Insurance AI Patents: 85% Were Narrowed Before They Issued

Key Takeaways

• Across 135 grants whose claim counts reconcile against the issued patent, 85% had an independent claim narrowed before allowance, at a median of 71 words added.
• Allstate’s route matching patent advertises seven scoring signals in its abstract behind an “at least one of” list. Claim 1 requires none of them, only “weighting set by training data”.
• Carriers surrender very different amounts to get a patent: USAA a median 180 words per narrowed claim against UnitedHealth’s 9, a twenty-fold spread.
• Eligibility is the dominant fight. 60% drew a Section 101 rejection and abstract-idea language appears in 76 office actions.
• In 13 grants, roughly one in twelve, the examiner named the allowable claim and the applicant took the offer without argument, so the scope was set by the USPTO rather than the filer.

Allstate’s US 12,716,731, granted on 25 August 2026, is described in its own abstract as a route matching system whose machine learning model scores map link candidates using travel time, match distance, heading difference, speed conformity, road curvature, road classification and travel distance versus time delay. Seven named signals, which is the sort of detail that makes a patent readable and quotable. Claim 1, the part that is actually enforceable, requires none of them. It says the model scores each link candidate “based on weighting set by training data” and stops there.

The seven signals are real, but they sit in the abstract behind the words “at least one of”, which in claim drafting means the patentee needs one, not seven. And the limitation that actually won the patent appears in neither the abstract nor any press summary. It was inserted during prosecution, after a rejection, and reads: “reevaluate the calculated total probability scores based on contextual data and considering links associated with timestamps at least three location data points ahead and behind each respective timestamp.” Fifty eight words that were not in the application as filed.

That gap is not unique to Allstate. We pulled the complete USPTO file wrapper for all 159 AI and machine learning patents granted to insurers and insurance-adjacent firms so far in 2026, parsed 330 office actions, and reconstructed what changed in each claim between filing and issue. On the 135 grants where our claim count reconciles exactly against the issued patent, 85% had an independent claim narrowed before allowance, at a median of 71 words added. The published description and the granted claim are describing different inventions, and only one of them is property.

The claim that issues is not the claim that was filed

Prosecution is a negotiation conducted in writing, and the file wrapper records both sides of it. Across the 141 applications where the office actions parsed cleanly, the examiner raised obviousness under Section 103 in 77% of cases and eligibility under Section 101 in 60%. The median application received two office actions before allowance. Almost none of this survives into the version of a patent that reaches a trade press summary or an internal competitive intelligence deck.

What the applicant gives up to close that negotiation is measurable. Across the verified set, claim counts moved from 2,657 at first rejection to 2,555 at allowance, which sounds like a rounding error until you look at what happened inside the surviving claims. The median narrowing added 71 words to an independent claim. The interquartile range runs from 3 words to 171, so a quarter of these patents absorbed a limitation longer than most abstracts while another quarter barely moved.

This is where the actuarial consequence sits, because the direction of the error is consistent. A capability assessment built from patent abstracts will overstate what a competitor owns, and it will overstate it by more where the prosecution was hardest.

If you are pricing against an assumption that a rival holds a defensible telematics scoring method, the Allstate file says the defensible part is a specific reevaluation window of three data points ahead and behind, not the seven-signal feature set the abstract advertises. Any carrier can score on travel time and heading difference. What Allstate owns is much narrower, and a diligence process that stops at the abstract will not see the difference. The same logic runs through vendor negotiations and freedom-to-operate reviews, where the question is never what a patent describes but what its claims exclude you from doing.

Allstate is not the extreme case. USAA’s US 12,524,486, a data comparator system, absorbed 180 words into its independent claim as a single unbroken insertion, and what it absorbed is a specific control flow: determine a value difference between two values, test whether that difference “meets or breaches a difference threshold”, and only then query services to identify a reason for the difference and insert that reason into the record.

Read as a capability, this is reconciliation with exception handling, which is not novel. Read as a claim, the property is the threshold test and the conditional query that follows it. Everything upstream of that limitation was available to the applicant before prosecution and is not what they were granted.

The narrowing is also legible in a way that does not require reading legal prose. Under the amendment rules, applicants must mark inserted text with underlining, and that markup survives into the XML the USPTO publishes. In 122 of these grants we recovered the added limitation verbatim, which means the question “what did this carrier have to concede” has a documentary answer rather than an interpretive one.

What separates the carriers

Volume in this cohort is concentrated in the way the existing trackers describe. State Farm accounts for 40 of the 159 grants and UnitedHealth, once its Optum subsidiaries are counted, for 34. Allstate and USAA follow at 14 each. What volume does not tell you is how hard each of them had to fight, and on that measure the portfolios diverge sharply.

