From assistants to agents
For most of the last five years, AI in patent practice meant a smarter search box. Semantic retrieval replaced keyword strings, and generative models started summarising office actions. Useful — but still fundamentally a tool that a human drove one query at a time. Agentic AI changes the unit of work: instead of a single query, the attorney delegates an outcome ("assess novelty against the top 20 references and propose claim amendments"), and the agent plans, retrieves, drafts, critiques, and returns a defensible package.
What agents do well today
Three workflows are already production-grade at leading firms: (1) prior-art triage — clustering 200+ hits into novelty, obviousness, and background buckets with citations; (2) office-action first-drafts — mapping examiner rejections to claim elements and proposing amendment strategies grounded in file-wrapper history; and (3) docket hygiene — reconciling annuities, deadlines, and family relationships across jurisdictions with proactive escalation.
Where humans still lead
Claim strategy, inventor interviews, examiner interviews, and any communication with the client remain firmly human. Agents are excellent at compressing hours of retrieval and comparison into minutes; they are poor at judgement calls where commercial context, competitor posture, and litigation risk intersect. The most effective teams use agents to prepare the record so attorneys spend their time on decisions, not document assembly.
What leaders should do in the next 90 days
Pilot one workflow end-to-end (we recommend prior-art triage), measure the time saved per matter and the quality delta against your senior associates' baseline, and publish an internal guideline on what an agent may and may not sign. Do not roll out generative drafting without a review protocol — the failure mode is not hallucination, it is confident mediocrity that passes casual review.
