AI for Patent and IP Firms: Dockets, Deadlines and Updates

AI for a patent law firm works best as a monitoring and drafting layer around the docket, not as a replacement for it. An agent can read incoming office correspondence, compare it with the docketing system, alert the right attorney and prepare client status updates for review. The attorney and the docketing team keep every decision that carries legal consequence.

Key takeaways

• The docketing system stays the source of truth. The agent reads it, compares against it and proposes changes, but does not silently rewrite it.
• The most useful first jobs are catching mismatches between incoming notices and docket entries, tracking whether alerts were acknowledged, and drafting routine client updates.
• A good alert names its source document, its owner and what happens if nobody responds. Fewer, better alerts beat constant noise.
• Anything sent to a client, filed with an office or interpreted as legal advice passes through an attorney first.
• Unpublished inventions need per-matter permissions, treatment of every inbound email as untrusted input, and logs that a partner can read.

What should an AI agent do in a patent firm, and what should it leave alone?

Patent and IP practice is deadline work with judgment attached. Responses to office actions, maintenance windows, priority dates, annuity payments and foreign filing steps all depend on dates, and each date depends on a rule, an instruction and sometimes an extension. Most of the risk is not in the arithmetic. It is in a notice that sits unread, an instruction that never reaches the docket, or a status question that waits in an inbox.

That shape suggests a clear division of labor. The agent handles reading, comparing, flagging, reminding and drafting. People handle calculating the controlling date, choosing a strategy, filing, signing and advising.

A simple decision guide for each task you consider handing to an agent:

• If the task is reading and sorting incoming documents, the agent can do it, with a person sampling the results.
• If the task is comparing a notice against a docket entry and reporting a mismatch, the agent can do it and route the result to docketing staff.
• If the task is proposing a new docket entry, the agent drafts it into a review queue and a person approves it.
• If the task is sending something to a client, the agent drafts it and an attorney approves it, at least until a narrow template has proven itself.
• If the task is deciding what a deadline is, whether to extend, abandon or amend, or what a claim means, it stays with the attorney.

This is the same principle described in the setup guide for AI agents in law firms: the agent supports the people who own the outcome, and the approval boundaries are written down before anything goes live.

How does docket monitoring work in practice?

Docket monitoring starts with defining what the agent can see. In a typical setup it has read access to the docketing system, to the shared mailbox where office correspondence arrives, and to the matter folders where instructions and filings live. It does not need write access to the docket to be useful.

Once those sources are connected, the agent performs a handful of recurring checks:

• A new notice arrives and the agent looks for a matching matter and a matching docket entry. If none exists, it flags the notice as unmatched.
• A notice carries a response period and the docket shows a different date. The agent reports the discrepancy with both sources side by side.
• A client instruction arrives by email, for example to proceed with a foreign filing, and the agent checks whether a corresponding task exists.
• A docket entry has had no activity as its date approaches, and the agent lists it for the responsible attorney.
• A matter changes status in the system, and the agent notes whether the client has been told.

Each finding goes into a review queue with a link to the source document and a short plain-language description. A docketing professional works that queue. If the agent proposes a new entry, the proposal is clearly marked as a proposal and the person who approves it is recorded.

This design avoids the most common failure of automation in legal settings: a tool that is confident, quiet and wrong. Because the agent only proposes, an error costs a minute of review, not a missed date.

What makes a deadline alert trustworthy?

Alerts fail when people stop reading them. A firm that already receives many reminders does not need another stream of them. It needs alerts that carry enough information to act on and that follow up when no one does.

A useful alert includes:

• The matter, the event and the controlling date as recorded by the docketing team.
• A link to the source document that triggered it.
• The named owner, not a group mailbox.
• The next step expected, such as confirm strategy, request client instruction or prepare response.
• An escalation path with a clear trigger, such as no acknowledgement by a set point before the date.

The agent's most valuable contribution is often the follow-up. It tracks whether the owner acknowledged the alert, whether the client instruction arrived, and whether the task moved. Silence is never treated as acknowledgement. If an alert goes unanswered, the agent escalates to a second person according to a rule the firm wrote, not one the agent invented.

A weekly reconciliation helps too. The agent produces a short report of upcoming items with no recorded activity, unmatched notices and alerts still waiting for acknowledgement. A partner or docket manager reads it in a few minutes. Over time this report also shows where the process leaks, which is useful whether or not any AI is involved.

For teams that need broader evidence trails around reviews and alerts, the approach overlaps with what is described in AI for compliance teams.

How can clients get status updates without attorneys writing each one?

Clients of IP firms ask similar questions repeatedly. Was the application filed? Has the office responded? What happens next, and when do you need something from us? Attorneys know the answers, but writing each reply interrupts work that needs concentration.

An agent can prepare these updates from facts already in the docket and the matter file. It works in two tiers.

