AI Can Draft the Application. It Can't Carry the Portfolio.

A “patent application” generated by AI in a few hours — properly formatted, figures and all — goes across the boss's desk, and by the next investor meeting it has become “we're ready to file.” Over the past two years this has stopped being unusual.

The problem isn't whether the paperwork looks complete. It's that the person holding it often can't say what that document actually protects.

AI genuinely helps. It can organize an invention disclosure, run a first-pass search, generate drawings, and produce something that reads like a real application — fast. But fast and enforceable are different properties. What an investor wants to know has never been “did you file”; it is “will your position survive a challenge.”

What follows takes a five-stage framework recently set out by U.S. practitioners and re-grounds it in what Taiwanese R&D companies actually run into, comparing the Taiwanese and U.S. rules point by point. If you're mid-R&D, preparing to raise, or negotiating a technology collaboration, you should finish this able to tell which checkpoint your company is stuck at.

Checkpoint 1 — Control Disclosure: What You Think Is Confidential May Already Be Public

Patent work starts with control, not with drafting.

The first question is simple: who knows about this invention right now? The second matters more: under what duty of confidentiality did they learn it?

Talking to a patent attorney, chatting with a prospective partner, an investor, a customer, a contract manufacturer or a university lab, and pasting the idea into a public AI tool are not legally equivalent acts. The first normally sits inside a professional confidentiality framework. The rest depend on whether there is a contract and on trade custom — and if the luck runs out, you discover after the fact that nothing was protected at all.

Taiwan's Patent Act Article 22(3) gives inventions and utility models a 12-month novelty grace period; 35 U.S.C. § 102(b)(1) also runs 12 months. The symmetry is superficial. Grace period conditions in Europe, China and Japan are generally stricter than Taiwan's, and each jurisdiction recognizes only its own. Successfully invoking the grace period in Taiwan does not mean Europe or China will treat the same disclosure as non-novelty-defeating.

For companies that deal constantly with overseas customers and contract manufacturers, or that exhibit at international trade shows, the practical rule is the old one: file first, disclose second. A Taiwanese grace period is not a worldwide insurance policy.

Feeding invention content into a consumer AI tool is itself a disclosure. Whose servers hold that conversation, whether it will be used to train a model, whether those servers sit offshore — all of that belongs on the checklist before anyone presses Enter. Enterprise tools with explicit no-retention, no-training commitments are a different matter. But has anyone at the company actually read those terms?

Ownership gets settled at this checkpoint too. Article 7 assigns the right to apply and the patent right in service inventions to the employer. Article 8 gives the employee a procedure: after the employee asserts in writing that an invention is a non-service invention, if the employer does not object within six months, ownership is fixed and cannot be reopened later. Read the other way, a company without a clear service-invention determination process and a written record-keeping mechanism can be cut by its own rule — six months of silence and it's decided. Ownership for every inventor, consultant and intern should be investigated before filing.

Practice rule: establish who knows what, and under what duty, before deciding what goes out the door. A Taiwanese grace period is not a global grace period.

Checkpoint 2 — AI-Assisted Invention Disclosure: AI Can Organize, but It Cannot Be Named as Inventor

AI is well suited to helping inventors and commercial teams turn scattered thinking into a usable disclosure document. The point of that exercise is to get the patent attorney up to speed faster — not to substitute for the attorney's judgment.

A solid disclosure answers a specific set of questions. What problem does the invention solve? Where is it better than existing approaches? What has actually been built (as opposed to planned)? What are the key components or steps? What alternatives and fallback designs exist? What data or drawings support it? Which approaches were tried and failed? Who conceived each technical feature? And where, exactly, did AI enter the process?

For hardware and process inventions, add two more: what the currently commercialized version looks like, and where it might move in the next 6 to 12 months. Companies pivot and products get revised. A disclosure written too narrowly leaves no room to adjust later.

