Nov 7, 2026 · By Oleg K. Salakhov, Partner, Berkeley Law & Technology Group, LLP
Advice on responding to patent eligibility rejections is not in short supply these days. Practice guides, webinars, and on-line articles explain, in considerable detail, how to argue with an examiner once such a rejection arrives. Far less attention appears to go to the patenting strategy decisions that come earlier: how to structure a patent application so that it is examined in the right place to begin with. As a result, much of the guidance treats symptoms, and, comparatively, little addresses prevention. This article focuses on that earlier stage: how the USPTO routes software and AI patent applications to its art units, and how your patenting strategy, and specifically your drafting choices, can influence whether your application ends up in the emergency room or at a routine wellness checkup.
Every patent practitioner has a war story or three about the USPTO’s art units where patent applications go in and rarely come out. Most of those stories share a similar ending – the fate of the application was potentially decided the day it was classified and routed, not the day the examiner wrote the first rejection.
However, if you look at this from the right angle, there is some good news. Namely, patent application routing can be largely influenced by your own actions – more specifically, your words – that you can control: your title, your abstract, your figures, and your claim 1. Write them well and you increase your likelihood of skipping a fight you are statistically likely to lose. Write them carelessly and you buy – at your clients’ expense – years of office actions, RCEs, and, possibly, appeals.
So, where exactly are these patent routes you want to avoid? Let’s look at the map.
Where Software Applications Go to Struggle
If you look at software as a whole, it does relatively well at the USPTO: across roughly 250 software art units, average allowance rates run from 36% to 94%. The trouble is concentrated in a few neighborhoods, and, as the table below shows, the spread is wide. To steer clear of these neighborhoods, you need to know how the steering works.
| Neighborhood | Art Units | Statistics |
| Business Methods | 3620s, 3680s, 3690s | All Business Methods Art Units: 93% of rejections cited AliceArt Unit 3689 fell to about 3% allowance after AliceArt Unit 3626 is currently at a 38% three-year grant rate |
| AI and Machine Learning | 2121 – 2129 | Median allowance rate is 62%; the lowest art units (2122, 2123, 2128) average 51% – 55% |
| AI-Related Art Units in the 2140s Group | 2142, 2147 | Average allowance rate is 44% – 45% in one analysis; another reports a 68% three-year grant rate for each art unit. Art Unit 2142 mainly examines user interfaces. |
| Database and Retrieval | 2169 | Average allowance rate is 47% |
| Friendlier Technical Art Units | 2661, 2676, 2657, 2436 | Allowance rates range from 81% to 86% |
How Routing Works
There is a pipeline for routing patent applications to art units, and every stage in this pipeline reads your words.
- AI suggests, humans decide. Since December 2020, the USPTO’s AI auto-classification tool has been used to read your claimed subject matter and suggest CPC codes, but human contract classifiers still assign the official ones.
- The claims carry the most weight. The MPEP ties the routing codes to what you claim overall. In practice, however, your claim 1, title, abstract, and figures appear to dominate, with claim 1 carrying the most weight.
- The old class system still picks your art unit. The Office tried routing by CPC codes in 2020, then paused in August 2022 after examiners got cases they weren’t qualified to examine. So, it looks like art units are still being assigned by legacy USPC class, despite what MPEP § 909 says. A recent GAO report says that a new system is being explored, but none has been launched or publicly confirmed.
- The same codes guide the prior art search. Classification determines not only where your application is routed, but also where the examiner searches for prior art. So, misclassification can cost you twice: dealing with the wrong examiner and with the wrong prior art.
But, this pipeline also hands you a strategy. Since legacy USPC classes appear to still pick your art unit, two of them work as trapdoors: class 705 (business methods) and class 706 (AI). As such, their official definitions can serve as your list of trigger words to avoid up front.
The USPTO’s class 705 definition potentially points to two ways out:
- Keep the business context in the background. Class 705 is for claims that actually recite business data processing. If your independent claims, especially claim 1, focus on the technical processing, your application is less likely to be classified in class 705.
- Anchor your claims in a particular device. Data processing combined with a significantly claimed device is generally classified with that device. Tie your independent claims to sensors, vehicles, cameras, or network hardware, and your application is more likely to be examined as that particular technology.
Class 706’s own definition offers a similar way out, along with one caveat:
- Frame AI as part of a specific technology. Class 706 is the default home for artificial intelligence as such, and it can also encompass AI combined with other fields, especially AI techniques no other class covers. If your independent claims, especially claim 1, tie your AI model to a specific field it improves (like image analysis, speech processing, network security, vehicle control), your application is more likely to be classified in that field’s class than in class 706.
- But watch out for generic “pattern recognition.” The class 706 definition expressly lists machine vision, acoustic recognition, and sensor fusion as combinations it can take. So, recite a specific task and its technical result in your independent claims, especially claim 1, and reflect them in the title and abstract (for example, detecting cracks in turbine blades from infrared images with fewer false alarms), rather than claiming a system that “recognizes patterns” in any kind of data, technical or otherwise.
