Choose patent AI by whether it is a workspace with evidence, jurisdiction handling, matter records, and human final review—not by who writes the prettiest paragraph. Fluent drafts are available almost anywhere. What splits practice tools is whether that draft is anchored to source passages, stays inside one matter, and is signed off by a person.
Humans own claim scope and response strategy. The criteria below are a pre-purchase checklist. They do not promise patentability or allowance.
Series map: Patent AI practice guides (I0).
Why a general ChatGPT or Claude chat is not enough
General LLMs are useful for draft sketches, summaries, and translation outlines. In patent matters, four practice risks keep showing up:
- Lost matter context — claim versions, cited art, and prior instructions scatter when the chat changes.
- Weak cite-to-passage tracing — you get a “similar document” summary without cell-level source highlights.
- No version or audit trail — who locked or changed key wording is hard to reproduce.
- Wrong office format — converting into US or PCT form can quietly drop limitations, numbers, and figure refs.
This is not an argument against using AI. The question is whether the product is built around those four risks.
The market also splits along a real buying axis: Word-centric plugins versus browser matter workspaces. Both can be legitimate; they are not the same job. For shortlisting, four buckets are enough—named once as examples, not as a ranking.
| Bucket | Typical job | Example (once) |
|---|---|---|
| Browser drafting / prosecution workspace | Interactive draft + OA workflows | Solve Intelligence |
| Patent intelligence + drafting | Claim charts, invalidity/infringement, plus drafting | Patlytics |
| Word-centric drafting / prosecution | Attorney stays in Word; draft + prosecute | DeepIP |
| Established search / IP platform | Search, families, portfolio | Questel |
| Matter workspace | Prior-art compare + spec + review + audit + jurisdiction convert in one matter | Patenty |
The seven criteria below do not crown a winner. Any vendor can be run through the same 30-minute protocol.
Criterion 1: Can you map invention elements to prior art with source passages?
Why it matters. Prior-art judgment is not “is this document similar?” It is which element maps where. A one-line similarity score is not practice evidence.
Check for:
- A matrix (or equivalent chart) with rows = invention elements and columns = references
- Opening a cell lands on a clickable source passage
- Partial mappings can record what corresponds and what does not
- AI suggestions are distinguishable from attorney confirmation
“AI says similar” alone is close to a fail. The matrix itself is unpacked in element-by-element prior art (I2) (upcoming).
Criterion 2: Do claims → specification stay in the same matter?
Why it matters. Drafting a claim in chat A and pasting the specification into chat B is how element IDs, defined terms, and dependencies drift.
Check for:
- Claims and the written description live in the same matter
- Element definitions reuse into the spec and the matrix
- You can follow versions without copy/paste as the default workflow
If paste-between-chats is the normal path, you effectively bought two systems.
Criterion 3: Do reviews leave a diff?
Why it matters. Firms and in-house teams eventually ask who finalized a phrase. A final file with no history breaks the audit.
Check for:
- Proposed edits versus accepted text as a diff
- Ability to lock key wording (or an equivalent guard)
- Contributors, versions, and an exportable audit package (export centers on spec/claim/review decisions—confirm package scope in demo)
Fluent rewrite is less important than reproducible change history.
Criterion 4: Does prosecution (office actions) share that context?
Why it matters. If filing drafts and office-action work live in different tools or chats, cited passages and claim history must be rebuilt from scratch.
Check for:
- OA identifiers, claims under rejection, and cited art attach to the same matter
- The pre-filing element matrix can be reused in prosecution
- Response drafts cannot silently target an outdated claim set
Before/after checks live in the office-action response checklist (I3) (upcoming). Outsourcing response strategy to the model is not the premise.
Criterion 5: Can you catch omissions in jurisdiction convert / translation?
Why it matters. Foreign-filing quality is not “does the English read well?” It is whether limiting language, numbers, dependencies, and figure references still line up 1:1 with the source.
Check for:
- Side-by-side source ↔ target review after US/PCT (or other) conversion
- Flags for omission types—e.g. a voltage, temperature, or ratio range that survives only in part
- Figure refs and element IDs still pointing at the same rows after translation
Do not invent “X% of conversions drop a number.” Omission types and a short review protocol are covered in US/PCT translation omissions (I4) (upcoming).
Criterion 6: Are data, training, and retention policies documented?
Why it matters. Specifications, prior art, and client materials are sensitive. Marketing copy matters less than what the policy documents say.
Ask the vendor (do not assume; treat this as a question list):
- Is customer material used for model training—yes or no—stated in FAQ or policy?
- Can you obtain subprocessors, retention, and deletion procedures in writing?
- Is there a security / compliance page (or equivalent pack) you can link—e.g. trust?
This article does not assert any vendor’s security posture or training practice. It only fixes what to verify before matter data goes in. For Patenty specifically, confirm the live site FAQ and policy pages at evaluation time; do not rely on second-hand summaries.
Criterion 7: Is attorney final review a product premise?
Why it matters. Messaging that implies auto-file, automatic patentability calls, or submit-without-review is a practice and ethics red flag.
Check for:
- UI and copy that assume human final review
- No guarantees of allowance, grant, or invalidity outcome
- Clear separation between AI output and the confirmed work product
People own the final call on claim scope and response strategy. Tools only accelerate that judgment.
30-minute evaluation protocol
On any demo or trial, run one sample invention through the steps below. The protocol is brand-agnostic.
| Step | Do this | Pass signal |
|---|---|---|
| (a) | Split the invention into ≤5 elements | Elements persist inside the matter |
| (b) | Open 1–2 prior-art passages from a cell | Cell → source passage opens |
| (c) | One independent claim → spec draft | Continues in the same matter |
| (d) | Plant one intentional error; fix it; inspect the diff | Change + contributor remain |
| (e) | Open training / security policy pages | You leave with document links and wording |
Items that fail are not “coming later in the UI.” Treat them as holes in today’s workflow.
How this looks in practice (Patenty)
Here is only how the seven criteria can be checked inside a Patenty matter workspace. Competing brand names are not re-listed.
- Criterion 1 — Element comparison matrix with per-cell source passages
- Criterion 2 — Claims, specification, and review continue in the same matter
- Criterion 3 — Review diffs, key-phrase lock, version compare/restore (audit export centers on spec/claim/review decisions—confirm package scope in demo).
- Criterion 4 — Filing context carried into OA / prosecution in the same matter
- Criterion 5 — Jurisdiction convert with side-by-side omission review
- Criterion 6 — Documented handling per the live site FAQ and policies (verify current trust and related pages at evaluation; do not rely on this article for the claim)
- Criterion 7 — Professional final review required; no auto-grant or patentability guarantee
Screen evidence is fastest in a shared demo. The goal is not a feature dump; it is whether protocol steps (a)–(e) can be reproduced in one matter.
Soft CTA
If your team is evaluating vendors, send the 30-minute protocol as written. To see the same judgment recorded in a matter workspace, request a product / plans demo. Filing and claim scope remain the attorney’s responsibility.