So who is the inventor when AI helped? Taiwan already has a clear answer. In Case No. 110-Xing-Zhuan-Su-3, the Intellectual Property and Commercial Court dismissed an application naming the AI system “DABUS” as inventor, holding that an inventor under the Patent Act must be a natural person; AI is legally an object, not a person — the subject matter of rights, not a holder of them — and the Taiwan Intellectual Property Office (TIPO)'s refusal to accept the filing was lawful. This aligns with the Federal Circuit's position in Thaler v. Vidal (2022): AI is a tool, no different in character from lab equipment or modeling software, and the person who actually conceived the technical solution is the inventor.

Using AI does not cost a human contributor inventorship. But keep the record, and make it specific about who contributed what. Inventorship errors can be corrected after the fact — but investors and litigation opponents are both very good at digging into that particular backstory.

One duty does not get lighter because AI was involved: the duty of candor. If an AI-assisted search surfaces prior art, literature or competitor material that could bear on patentability, it goes to the patent attorney for assessment — not into the mental file marked “the outcome favors us anyway, let's not mention it.” The purpose of a disclosure document is not to make the invention look better than it is; it is to make the record complete enough to survive examination and, later, litigation.

Practice rule: use AI to organize facts and find gaps. Don't use it to bury unfavorable ones, and don't let it stand in for the attorney's judgment.

Checkpoint 3 — Patentability Assessment: Can We File and Should We File Are Two Different Questions

Once the disclosure is assembled, the next question is not “can we file” — anyone can submit an application. The real questions are: is it worth filing? Is the technology mature enough to support an application? Is a patent even the right instrument for this commercial problem?

A serious assessment starts with a broad search, not a handful of Chinese and English keywords that happened to come to mind. Published applications from Taiwan, the U.S., Europe, China, Japan and Korea, plus PCT publications, all belong in scope. WIPO's 2025 statistics show China, the U.S. and Japan as the three largest sources of patent filings globally — those three deserve priority in any search.

The statutory tests run on similar logic in Taiwan and the U.S., with different vocabulary: novelty, inventive step (the counterpart of U.S. non-obviousness), industrial applicability — and the threshold R&D companies most consistently underestimate, patent eligible subject matter.

Do the assessment with discipline and you can hold a real conversation with investors, partners and the board. Not “we've filed,” but: here is the scope we're pursuing, here is the prior art we searched, here are the rejections we expect. No patent attorney can guarantee an examination outcome, but an experienced one can usually predict in advance which wall a case will hit — novelty, inventive step, eligibility, or insufficient disclosure.

Practice rule: ask whether it's worth filing before asking whether it can be filed. For medical and diagnostic inventions, confirm eligibility early — don't wait for the office action to deal with it.

Checkpoint 4 — Professional Drafting: Taiwan Has No Provisional, but It Has Domestic Priority

This is the checkpoint where saving money most reliably turns into spending money.

Using AI to help assemble an invention disclosure is fine. Using AI to write a complete application and filing it as-is is a different order of risk.

The U.S. provisional application has a widely misunderstood property: it is never substantively examined. Precisely because nobody checks on filing day whether the content is adequate, a thin provisional can sit quietly in the file for a year while everyone assumes the company is protected. The problems all surface later — at conversion to a non-provisional, when claims broader than the original disclosure are asserted, when an investor asks whether the case actually covers the shipping product, or when a competitor attacks the priority basis.

In New Railhead v. Vermeer (Fed. Cir. 2002), a provisional omitted a technical feature later claimed; the feature lost the early filing date, and the patent was ultimately invalidated over the company's own prior sales activity.

Taiwan has no provisional in the American sense, but it has a functionally adjacent tool: domestic priority under Article 30. A company can file a relatively streamlined basic application, then within 12 months claim domestic priority in a later application — using that window to add experimental data and adjust claim scope while preserving the original filing date.