The Playbook: Words That Can Open (and Close) Trapdoors
Keep class 705 words out. These trapdoor terms are drawn from class 705 and CPC G06Q subgroup titles. Keep them out of your title, abstract, and, especially, claim 1, and use them in figures sparingly, alongside technical diagrams. Reserve them for dependent claims and the specification, where they are less likely to drive unwanted classification.
| Category | Trapdoor Terms |
| Payments | payment, billing, invoice, fee, pricing, funds transfer, credit/debit, banking, checkout, payment gateway, settlement, digital wallet |
| Commerce | purchase, transaction, merchant, customer, shopping, order, cart, auction, leasing, marketplace |
| Marketing | advertisement, coupon, discount, incentive, reward, loyalty, market segmentation, product recommendation |
| Finance | investment, portfolio, trading, loan, credit risk, insurance, tax |
| Management | scheduling, task assignment, workflow, human resources, inventory, logistics, reservation, accounting, business forecasting |
| Business Sectors | health care management, patient records, real estate, legal services, travel, social networking, fundraising |
How much can one step matter? In a 2020 experiment reported by IPWatchdog using a routing prediction tool, adding a single step of billing the user’s account to an electric-vehicle charging claim moved the claim’s predicted routing from art units allowing about 82% into business method territory.
Avoid leading claims with “AI”
Class 706 covers artificial intelligence as such, including machine learning, neural networks, fuzzy logic and knowledge-based systems. Terms such as “artificial intelligence,” “machine learning model,” “neural network,” “training,” “inference” and “knowledge base” in your title, abstract and independent claims, especially claim 1, can pull your application toward class 706 and the 2120s art units that examine it. That is a risky route to take. Those art units have a median allowance rate of 62%, below the 70% median for software art units overall, and well below the 81% – 86% of the friendlier art units discussed below. They are also far more likely to reject on patent eligibility grounds. The exception is when the invention is an improvement to the AI model itself. In that case, class 706 is likely the correct classification, and changes in claim wording are unlikely to route the application elsewhere.
Use the words of the technical field your invention improves
The table below lists technical fields, terms commonly used in each field, and allowance or three-year grant rates for art units that typically examine applications in that field. These are compiled from the titles of the CPC classification groups and from published allowance and grant rates. Available statistics for the 2660s, 2650s and 2400s art units are allowance rates (allowed applications divided by allowed plus abandoned applications), and for the 2180s, 2170s and 3600s art units are three-year grant rates (applications granted within three years of the first office action).
| Target area | Words | Allowance or Grant Rate |
| Computer Vision (2660s) | image data, pixel, frame, feature extraction, object detection, segmentation, image sensor | Art Unit 2661: 86%Art Unit 2676: 81% |
| Speech and Audio (2650s) | audio signal, acoustic features, waveform, spectral, microphone | Art Unit 2657: 81% |
| Networking and Security (2400s) | packet, network node, protocol, latency, bandwidth, encryption, authentication, intrusion detection | Art Unit 2436: 82% |
| Computer Architecture (2180s) | processor, memory, cache, instruction execution, parallel processing, hardware accelerator | Art Unit 2184: 87%Art Unit 2182: 75% |
| User Interfaces (2170s) | rendering, display region, graphical element, touch input | Art Unit 2175: 86%Art Unit 2176: 86%Art Unit 2179: 70% |
| Vehicles and Navigation (3661, 3662) | vehicle control, navigation, vehicle position, route, vehicle sensor | Art Unit 3661: 87%Art Unit 3662: 77% |
| Control Systems and Robotics (3658) | signal, actuator, control loop, robot control | Art Unit 3658: 75% |
Not every technical field is a safe choice, however. If your independent claims, title, and abstract describe your invention mainly in database terms (such as index, query or data structure), your patent application can still be routed to art unit 2169, which examines database and information-retrieval inventions. One analysis reports an average allowance rate of 47% for that art unit, well below the 70% median for software art units overall. Another reports a higher three-year grant rate of 71% for the same art unit. Although these figures vary by source and method, both place art unit 2169 below most of the friendlier art units.
Match each part of the application to its role in classification.
- Title, abstract and independent claims, especially claim 1: use technical vocabulary and state the technical improvement. For example, “reducing inference latency on an edge processor by pruning attention heads,” not “AI-powered customer churn prediction.”
- Figures: lead with hardware and data-flow diagrams that show how your system works. Business process flows and customer-facing screens can still be valuable, for example, to support business-use claims, but present them in addition to the technical diagrams, not in place of them, and show how each business step is carried out by the technical components.
- Dependent claims and specification: describe the business use (such as fraud detection, pricing or health care) here, so that your claims cover your product, while recognizing that these parts generally carry less weight in classification. For example, claim 1 can recite a graphical user interface “for digitally rendering graphical representations of dynamic temporal fluctuations,” and the specification can define – and a dependent claim can specify – that the graphical representations are indicative of particular stock prices.
