
Superintelligence for Trademark and Patent Firms: Managing Deadlines
How to track follow-up deadlines with AI at a trademark and patent firm, without letting the system judge technical merit. Limits and steps.
Want to apply this in your company?
Book a call with Juan →At trademark and patent firms, AI is most useful for tracking office action deadlines and follow-ups, not for weighing in on the merits of an application. A safe design separates two layers: a deterministic part that reads dates, statuses and official publications and triggers alerts, and a human part that decides what to answer. When that boundary is set before rollout, the system reduces missed deadlines without taking on a role that belongs to the responsible professional.
Key takeaways
- A deadline is a date and a rule, not an interpretation. The counting method must follow explicit rules and be checked by a person.
- Novelty, inventive step, distinctiveness and likelihood of confusion stay outside the reach of an AI agent.
- Every alert needs an owner, a date of acknowledgment and a record of who decided what.
- Start with one real workflow, one pilot office and a success metric agreed on before you begin.
- Client data and application documents fall under LGPD rules from day one.
Why an office action deadline is not an interpretation problem
A formal office action has a date, a text that describes it and a consequence if the deadline passes without a response. These elements can be verified. What needs to be answered, with which arguments, and whether it is worth responding at all depends on technical analysis. When a system mixes the two, the result is usually a well-formatted alert that suggests a reply with the tone of a legal or technical opinion, and that creates a risk nobody noticed during implementation.
For that reason, the first design decision is not choosing a model. It is listing what the agent may do on its own. In general, the agent can locate the publication or decision, identify the type of communication, calculate the deadline using the firm's adopted rule, notify the responsible person and log everything. Analysis of the merits, drafting the response and the decision to abandon or appeal remain with the team.
What the AI can read and what stays with the responsible professional
The most useful division is a decision table written as a list, which the team reviews before the pilot:
- The agent can do: read status from authorized sources, identify the type of communication, calculate deadlines under a fixed rule, prepare reminders and record acknowledgment.
- Human review is required: confirm the calculated date, classify ambiguous requirements and assign who takes each case.
- Only the professional handles: assess the merits, write the response, define the defense strategy and authorize any submission.
This design also prevents the model from resolving an ambiguity on its own. If a communication does not fit any known category, the workflow should stop and call a person instead of trying to classify it.
Step by step: setting up deadline control
A deadline control process without AI already requires discipline. With AI, that discipline has to be written down. A practical sequence looks like this:
- Map the current workflow: where communications arrive, who reads them, where they are recorded and how follow-up happens.
- Write the firm's counting rule down and ask the responsible professional to validate it.
- Build a list of known communication types and the expected action for each one.
- Set a lead time for alerts and define who receives each alert when no one responds.
- Record every alert, every acknowledgment and every human decision in a searchable log.
- Test with past communications that were already resolved, comparing the agent's result with what the team actually did.
- Only then open the pilot with live cases, with daily supervision at the start.
The log is the most important piece. Without it, nobody can audit a date error or explain why an alert was ignored.
Illustrative scenario: an office action that arrives on a Friday
Illustrative scenario: an office action is published late on a Friday afternoon. The agent identifies the file, calculates the deadline using the firm's rule and sends a notice to the case owner and to the coordinator. The notice includes the file number, the type of communication, the calculated date and a link to the original document, with no paragraph suggesting a response. On Monday, the responsible professional opens the document, decides it needs a technical opinion and marks in the system that they have taken the case. The agent records that decision, stops sending general reminders and keeps only the day-before alert. None of this required the AI to understand the merits. It only required that every step had an owner and a record.
When the agent should stop and call a person
A well-designed agent needs to know when it should not continue. Some situations call for an immediate stop and human handoff:
- The calculated date conflicts with another source or with the recorded rule.
- The document is incomplete, unreadable or in a format other than expected.
- The communication does not belong to any known type.
- A client request changes the deadline, the urgency or the scope.
- Someone tries to make the agent generate merit arguments or send a response without review.
That last point deserves attention. Text inside external documents and messages can contain instructions that try to change how the agent behaves. The OWASP guidance on risks in applications built with language models covers this kind of failure, and the practical answer is to restrict what the agent is allowed to do, rather than relying only on the model behaving well.
Client data, access and LGPD
Industrial property files contain data about individuals, companies and sometimes information that has not been made public. The agent should receive only what it needs for its task, with read-only access where possible and no permission to send documents or change records. Which legal basis covers each form of processing, how long data is kept and how a data subject exercises their rights are questions the firm needs to answer together with its legal counsel. The ANPD website publishes official guidance on these topics. If the solution uses models hosted outside Brazil, it is also worth reviewing the requirements for international data transfer.
The NIST AI Risk Management Framework offers an organized way to record risks, controls and owners, which can support that review without turning the project into bureaucracy.
How to apply the method gradually
Our working method follows four stages: map the current process and its baseline, limit data and actions, pilot a real workflow with records and an agreed success metric, and transfer the operation to the team, then decide whether the solution should be expanded, adjusted or stopped. In industrial property firms, the limiting stage is usually the slowest, because the team has to agree on what the agent will not do. That is uncomfortable, but it is exactly what makes the pilot safe. For those who want to test the logic before connecting to live systems, simulating communications with fictitious data also helps, as described in the content on testing with mocks and sandboxes.
If the question is what separates narrow AI from something broader, it helps to distinguish the concepts before discussing tools. A deadline system needs a restricted and predictable AI, not any promise of general capability. A more detailed explanation of the topic is available in Is ChatGPT superintelligence? What it is and what it is not.
If your firm wants to compare this approach with other sectors that handle deadlines and triage, the case of superintelligence for notary offices: AI triage and deadline reminders has useful parallels, especially in separating what is record keeping from what is decision making.
Frequently asked questions
Can AI calculate an office action deadline on its own? It can calculate the date from a written rule that the firm has validated, but a person should check the result before it becomes a commitment. Because the counting rule changes with the type of communication, the most common error is applying a general rule to a case that has an exception. The calculation log should show which rule was used.
Can the agent suggest a response to an office action? In general, it should not. Suggesting response text touches the technical merits, which is the responsibility of the qualified professional. A safer design limits the agent to alerting, summarizing the original document and routing the case. If you want drafting support, it should be an internal draft, reviewed and signed by whoever is responsible for the file.
How long does it take to put a deadline control system into operation? It depends on the volume of files, the quality of the source data and the clarity of internal rules, which are usually the bottleneck. A pilot with a limited set of files and daily supervision is typically the most realistic path. The decision to expand should come from a metric agreed at the start, not from a feeling that the system seems to work.
Do I need a special superintelligence system for this? No. Deadline control is a problem of process, data and records, and most of the value comes from clear rules, reliable alerts and defined owners. Terms like superintelligence attract attention, but a firm that wants to operate safely needs tools that are predictable, auditable and bounded by well-written limits.