The difference is in whether the tool is available to you at all, and the precondition is as strict as it is in the U.S.: the features the later application claims must be supported by content that the earlier application disclosed in a “clear and sufficient” manner under Article 26. What the earlier case didn't say, the later case doesn't get the early date for. The New Railhead lesson transfers intact into Taiwan's domestic priority framework — a different set of provisions saying the same thing.

A well-drafted application describes alternatives, includes meaningful drawings, preserves fallback positions, avoids unnecessary prior art admissions, writes in the version the company actually sells, and leaves room for where the business is going. Investors know companies pivot — products get revised, data updates, models iterate, processes get optimized. No application predicts every turn. But a solidly drafted one holds far more ground than a document an AI generated from a fundraising deck.

Related point: choosing a patent attorney is not simply a procurement decision. The application filed today may serve as the baseline document for continuation filings, overseas portfolio strategy, licensing negotiations, pivot-driven adjustments and enforcement decisions for the next twenty years.

Firm size isn't the variable. The variable is whether this attorney understands the technology, understands the company's commercial objectives, knows what examination is going to throw at it, and knows what investors will ask during diligence.

Practice rule: domestic priority is Taiwan's buffer mechanism, but whether it holds depends on how completely the first application was written. A cheap application that discloses your position and then fails to support the scope you need later is the expensive kind.

Checkpoint 5 — The Patent Roadmap: Filing Is the Starting Line, Not the Finish

Filing is not the end of the exercise; it is the beginning of a management period. Invention patents run 20 years from the filing date.

Whether it's the domestic priority year or the Paris Convention priority year, treat it as a build period, not a waiting period. During that window the company keeps improving products, updating models, accumulating data, optimizing processes. If those improvements aren't caught by follow-on filings, what the original application protects may no longer be what the company actually does.

Overseas decisions also need to come early. Paris Convention priority lets a company claim the same priority in other countries within 12 months of the filing date — and this 12 months is a different thing from the grace period discussed earlier. The two should never be conflated. Via the PCT route, you instead secure a single international application and have 30 months from the priority date to decide which national phases to enter, pushing the market-selection decision back. But the PCT does not produce a “world patent.”

Annuities escalate year over year, deliberately: the design pushes commercially worthless patents toward early abandonment and back into the public domain. Taiwan and the U.S. share that logic with different rhythms — Taiwan collects annually, the U.S. takes larger maintenance fees at specified intervals. Same point either way: work out early which technologies are worth feeding long-term.

An office action is part of normal prosecution, not a signal of failure; every ground of rejection needs a corresponding argument or amendment. After allowance, the divisional application is Taiwan's standard tool — functionally close to the U.S. continuation — letting a company preserve the option to claim other technical features before the parent issues. How precisely that timing is judged often determines how much flexibility the portfolio has for years afterward.

Practice rule: manage the post-filing period as a build plan with a schedule, not as a file-and-wait.

A Checklist for Founders and R&D Leads

Before the next investor meeting — or before deciding whether a given technology goes into the fundraising deck — five questions are worth asking:

Who has this technology been discussed with, and under what duty of confidentiality? Does the disclosure document state clearly who conceived each feature and where AI entered the process? Does the search cover the major markets outside Taiwan? Do the pending applications actually cover the product the company sells today, or are they still sitting on the earliest concept? And since filing, is anyone tracking the build-period actions that are supposed to be happening?

AI hasn't made patent fundamentals obsolete. It has made them matter more. AI can help a company assemble a more complete disclosure, run a broader search, and make the first meeting with a patent attorney far more productive. All of that is good. But the same tools can make a thin application look substantial, and let a company believe it is protected before the work has actually been done.

Our approach with technology-manufacturing clients is straightforward: AI is for preparation, not for appearances. The companies that actually reassure investors are the ones that control disclosure, do the disclosure and assessment work properly, hand drafting to an attorney who understands both the technology and the business, and manage the whole patent term as a long-range plan.

This article is general commentary, not legal advice for any specific matter; actual filing strategy and portfolio design should be assessed case by case against the underlying technology and commercial goals.