Federal AI deployment has moved past the pilot stage. The USPTO’s launch of the Trademark Classification Agentic Codification Tool, Class ACT, is a production deployment, not a proof of concept. And the task it automates is genuinely consequential for the 700,000-plus trademark applications the office processes annually.
Class ACT handles three specific pre-processing functions: assigning international classes to unclassified applications, generating design search codes, and producing pseudo marks. These are the administrative classification steps that currently precede human examination, and that, according to the USPTO’s official announcement, consume substantial time in the pre-examination queue. Per the USPTO’s official announcement, the tool can immediately assign international classes to unclassified applications and generate the design search codes and pseudo marks that follow.
The agency’s stated time expectation is stark. The USPTO states Class ACT is expected to reduce classification preparation time from approximately five months to five minutes. That framing deserves precise reading: this is the agency’s own projection, not a measured outcome. The tool is newly deployed. Five-to-five is the target, not the confirmed result. How that projection holds against the full complexity of real application volume, specialty goods categories, compound marks, translation edge cases, will take months of production data to assess.
What’s confirmed is the scope of what Class ACT automates. International trademark classification (the Nice Classification system) and design search code assignment (the Vienna Classification) are structured but judgment-intensive tasks. Automating them at scale, reliably enough for pre-examination routing, represents a meaningful capability claim for an agentic system. The USPTO plans to introduce additional AI-enabled trademark tools, a signal that Class ACT is positioned as the first in a series, not a standalone deployment.
A note on technical framing: Class ACT is an agentic classification tool, not a generative AI or large language model deployment. It performs structured, task-directed work within defined classification schemas. That distinction matters when evaluating governance implications. This is agentic AI in a narrow, well-defined domain, the kind of deployment that AI governance frameworks like the NIST AI RMF were designed to assess.
What to watch: Three things. First, accuracy rates. If Class ACT’s classification outputs require frequent correction in the examination stage, the time savings at pre-processing get offset by downstream costs. The USPTO’s ability to report accuracy metrics transparently will be a test of government AI accountability norms. Second, practitioner workflow impact. IP attorneys and trademark agents who currently advise clients on classification strategy during the application process will need to understand how Class ACT’s outputs interact with their guidance. Third, whether other federal agencies replicate this deployment model. The USPTO is not a typical tech-forward agency. If Class ACT performs, it provides a replicable template for agentic AI in high-volume administrative processing across government.
TJS synthesis: Class ACT is notable not just for what it does but for where it happens. The USPTO operates under public accountability constraints that private sector AI deployments don’t face, every classification error affects a real applicant’s application timeline. A federal agency deploying an AI agent to automate pre-examination processing at this scale is a governance benchmark event, not just a technology story. Watch how the USPTO reports performance data. The transparency standard it sets, or doesn’t set, will matter as other agencies follow.