Key Takeaways
- Of 161 insurance AI patents granted in 2026 whose office actions we parsed, 99 (61%) had been rejected under Section 101 before they issued. Every one of them issued anyway.
- 84 of those 99 were told in as many words that their claims were directed to an abstract idea. That language is not a signal of a doomed application.
- The cost is time. Median 13 months from first rejection to grant where eligibility was raised, against 10 months where it was not, and a median of two office actions rather than one.
- The live pipeline shows the same picture, not a worsening one: 35 of 51 pending applications with parsed grounds have drawn a Section 101 rejection, a rate statistically indistinguishable from the granted cohort.
- Six pending applications have been rejected only for double patenting, a judicially created ground with no statute number, where the examiner reads the claims as too close to the applicant’s own earlier filing.
Travelers filed US 12,711,519, “Dynamic Web Content Insertion”, in April 2023. In March 2025 the examiner rejected it under Section 101, Section 103 and Section 112, and told Travelers its claims were directed to an abstract idea. On 18 August 2026 the patent issued. That sequence is not an anomaly in this corpus. It is the median case.
What the examiner is actually saying
Section 101 asks whether an invention is the kind of thing the patent system covers at all. Since Alice Corp. v. CLS Bank in 2014, an examiner reaching for it against software will usually say the claims are “directed to an abstract idea” and then ask whether anything in them amounts to significantly more. For insurance machine learning work, which is frequently a model applied to a business process, that is the obvious line of attack, and examiners take it. The doctrinal side of that fight, including the Recentive Analytics ruling and the November 2025 USPTO guidance, is covered separately in our Section 101 reset; this piece is about what the examiners actually did.
They take it constantly. We pulled the complete USPTO file wrapper for the AI and machine learning patents granted to insurers and insurance-adjacent firms in 2026, and parsed the office actions in each. Across the 161 grants where the grounds parse, 99 drew a Section 101 rejection and 84 were explicitly told their claims were an abstract idea. All 161 are granted patents. The rejection that reads as fatal was survived by nearly two thirds of the patents that made it.
This matters because the rejection and the outcome are routinely conflated. A Section 103 rejection says somebody got there first, and the applicant answers it by narrowing the claim over the cited art. A Section 101 rejection says the subject matter may not be patentable at all, which sounds categorical, and in the popular reading of Alice it was treated that way for years. The file wrappers say otherwise. Among patents that issued, the presence of an eligibility rejection carries almost no information about the destination.
The cost is time, not the patent
What eligibility pressure does buy is delay. Measuring from the first rejection to the grant date, patents that faced a Section 101 challenge took a median of 13 months, against 10 months for those that did not. The median count of office actions moves from one to two. That is about three extra months and one extra round of prosecution to answer the eligibility question, and it lands on a majority of filings in this class.
| Ground raised | Granted 2026 (n=161) | Pending now (n=51) |
|---|---|---|
| Section 101, eligibility | 61% | 69% |
| Section 103, obviousness | 78% | 76% |
| Section 102, anticipation | 29% | 29% |
| Section 112, definiteness | 30% | 20% |
| Abstract-idea language | 53% | 57% |
For anyone valuing a competitor’s AI portfolio, or their own, that reframes what an eligibility rejection is worth as a signal. Treating a live Section 101 rejection as an impairment indicator will misclassify most of the filings it touches, because the base rate among eventual grants is 61%. The defensible read is closer to a scheduling adjustment: an application under eligibility challenge should be expected to sit about three months longer and absorb one more office action before it resolves, and a portfolio weighted toward such filings converts to enforceable property more slowly rather than less often.
The corollary is that the rare signals are the informative ones. An examiner naming allowable subject matter, which is the examiner identifying the dependent claim carrying the limitation that would earn the patent, appears in 16% of the granted cohort and in only three of the pending applications we can currently read. That is a roadmap when it appears, and its absence is the ordinary condition rather than a warning.
What the pipeline cannot tell us yet
The comparison has a survivorship problem that has to be stated plainly. The granted cohort is, by construction, made of survivors: every one of those 161 patents issued, so measuring the eligibility rate among them says what rejections issued patents overcame, not what share of all rejected applications ever issue. Applications abandoned after a Section 101 rejection never enter that denominator. The honest claim is the narrow one: a Section 101 rejection is common among patents that issue, so it cannot on its own predict that a given application will not.
The pending side is a different sample rather than a control. Of 179 published applications in our live pipeline, 65 have reached first action and 51 have grounds we can parse. The rates are close to the granted cohort across every statute, which is what we would expect if examiner behaviour is stable, but these applications have not resolved and some will be abandoned. Only the passage of time will attach outcomes to them, and we intend to publish that follow-through as it lands.
Two limits on the reading are worth naming. Fifty one pending applications is a small sample, and per-filer splits within it are smaller still, so we have not broken the pending rates down by carrier here. And our parser reads rejection grounds from the examiner’s own sentences: six pending applications returned an office action with no statute at all, which turned out in every case to be a nonstatutory double patenting rejection, a judicially created ground that carries no section number. That is a real rejection rather than a parser miss, and it is the sort of thing that only becomes visible when the file wrapper is read rather than the abstract.
Sources
- USPTO Patent Center, Travelers, US 12,711,519, “Dynamic Web Content Insertion”, filed April 27, 2023, first rejection March 21, 2025 under Sections 101, 103 and 112, granted August 18, 2026
- USPTO Open Data Portal, patent file wrappers for the 193 insurance AI grants issued in 2026, office actions parsed for statutory grounds (retrieved August 2026)
- USPTO Open Data Portal, file wrappers for 179 published applications still in examination across 30 insurance and insurance-adjacent filers (retrieved August 2026)
- MPEP 2106, subject matter eligibility, and the two-step Alice framework applied to abstract-idea rejections
- Alice Corp. v. CLS Bank International, 573 U.S. 208 (2014)
Updated 29 August 2026. This piece first ran against the 159 grants that carried CPC G06N, the classification the AI Patent Watch tracker used at the time. That definition was widened the same week to include grants whose title names a machine learning technique, because G06N was missing most generative-era work, and the 2026 cohort grew to 193 with 161 parsing a statutory ground. Every figure here is recomputed on the wider cohort. The conclusion is unchanged and two numbers moved: the Section 101 rate from 63% to 61%, and the delay a challenge costs from about five months to about three.
Grounds are parsed from the examiner’s own office action text. Percentages are computed over the 161 granted and 51 pending applications where a statutory ground parses, not over the full corpus. Live figures for the pending pipeline are published at AI Patent Watch.