An IP firm can use AI to read incoming correspondence for docket entries, screen search results and translate foreign documents, and produce first drafts of specifications, responses and letters, through specialist tools or a business-plan assistant. A lawyer must still verify every deadline, read every cited reference and own every filing.
That is the pattern in all three jobs. AI is good at reading, summarising and drafting. It is unreliable at calendar rules, at knowing what has been published and at judging its own output. Specialist tools such as Alt Legal, DeepIP and Solve Intelligence sell the first strength, and the lawyer's checks cover the rest.
Two kinds of software, and why IP work needs both
A docketing system is a rules engine. It encodes each office's periods, holidays and extension rules, computes dates exactly and knows nothing about what a letter means. A language model works the other way round: it understands what a letter is about, can summarise it in plain words, and cannot be trusted with a calendar. IP practice needs the meaning and the date, so the dangerous mistake is asking one tool to do the other's job.
The same split explains AI search tools. They usually combine retrieval, which finds documents that look similar to a query, with a language model that summarises what was found. Errors can enter at either stage: the retrieval can miss the one document that matters, and the summary can misstate a teaching in a document it did retrieve. Keep those two failure points in mind and most of the checks below follow naturally.
Docketing: let AI read the post, never set the date
The useful job is extraction. A clerk currently opens each incoming letter, finds the document type, the number, the dates and the response period, and types them into the docket. AI can read the same letter and propose those fields, with the sentence each came from. The clerk then checks the proposal against the letter, the system applies its own rule to calculate the date, and a second person verifies the entry.
Alt Legal, which sells trade mark docketing across 180-plus jurisdictions and prices it by docket size (from $60 a month for up to 50 matters), describes its AI Filtering Assistant as a way to "quickly and easily search your docket", while its rules engine calculates the due dates. That is a sensible role for AI: finding and explaining, not calculating. Whatever you buy, ask the vendor which of those two jobs the AI is doing, and which rules the date calculation uses.
A prompt that keeps the tool in its lane:
You are helping a docketing clerk. Read the letter below
and return a table with these fields: document type; date
printed on the letter; application or registration number
as printed; response period exactly as the letter words
it; action required; any deadline the letter states as an
explicit date; any extension the letter mentions.
Do NOT calculate due dates. If a field is not printed,
write NOT STATED. For each field quote the sentence it
came from.On a hypothetical examination report dated 14 September 2026, a good result lists the period as "within four months of the date of this report", gives the explicit deadline as NOT STATED, and quotes a sentence saying the period "may be extended on request and payment of a fee". The clerk then enters the due date using the system's rule for that office and document type, and the verifier ticks it off. Two failures tend to appear. The tool may add a sentence of its own such as "response due 14 January 2027" despite the instruction, which is a date calculated from the printed date. That is wrong whenever the office counts from posting or receipt. Or it may miss an extension clause that sits in a footnote. Both are why the verification step exists.
Fixed dates make good sanity checks
Some deadlines are standard enough to act as a tripwire on anything the AI or a clerk enters. WIPO's summary of the Paris Convention sets priority periods of 12 months for patents and utility models and 6 months for industrial designs and marks. WIPO's PCT FAQ says that, for most contracting states' offices, national-phase steps must be taken before the end of the 30th month from the priority date. A docket entry that disagrees with these for a matter at that stage deserves a second look, not because the standard is always right, but because a disagreement is worth explaining.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowA pilot on 40 old letters
Run a test before trusting extraction on live matters. Say a six-person IP practice takes 40 letters it has already docketed, across five office types, and runs them through the prompt above. The clerk compares each output field by field with the entry that was made at the time. In this illustration 37 letters match completely, two miss an extension note and one misreads a date written in a different format. That is 92.5 per cent of letters with no errors, and the three failures are the kind that cost a client rights if nobody checks.
The time maths looks like this. If the practice receives 60 documents a week and manual docketing takes six minutes each, that is six hours a week. With extraction and checking at 3.5 minutes each, it is 3.5 hours, a saving of 2.5 hours a week or about 130 a year. Notice that the checking minutes are inside the 3.5. A version of the process in which the clerk stops checking would look faster on paper and is the one that ends up in a negligence file. The companion routine, with the human check built in, is in law firm docketing with AI: deadline reminders plus a human check.
Add one report to the process: every week, list the documents received and the docket entries created, and look for letters with no entry. That reconciliation catches the failure no extraction prompt can, the letter nobody read.
