A patent landscape of 3,146 agentic AI families shows that operators, banks, autonomous-vehicle companies, and Chinese research institutions are building the moat.
Patent intelligence from 3,146 families filed since 2019. Wall Street is in the top 10. The foundation-model labs are not. Here is what the data says, and what to do about it.
If you only read the press releases, agentic AI looks like an OpenAI, Anthropic, and Google story.
The patent data tells a different one.
Across the global patent system, 3,146 patent families have published since 2019 that claim or describe agentic AI techniques. The filers are not the lab leaders most people associate with the field. They are operators, banks, autonomous-vehicle companies, and Chinese research institutions. And the velocity is now extreme: filings nearly tripled in 12 months and the 2026 figure is already on pace to exceed 2025 by half.
If you are building, buying, or funding agentic systems and your prior-art surface is just Big Tech, you are missing about two-thirds of the relevant landscape.
This is what the patent office is recording. And it changes how to position.
Source for every number that follows: Minesoft Origin API, query run May 12, 2026, extended-family grouping. Full Boolean query is in the appendix below.
Agentic AI patent families published per year, 2019 through May 12, 2026:
That is a 4.86x jump from 2023 to 2025. Four and a half months into 2026, the count is already 87% of the full 2025 total. At the current run rate, 2026 will land somewhere in the 2,500 to 2,800 range, roughly 2.3x the prior year.
For context, this is faster than the early curve of any AI sub-field tracked in the last decade except for generative-AI image synthesis (which was driven by a small number of foundation-model labs, not the broad operator base showing up here).
The growth is real, sustained, and broad. It is not driven by one company.
Here are the top 14 filers by extended family count, color-coded by camp:
Two things stand out.
Physical autonomous agents. May Mobility (109 families), Gatik AI (82), and State Grid Corp of China (72) are filing on perception, planning, fleet coordination, and field operations. The vocabulary is “autonomous agent” but the technology is moving robots and vehicles. ABB sits in this camp on the industrial side, and the AV cluster grows further down the list with companies that fall outside the top 14.
Enterprise software agents and orchestration. Oracle (93), Intel (49), Morgan Stanley (31), Citigroup (25), IBM (20), Salesforce, and Microsoft. These are filings on agent orchestration, tool-calling, multi-agent coordination, memory systems, and the integration between agentic systems and existing enterprise data. This is the camp most people picture when they hear “AI agents.”
The same Boolean query pulls back both because the language overlaps, but the technology stacks do not. Anyone scoping freedom-to-operate on a software-agent product needs to be alert to the AV cluster – broad claim language on “planning” or “task decomposition” will pull AV filings into your search results whether you want them or not.
Morgan Stanley (31 families) plus Citigroup (25 families) combined are filing more agentic-AI patents than IBM (20). US Bancorp shows up at 3 and rising. This is not vanity filing. Banks file when they expect to defend, license, or use IP as a differentiator on enterprise services.
Reading the filing topics, the bank patents cluster around compliance workflows, document handling, customer service orchestration, and regulatory reporting. These are the highest-value, highest-friction parts of bank operations. Agentic AI is a labor-savings story for them, and they are protecting the implementations.
OpenAI, Anthropic, and Mistral together account for fewer than 10 published families in the dataset. Google and Microsoft have small but growing positions. The simplest explanation is the 18-month publication lag: foundation-model lab work filed in 2024 and 2025 will not surface in the search corpus until 2026 and 2027.
But that lag itself is the strategic point. The published prior-art landscape today is biased toward operators and incumbents. If you are filing now and you want to clear the operator filings, you have a window. If you wait for the lab filings to land, the landscape will be much more crowded.
Zhejiang University, Tsinghua, Beihang, the Chinese Academy of Sciences, and the Beijing General AI Research Institute all sit in the top 15 globally. Inspur Group, a state enterprise, ranks at 33. This is a different shape than the US picture. In China the IP is being generated inside research institutions and rolled into state-aligned corporate hands. In the US it is enterprise-led from the start.
For anyone tracking competitive risk, the implication is that Chinese filings will tend to surface through licensing arrangements and state-backed commercial vehicles rather than through visible startup acquisitions.
The 3,146 families cluster across recognizable CPC subclasses. Here is the top-10 distribution:
A family can land in multiple subclasses, so the counts do not sum to 3,146. But the distribution is the story. Roughly one-third of agentic AI filings are physical-systems patents wearing software-agent vocabulary. Another third sit in core AI substrate. The remaining third covers the enterprise, communications, and business-method work most people would call “agents.”