• Factual status updates: a filing was made, a notice was received, a date is approaching, an instruction is needed. These follow an approved template and draw only from recorded data.
• Narrative updates: anything that explains what a rejection means, what options exist or what is recommended. The agent can draft a skeleton from the attorney's own notes, but the attorney writes or rewrites the substance.

Both tiers pass through an approval step. As a template proves reliable over many reviewed drafts, a firm may decide to let a narrow class of purely factual messages go out after a lighter review. That decision belongs to the firm, and it can be reversed.

The agent also keeps client memory in a structured way: who the day-to-day contact is, who must be copied, which language the client prefers, whether the client wants a monthly summary or only action items. Be Human works in English and Portuguese, which matters for firms whose clients and foreign associates communicate in both.

An illustrative scenario:

• A mid-sized firm receives an office notice for an application. The agent matches it to the matter and sees that the docket already has a response date, so no mismatch is flagged.
• It drafts a factual client message saying the notice arrived, the response period is being tracked, and the attorney will send a recommendation. It places the draft in the responsible attorney's queue.
• The attorney edits one sentence and approves it. The agent sends it, records the approval, and sets a reminder to follow up if the client has not replied before the internal decision date.

Nothing in this flow required the agent to interpret the notice. It moved information to the right people faster and left a record.

How do you protect unpublished inventions and privilege?

Unpublished applications are among the most sensitive material a firm holds. Any agent that touches them needs narrower access than a general assistant would get.

Start with permissions. Give the agent access matter by matter, or by team, rather than across the whole firm. The OWASP [authorization guidance](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) recommends least privilege and checking permissions on every request, which fits well here: if an attorney cannot open a matter, an agent acting for that attorney should not either. Firms that maintain internal walls between clients will recognize the same concern discussed in AI for engineering firms and client walls.

Next, treat everything arriving from outside as untrusted. Emails, attachments and web pages can contain text that tries to steer an AI system, for example an instruction hidden in a document telling the agent to forward files. OWASP's [prompt injection prevention cheat sheet](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html) describes practical defenses, including separating instructions from data and limiting what tools the agent can call. In a patent firm that means the agent can read a notice but cannot send external email, move files or change the docket on the strength of what that notice says.

Then keep logs a human can read: what the agent read, what it proposed, who approved and what was sent. These logs also support the governance practices outlined in the NIST [AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework), which emphasizes mapping, measuring and managing risk across the system's life.

Finally, check outside counsel guidelines, client agreements and your jurisdiction's professional rules on technology and confidentiality before connecting any system. Those rules vary, and this article is not legal advice.

How do you start without disrupting the docket?

A first engagement should be small enough to supervise and useful enough to matter. One workable sequence:

• Step one: choose one role, for example docket watcher, and write what it should and should not do in a single page.
• Step two: list the sources it may read and confirm that access is read-only wherever possible.
• Step three: define the review queue, who works it, and what counts as an acknowledged alert.
• Step four: collect a set of past notices and instructions with known outcomes, and test the agent against them. Include awkward cases such as misfiled notices and ambiguous instructions. Anthropic's guidance on [evaluating AI agents](https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents) is a useful reference for building such tests.
• Step five: run the agent in shadow mode, where it produces findings but nobody acts on them, and compare them with what the docket team found.
• Step six: turn on alerts for a small set of matters, review misses and false alarms weekly, and widen scope only when the numbers of surprises fall.
• Step seven: add client status drafting, with every message approved, once the monitoring is stable.

Be Human designs this first role with the firm, including the memory and tools it needs, the approval boundaries and an observable success criterion, such as unmatched notices surfaced before the docket team found them. In managed operations, Be Human can then monitor the agent, run recurring evaluations, maintain integrations, review permissions, respond to incidents and make controlled improvements, with the firm's attorneys and docket staff keeping control of decisions.

Frequently asked questions

Does an AI agent replace docketing software or docket staff? No. Docketing software remains the system of record, and trained docketing professionals remain responsible for controlling dates. The agent adds a second pair of eyes that reads continuously, flags mismatches and chases acknowledgements.

Can AI be trusted to calculate deadlines? It should not be the final authority. Rules differ by office and situation, and extensions and exceptions need professional judgment. A safer use is checking that a calculated date in the docket matches the one stated in the notice, and flagging any difference for a person.

Is it safe to use AI with unpublished applications? It can be, if the setup limits access by matter, treats inbound content as untrusted, restricts what the agent can send or change, and records its actions. Data handling terms with any model provider should be reviewed before connecting anything, and the answer depends on the specific configuration and your client commitments.

How small can a first project be? Small enough to be one role on one workflow, such as watching a shared mailbox for notices that have no matching docket entry. That alone produces a visible result, lets the team judge accuracy, and creates a base for deadline alerts and client updates later.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.