Searches: faster screening, the same blind spots
AI changes how fast you can screen a result set, not what a search can know. Four uses are worth having:
- Translating foreign documents. WIPO's PATENTSCOPE site, which covers published international applications and documents from participating national and regional offices, lists a separate WIPO Translate service for patent documents. Machine translation of a cited reference is good enough to decide whether to commission a proper translation.
- Screening and summarising. A model reads 50 abstracts and flags the ten that touch the claimed features. DeepIP's page describes semantic search across what it calls a 120M+ patent database. PATENTSCOPE's AI-assisted search, announced on 2 July 2026, does a different job: it turns plain-language instructions into structured keyword queries, using a model WIPO says it hosts on its own premises. One helps you build the search, the other ranks what it finds, and neither decides relevance.
- Widening the strategy. Ask the tool for synonyms, alternative terminology and classification codes you had not tried, then run those queries yourself.
- Drafting claim charts. Solve Intelligence's page lists evidence-backed charts. Check the evidence against the document itself.
Now the blind spots. First, similar is not the same as anticipating: a semantic result list ranks documents that resemble the query, and the reference that defeats a claim may share little vocabulary with it. Second, a clear result proves little. WIPO says applications are generally published 18 months from the filing date in many countries, so the newest filings in your field cannot be in any database yet. Third, summaries can distort. Read the passage the model quotes, not the paragraph it wrote about it. Fourth, what you did matters as much as what you found, so keep a log. This one is filled in, with illustrative entries:
| Date | Database | Query or class | Hits screened | Kept | By |
|---|---|---|---|---|---|
| [date] | PATENTSCOPE | Keyword set A plus IPC subclass [code] | 120 | 9 | [initials] |
| [date] | Commercial semantic search | Summary of the invention as the query | 50 | 6 | [initials] |
| [date] | Non-patent literature | Key terms from the claims | 40 | 2 | [initials] |
Under the table, record the publication date of the newest documents each database covers. If the question later arises why a reference was missed, that line and the log are your answer.
Here is the distortion risk in a concrete case. A screening prompt returns: "Reference D1 discloses a hinged bracket attached to a housing (paragraph 22)." The reviewer opens paragraph 22 and finds the bracket is slidable, while a different paragraph describes a hinge on a lid. The summary merged two features. Nothing was fabricated, and a reference is now wrongly rated as close. The fix costs about a minute per kept reference: open the quoted paragraph, read the sentence either side, and log the reference as verified only then. The same discipline applies to translations. Machine translation is fine for deciding whether a foreign document matters. For one you will rely on, have a person check the technical terms, which are a common source of slips.
For trade marks the same logic applies with different tools. WIPO's Global Brand Database covers international registrations under the Madrid system, geographical indications under the Lisbon system, certain protected emblems and marks from participating national and regional offices, and it offers image-similarity search. Its own page says it may be prudent to search national or regional registers too, and to consult a specialist. Picture a client launching a new range of cleaning products who wants a name checked in the relevant goods class. AI can group hits by class, rank them by visual and sound similarity and draft a neat summary table. It cannot see unregistered use, and it cannot decide whether two marks are likely to confuse buyers. That judgement stays with the lawyer, and the longer treatment is in AI trade mark search for small IP firms: what it finds and misses.
First drafts: from an invention disclosure to a specification
Drafting tools turn an inventor's disclosure into a first-pass specification, figures and claims. Solve Intelligence's page says it drafts complete applications, generates figures and analyses office actions. DeepIP's page describes a Word add-in that drafts from disclosures and responds to rejections, and its FAQ states plainly that the tool is not a substitute for professional legal counsel and that users should validate its outputs. A general assistant on a business plan can do a rougher version of the same work. ChatGPT Business and Claude Team both list standard seats at $20 a user a month on annual billing or $25 monthly, with a two-seat minimum, while Solve Intelligence and DeepIP show no prices on their pages.
The same rules govern all of them. Draft from the inventor's disclosure only. If the tool proposes an embodiment the inventor never described, put it to the inventor before it goes in, because in most systems matter added after filing cannot be relied on. And expect characteristic errors. Here is one, shown before and after:
AI DRAFT, claim 1: A fastening device comprising a housing
and the bracket, wherein the bracket is attached to the
housing by a hinge (14).
CORRECTED: A fastening device comprising a housing and a
bracket, wherein the bracket is attached to the housing
by a hinge."The bracket" has no earlier introduction, so it lacks antecedent basis, and the stray numeral belongs in the description, not the claim. Draft tools produce such slips often enough that a standing checklist pays for itself: antecedent basis in every claim, one reference numeral per part throughout, terminology identical between claims and description, and support in the description for every claimed feature.