Anyone framing agentic AI as a software-only field is reading half the data.
Reading sample filings across the top assignees, the recurring claim themes are:
The opening in this landscape is at the integration layer. Generic orchestration is already heavily filed. Memory is lighter. Domain-specific integrations are filed in isolation by operators but rarely cross-claimed against the underlying agent architecture.
The Chinese filings concentrate in domestic-only filings from universities and state enterprises, with limited PCT activity. The US filings skew toward enterprise software, finance, and AV, with healthy PCT and EP activity.
Practical reading: if you are operating only in the US, your active competitive surface is much smaller than the global 3,146 number suggests. The relevant prior art for a US filing is roughly the 665 US families plus the 120 PCT filings plus any inbound nationalizations from CN and EP. That is closer to 900-1,000 active references, not 3,146.
But the WO and EP filings tell you which Chinese inventions are likely to follow you across borders. Watch those entries closely. They are the leading indicator for which Chinese research will arrive in the US examination corpus 12-24 months later.
The picture differs by role.
The white space is not in orchestration as a generic concept. The big players already have positions there. Oracle and Salesforce in particular are filing heavily in the workflow-orchestration core.
The opening is at the integration layer between agentic systems and specific industry workflows. Compliance, claims processing, manufacturing operations, healthcare intake, regulatory reporting. The domain-specific agent applications are filed in isolation by operators but rarely cross-claimed against the underlying agent architecture. If your invention sits at the integration point and you can claim both the technical method and the domain-specific application, you have a strong filing surface.
Memory and long-horizon state is lighter than the other clusters. Filings there will face less prior-art crowding right now. That is a window that will close.
An agentic AI startup with no patent strategy in 2026 is a bigger flag than it was in 2024. The incumbents are filing fast. The 4.86x jump from 2023 to 2025 means competitor density doubles roughly every 8-10 months at the current rate.
If a portfolio company is shipping orchestration tech with zero filings, that is a position you will not be able to defend in a strategic exit. The acquirers’ diligence teams will run exactly the search shown above and notice the gap.
The same data also tells you which technical bets have IP defensibility ahead of them. Watch the integration-layer companies and the memory-architecture companies. They have room.
The Wall Street filings are a quiet signal. Banks are not filing for fun. They file because they expect to license out, defend against operational copying, or use IP as a moat for differentiated enterprise services.
Worth tracking specifically who they cite in their continuations. Backward citations from a Morgan Stanley or Citigroup family are a curated list of the prior art the bank’s outside counsel considers material to operationalizing agentic AI in financial services. That is free competitive intelligence.
On the operator side, the AV agent cluster (May Mobility, Gatik AI) is undervalued relative to its IP weight. May Mobility alone has more agentic-AI filings than IBM. That has not yet shown up in public-market multiples.
Prior-art coverage on agentic AI claims is much wider than the “ChatGPT and friends” model would suggest. Operators, banks, and Chinese universities all need to be in your search corpus. Specifically:
The agentic AI race is being run on two different tracks. The press covers the foundation-model labs. The patent office is recording a different story, where the active filers are operators, banks, autonomous-vehicle companies, and Chinese research institutions.
Three takeaways:
If you are building, buying, or funding agentic AI systems, the next 12 months are the right window to file ahead of the foundation-model lab work that will surface in 2026 and 2027.
All data in this article comes from a single Minesoft Origin API query run on May 12, 2026.
Query:
ATAC=(("AI agent*" OR "intelligent agent*" OR "autonomous agent*" OR "agentic") W5
(orchestrat* OR planning OR reason* OR "tool use" OR "tool calling" OR multi_agent
OR autonom* OR "task decompos*" OR "task plan*")) AND PD>20180000
Grouping: extended family
Total families: 3,146
Date range: publications from 2019 through May 12, 2026
CPC subclass distribution was pulled via the same query against the cpcfirst4chars dimension. Geographic distribution via country. Top-filers via ultimate_owner (which rolls subsidiaries under their corporate parents).
Caveats:
This analysis is for strategic planning purposes and does not constitute legal advice. For legal opinions on patentability, validity, or freedom to operate, consult qualified patent counsel.
Need to understand what this means for your portfolio? ipCapital Group helps leadership teams turn patent landscapes into practical decisions about filing, freedom to operate, M&A diligence, and monetization.
Start a conversation with ipCapital Group or download the IP Strategy Playbook.
Work with ipCapital Group
From invention to monetization, our team has guided 2,000+ engagements across the full IP lifecycle. Start with a free 30-minute discovery call.
Written by
Seth Cronin