Filer	Grants	Narrowed	Median words added	Section 101 raised
State Farm	40	34 of 36	84	19 of 39
UnitedHealth	34	25 of 29	9	22 of 29
USAA	14	7 of 11	180	8 of 11
Allstate	14	8 of 12	42	4 of 12
Cotality	4	4 of 4	148	4 of 4
Hartford	4	3 of 4	36	1 of 3

Among grants where the counts reconcile, USAA gave up a median of 180 words per narrowed independent claim while UnitedHealth gave up 9. Allstate sits at 42, State Farm at 84 despite filing the largest volume. A twenty-fold spread in how much claim language a carrier surrenders is not noise at these sample sizes. It points at a real difference in filing posture: whether an applicant opens with claims drafted close to what they expect to defend, or opens broad and negotiates down.

Eligibility pressure is distributed just as unevenly. UnitedHealth drew a Section 101 rejection in 22 of 29 applications and Cotality in all four of its, against Allstate at 4 of 12. UnitedHealth is the instructive pair: the most eligibility pressure in the cohort and the least claim language surrendered, which is what a book of narrowly drafted claims looks like when the fight is about subject matter rather than prior art.

The shape of the amendment carries information as well as its size. Of the grants where we recovered an insertion, 27 arrived as a single contiguous block of new text, the signature of an applicant adding one limitation to clear one objection. Fourteen arrived as more than ten separate edits scattered through the claim, which is a rewrite rather than a concession, and those are concentrated among the largest word counts. UnitedHealth’s US 12,718,016 gained 360 words across 16 separate insertions. A patent that was rebuilt clause by clause and one that absorbed a single paragraph are different negotiations, and the word count alone conflates them.

These differences matter for anyone reading a competitor’s portfolio as a proxy for capability. A carrier that surrenders 180 words per claim is not necessarily inventing less than one that surrenders 42. It may be claiming more aggressively at the outset, or working in art that is more crowded. But the resulting property is different in kind, and a patent count treats those two outcomes as identical.

When the examiner writes the claim

The most striking pattern in the file wrappers is one that never appears in patent statistics at all. In 24 of these applications the examiner explicitly identified allowable subject matter, naming the dependent claims that would be granted if rewritten in independent form. In 13 cases the applicant simply took the offer: cancelled exactly those claims and folded their content into the independent claim, with no substantive argument anywhere in the response.

Allstate’s route matching patent is one of these. The examiner named claims 2, 3, 6 and 7 as allowable; the applicant cancelled 2 and 3 into claim 1 and filed a response whose most substantive sentence thanks the examiner for the indication. The same pattern appears at UnitedHealth twice, USAA twice, and once each at Humana, CVS Health, Cape Analytics, State Farm, Genpact and EXL.

That is roughly one in twelve of these grants where the scope of the monopoly was set by a USPTO examiner rather than argued for by the applicant. It is a reasonable commercial decision, since arguing costs money and delays issue. But it inverts the usual reading of a patent as a statement of what a company believes it invented. In these thirteen, the claim reflects what an examiner was willing to allow, and the applicant’s contribution to the boundary was agreeing to it.

Where applicants did argue, the arguments cluster in a way that tracks the rejections. Of the 80 responses we could attribute, 31 argue that the cited prior art lacks the limitation, while 35 are eligibility arguments split between claiming a practical application and claiming a technical improvement to computer functioning. That eligibility share, just under half, sits close to the 60% Section 101 rejection rate measured independently from the office actions, which is a useful internal check: the arguments carriers make match the objections they receive.

The second-order point is what this does to the idea of an AI patent portfolio as a strategic asset. A portfolio assembled largely by accepting examiner suggestions is a different thing from one assembled by argument, even where the patent counts match. The file wrapper distinguishes them. The abstract does not, and neither does any tracker built on filing volume.

Reading the record instead of the summary

None of this requires speculation about what carriers are building. The prosecution history is public, machine readable, and specific: which claims were rejected, under which statute, over which references, what the applicant added, and whether they argued or accepted. For 137 of these 159 grants we can now state either what the applicant argued or a documented reason the file contains no argument, including two where the applicant filed no remarks at all and three that were allowed without a single rejection.

The coverage is not complete, and the gaps are worth naming rather than smoothing over. In 30 of the 159 we could not recover an inserted limitation on an independent claim, either because the amendment markup was absent from the file or because the narrowing happened in dependent claims. In 22 the record is ambiguous enough that we would not attribute an argument to the applicant. Those are 52 patents where the analysis above says less than it does elsewhere, which is a different statement from saying nothing happened.

The practical discipline is narrow. When a patent matters enough to inform pricing, diligence or a build-versus-buy decision, read claim 1 and the amendment that produced it, not the abstract. The abstract is written to describe; the claim is written to exclude. On this evidence they part company in 85% of cases, and the gap is widest exactly where the stakes are highest.