A safer use of the same technology comes before drafting. Ask the tool to read the disclosure and list the questions an examiner or competitor would ask, then put those to the inventor:
Read the invention disclosure below. List the questions an
examiner or a competitor's lawyer would ask that the
disclosure does not answer. Group them as: alternatives to
each key feature; ranges and limits for each parameter;
advantages the inventor could demonstrate; features that are
essential versus optional. Questions only. Do not answer
them and do not add facts.The tool is generating questions, not facts. The answers come from the inventor, which keeps the specification inside what was actually invented and gives the drafter genuine alternatives to claim.
Responses to office actions carry a different risk. A tool that analyses a rejection and drafts arguments is useful. One that adds case-law citations, as DeepIP's page says its responses can, needs every citation opened and read. A fabricated or misdescribed authority in a filing to an office is a professional problem. The routine for catching them is in how to stop AI inventing case law, and drafting-specific risks are in AI for patent practitioners: drafting help and the disclosure risk.
Confidentiality before filing, tool by tool
An unfiled invention is the most fragile information an IP firm holds. WIPO says that an invention made public before an application is filed generally counts as prior art, that some countries offer a grace period of usually between 6 and 12 months, and that if disclosure cannot be avoided it should be under a confidentiality agreement. Whether an AI vendor's handling of a prompt amounts to a public disclosure is a legal question that turns on the jurisdiction and the vendor's terms. Nobody should find out the answer after a filing is attacked. Sort tools by what you may put into them:
| Tool type | Examples | Unpublished invention disclosure? |
|---|---|---|
| Consumer chatbot or free plan | Any personal account | No |
| General assistant on a business plan | ChatGPT Business, Claude Team, Microsoft 365 Copilot | Only after reading the terms, recording them, and getting the client's agreement; business plans do not train on content by default |
| IP-specific drafting tool | DeepIP, Solve Intelligence | After diligence: ask about training, retention, sub-processors and certifications. DeepIP's FAQ says it does not retain client data or use it for model training; Solve Intelligence's page lists SOC 2 Type II and ISO 27001 and says nothing uploaded to or produced by the product is used to train any AI model. Get both in the contract, not just on a web page |
| Private or in-house deployment | Firm-controlled environment | Best fit for the most sensitive matters, with its own security review |
Do three things regardless of tool. Put an AI paragraph in the engagement letter so clients know how their disclosures are handled. Keep a short "AI use" note on each file: tool, date, task and who checked. And where an invention is subject to security or export controls, treat data location as a question for the lawyer in charge, not the IT contact.
A paragraph for the engagement letter might read, to adapt with your own adviser:
Use of AI tools. We may use AI tools that we have reviewed
for confidentiality to help with administration, searches
and first drafts. We never put unpublished invention
disclosures or other confidential material into public AI
tools. A lawyer checks and takes responsibility for all
work filed or sent on your behalf. Ask us for details of
the tools we use and how your information is handled.Clients with unfiled inventions tend to read this paragraph closely, and some will ask to exclude AI from their matters entirely. A tag on the file that says "no AI" and a docket flag that blocks extraction tools for that matter costs little and removes the argument.
What the lawyer still checks, point by point
This table is the practical answer to the second half of the question. Each row names something AI can do, what it cannot, and the evidence that the check happened.
| Task | AI may | The lawyer checks | Evidence kept |
|---|---|---|---|
| Docket entry | Extract fields and quote their source | Each field against the letter, the date against the system rule, by a second person | Verified-by initials and date in the docket |
| Prior-art search | Translate, screen, summarise, suggest terms | Reads every kept reference; confirms the quoted passages exist | Search log with coverage dates |
| Trade mark clearance | Group and rank hits, summarise | Registers beyond the database; unregistered use; likelihood of confusion | Clearance memo signed by the lawyer |
| Specification and claims | First draft, figure ideas | Support, antecedent basis, numerals, nothing beyond the disclosure | Tracked changes against the draft |
| Office action response | Analyse, draft arguments | Characterisation of each cited reference; every authority opened | Citation checklist |
| Inventorship | Nothing | The human contribution to each claim; read the guidance your office publishes on AI-assisted inventions | Inventor interview note |
| Anything filed or sent | Nothing | The signing lawyer owns the content | Sign-off on the file |
Four beliefs that cause trouble
- "No hits means it is novel." A search shows what was found in the databases searched. It cannot show what is unpublished, in another language or in a source nobody queried.