- Throughout the application: quantify the technical gain, such as improvements in speed, memory use, accuracy, bandwidth, or power consumption. Concrete technical results can help you address patent eligibility challenges, if raised, and help keep your application out of unwanted classification.
As patent practitioners know, proper patent application drafting improves the odds of success, but does not guarantee it. Once a patent application is filed, you have little direct control over which art unit examines it, but you can confirm where it landed and plan your prosecution strategy accordingly. If the application landed in a problematic business method art unit or a low-allowance AI art unit, plan for a longer prosecution: use examiner interviews early and prepare for a possible appeal. Wherever it lands, as a rule of thumb, expect a first Office action rejection for software and AI inventions. It is important to remember, however, that a rejection is the start of a dialogue with the examiner, not a verdict on whether your invention is patentable.
The fine print: drafting for routing is no substitute for technical substance
One last warning before you celebrate a friendly art unit.
A patent allowed in a friendly art unit still has to survive a courtroom, and the art unit has no bearing on how a court assesses validity. A claim set drafted only to dodge class 705, for example, with no real technical substance, may win the routing battle but is likely to lose the war. A thoughtful pre-filing strategy tends to serve you well both when the application is classified/routed and later in litigation. So, describe a genuine technical improvement, in technical language, amply supported by detailed technical flow diagrams. This improves your chances of being routed to a more favorable art unit and of obtaining a patent that holds up in court.
In short, the best way to avoid dealing with a problematic art unit is not to be routed there in the first place. So, choose your words before the USPTO chooses your art unit.
Contact Us
If your company is developing AI technology, or has software-related inventions and wants to evaluate whether it is patentable, or needs assistance preparing a patent application, or is stuck in the problematic art unit, experienced guidance matters. For more information about patenting this technology, contact Oleg Salakhov at Berkeley Law & Technology Group, LLP at osalakhov@bltg-ip.com or 503-439-6500.
Sources
- Lynch LLP – Software Patent Allowance Rates: What Examiner Records Show (Sept. 2026): https://lynchllp.com/insights/software-patent-allowance-rates
- Patentbots – Group 2120-2140 (Artificial Intelligence) grant rates: https://www.patentbots.com/stats/group/2120-2140
- Patentbots – Art Unit 3626: https://www.patentbots.com/stats/art-unit/3626
- Patentbots – Art Unit 2169: https://www.patentbots.com/stats/art-unit/2169
- Patentbots – Art Units 2184, 2182, 2175, 2176, 2179: https://www.patentbots.com/stats/art-unit/2184
- Patentbots – Group 3660 (Art Units 3658, 3661, 3662): https://www.patentbots.com/stats/group/3660
- Ironcrow – Art Unit prosecution statistics (2661, 2676, 2657, 2436): https://ironcrowai.com/prosecution-statistics/Art-Unit-2661
- IPWatchdog – The Most Likely Art Units for Alice Rejections (2015): https://ipwatchdog.com/2015/12/14/the-most-likely-art-units-for-alice-rejections/id=63829/
- IPWatchdog – Art Unit 3689 allowance data (2015–2016): https://www.ipwatchdog.com/tag/art-unit-3696/
- IPWatchdog – Avoid the Patent Pit of Despair: Drafting Claims Away from TC 3600 (2020): https://ipwatchdog.com/2020/06/25/avoid-patent-pit-despair-drafting-claims-away-tc-3600/
- Voice of IP – Section 101 rejections in working group 2120 (2024): https://www.voiceofip.com/p/breaking-section-101-rejections-soar
- Legal500 – USPTO AI auto-classification tool: https://www.legal500.com/intelligence/united-states/intellectual-property/the-uspto%27s-patent-classification-and-search-systems-have-jumped-on-the-ai-bandwagon
- U.S. Dept. of Commerce OIG – Patent Classification and Routing (OIG-23-026-A, 2023): https://www.oversight.gov/sites/default/files/documents/reports/2023-09/SECURED-OIG-23-026.pdf
- PatentNext – AI, Application Routing, and Consistency at the USPTO (Nov. 2025): https://www.patentnext.com/2025/11/ai-vs-tc-artificial-intelligence-application-routing-and-the-battle-for-consistency-at-the-uspto/
- GAO-25-107218 – Patent Examination and Quality Challenges (2025): https://files.gao.gov/reports/GAO-25-107218/index.html
- USPTO – MPEP § 909, Routing and Docketing of Applications: https://www.uspto.gov/web/offices/pac/mpep/s909.html
- USPTO – Class 705 definition: https://prod-web-alb-tg.uspto.gov/web/patents/classification/uspc705/defs705.pdf
- USPTO – Class 706 definition: https://www.uspto.gov/web/patents/classification/uspc706/defs706.htm
- CPC – G06Q scheme: https://www.cooperativepatentclassification.org/sites/default/files/cpc/scheme/G/scheme-G06Q.pdf