures to a working session with us and we will run them through the five gates while you count instructions and time the support traces. You will leave with the verification number, which is the one your model actually turns on.