- "A better model stops fabricated citations." Tools that ground answers in a database make them rarer. They do not make them impossible, and the cost of one in a filing is out of all proportion to the cost of opening the document.
- "AI docketing replaces the clerk." It moves effort from typing to checking. If the checking is real, you save time. If it is not, you have automated the creation of errors.
- "A business plan means confidential." A default of not training on your content is helpful. It is not the same as a contract with retention, sub-processor and deletion terms you have read.
Two weeks to a decision
A small practice can reach an answer in a fortnight with one lawyer and one clerk. Days one to three: pick the 40 letters and ten closed searches, and write down what the right answers were. Days four to seven: run extraction on the letters, screening on the searches and a draft on one closed disclosure. Days eight to ten: score it. For extraction, count matching letters. For search, count how many of the known relevant references the tool surfaced. For the draft, time how long it takes a lawyer to bring it to filing quality, against a manual first draft. Days eleven to fourteen: read the vendor's terms against the table above and decide.
| Test | What you count | Adopt if |
|---|---|---|
| Extraction on 40 letters | Letters with every field right, and errors the clerk check caught | The check caught every error, and at least 90 per cent of letters came through clean |
| Search screening on 10 closed searches | Known relevant references that appear in the tool's top 30 | It adds to your method; it never replaces it |
| Draft from 1 closed disclosure | Lawyer's editing minutes against a manual first draft | Editing time falls and the checklist finds no new-matter or antecedent-basis problem |
A sensible adoption rule, in numbers: adopt extraction if the checks catch every error in the test, whatever the error rate; adopt search screening only as a supplement to your existing method; adopt drafting only if lawyer editing time falls and no new-matter or antecedent-basis problem slips past the checklist. If you would rather work through the pilot with someone, that is what my AI implementation consultation covers, and the contact page is the place to start. For comparing tool types first, see best AI chatbot for lawyers: general models versus legal tools.
Official WIPO pages used above: PATENTSCOPE, the Global Brand Database and the patent FAQ on novelty and disclosure.
Questions IP practices ask about AI tools
Can AI calculate patent or trade mark deadlines?
Treat it as a reader, not a calculator. AI can pull the dates and wording printed in a letter, but the deadline of record should come from your docketing system's rules and be checked by a second person. Offices differ on whether a period runs from the date printed, posted or received, and extensions vary by office and document type.
Is it safe to paste an invention disclosure into ChatGPT?
Not into a consumer account. WIPO notes that public disclosure before filing can defeat patentability in many countries and advises a confidentiality agreement when disclosure cannot be avoided. Use only tools with contractual confidentiality and no training on your data, record the terms you relied on, and get the client's agreement first.
Do specialist IP tools publish their prices?
Some do. Alt Legal prices docketing by the number of matters, from $60 a month for up to 50 to $295 for up to 400, with a trade mark protection add-on and quotes for bigger dockets. Solve Intelligence and DeepIP showed no prices on the pages I read; DeepIP offers a free trial through a form. Ask about seat minimums, trial terms and what happens to your data when the contract ends.
Can an AI trade mark search replace a clearance search?
No. WIPO's own Global Brand Database page suggests it may be prudent to search national and regional registers as well and to consult a specialist. AI can sort and summarise results, but it cannot see unregistered use or pending filings that are not yet in a database, and it cannot decide whether two marks are likely to be confused.
Further reads
- AI Paralegal Software: What a Small Firm Should Test Before Paying — A test plan for AI paralegal tools before a small firm pays.
- Which Legal Tasks Should a Small Firm Never Hand to AI? — The tasks that stay with a lawyer, with the reasons.
- AI App for Lawyers: Score Accuracy, Privacy and Cost in a Short Trial — A scoring sheet for accuracy, privacy and cost in a short trial.
- AI Contract Review for Small Firms: What It Catches and Misses — Useful for licences and assignments, which IP firms review daily.
- Can Solicitors Use ChatGPT Without Breaching Confidentiality? — The confidentiality question for any client file, answered plainly.
- How to Summarise Bundles and Transcripts With AI Safely — How to summarise long documents without losing the detail that matters.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Alt Legal home and pricing pages, DeepIP and Solve Intelligence product pages (viewed 4 November 2026); WIPO pages on the Paris Convention, PCT FAQs, patent FAQs, PATENTSCOPE news and the Global Brand Database; ChatGPT Business and Claude pricing pages (checked November 2026). Worked figures are illustrative.