DETAILED ACTION
1. The present application, filed on or after March 13, 2013, is being examined under the first inventor to file provisions of the AIA .
Response to Amendment
2.. An RCE with accompanying Amendment was filed May 20, 2026 (hereinafter “Amendment”) and has been entered into the record and fully considered. The Amendment was filed in response to a Final Rejection dated February 20, 2026.
Despite the Amendment to the Claims and Applicant’s remarks, the Rejections under §101 and §103 as set forth in the Non-Final Rejection are hereby maintained. However, the Rejections under §103 to the Claims are on NEW GROUNDS necessitated by the Amendment.
An explanation of the maintained Rejections and a response to Applicant’s arguments are set forth below. Please see the “Conclusion” section of this Action below for important information regarding responding to this Action.
NOTE: While the interview of April 17, 2026 was helpful in advancing prosecution, additional specificity to the Claims is necessary under §101, and further distinguishing subject matter under §103. A follow up interview is recommended. Please use the AIR form, the link for which is found in the Conclusion section of this Action, for scheduling an interview.
Status of the Claims:
Claims 1, 3 – 9, 11 – 16, 19, and 21 are pending in this Application.
The independent Claims 1, 16 and 19 have been amended to be structurally and substantively similar to one another. Accordingly, the new grounds of rejection under §103 set forth in the Final Rejection is now applicable to ALL of the Claims.
The dependent Claims were not amended or only amended in a trivial manner.
Therefore, the following explanation of the maintained rejections with regard to Claims 1 is considered explanatory of the Rejection as a whole.
With regard to the Amendment:
Claim 1 was amended as follows:
PNG
media_image1.png
642
688
media_image1.png
Greyscale
PNG
media_image2.png
345
665
media_image2.png
Greyscale
Summary of the Amendment and Broadest Reasonable Interpretation:
Claim terminology is to be given its plain and ordinary meaning to a person of ordinary skill in the art, consistent with the specification. This is true, unless the terms are given a special meaning. See MPEP §2111.01
Here, no special meaning is detected. As noted in the Amendment, the
changes to Claim 1 relate generally to:
Adding a first and second large language model (LLM) applied in sequence
The first LLM receives training data as an input and generates “characteristics” found in historical travel analyst notes to identify “discrepancies.”
The output of the first LLM is a narrative that is fed into a second LLM to detect fraud.
These terms appear – subject to further consideration – to be defined in the specification based on their plain and ordinary meaning.
With regard to §101:
The Claim now adds some specificity in terms of the training of the model and the fact that both a first and second LLM model is deployed in sequence. However, these terms are recited at an extremely high level.
The first LLM, as recited in Claim 1, merely takes in “analyst notes.” It is not clear whether this input data is structured or unstructured. On the other hand, the data is defined as “classified.” How was such “classification” accomplished? By the first LLM? What are examples or parameters pertaining to the “characteristics” found in the data, and more particularly, what are the “discrepancies” identified. More importantly, “how” are the discrepancies identified by the first model? “How” are these models themselves trained? How are they fine-tuned? It should be noted that the use of sequential models – including LLM models – is very common. This process is often referred to as “LLM ensembles” or “cascading LLM’s.”
Accordingly, without greater specificity as to these claim features, the claims remain in an “apply it” condition with respect to the abstract idea of detecting fraud in travel-related transactions. Without greater specificity as to “how” certain functions solve a technical problem, the currently recited limitations can be achieved by any general purpose computer without special programming. In short, each step does no more than require a generic computer to perform generic computer functions. Considered as an ordered combination, the computer components of the Claim add nothing that is not already present when the steps are considered separately.
Accordingly, the Rejection of Claim 1 under §101 must be maintained,
With regard to §103:
With regard to Claim 1, a NEW GROUNDS of rejection is asserted necessitated by the Amendment.
Claims 1, 3 – 9, 11 – 16, 19, and 221 are rejected under 35 U.S.C. §103 as being unpatentable over U.S. Patent Publication No. 2021/0182830 to Edwards et al. (hereinafter “Edwards) in view of U.S. Patent Publication No. 2025/0045735 to Araujo et al. (hereinafter “Araujo”) and further in view of Non-Patent Literature to anonymous, “ChatGPT -AI Chat bot – A Complete Guide,” The Encrypt, December 2022 (hereinafter “ChatGPT”) and still further in view of the U.S. Patent Publication No. 2025/0278734 to Singh et al. (hereinafter “Singh”).
Singh is directly on point with the use of sequential models in fraud detection.
The title is: Systems and methods for utilizing artificial intelligence models to prevent fraudulent activities
The Abstract reads as follows:
A device may receive input data identifying input features associated with a user, and may process the input data, with an initial model, to predict one or more fraud stages for the user. The device may identify one or more sets of models from a plurality of models and based on the one or more fraud stages, and may process the input data, with the one or more sets of models, to determine one or more fraud parameters associated with the user. The device may identify a fraudulent activity associated with the user based on the fraud parameters, and may utilize a large language model to generate steps to resolve the fraudulent activity based on historical fraud resolutions. The device may provide the steps to resolve the fraudulent activity to a representative or the user for implementation.” (Emphasis Added)
It is respectfully submitted that the “fraud parameters” teachings of Singh are tantamount to the characteristics of travel-related transactions that may be associated with fraud. In addition, such fraud parameters identify “discrepancies” – i.e. anomalous behavior – that is likely to indicate fraud.
Furthermore, Singh is on point with respect to the sequence of model training as recited in the Claim:
“[0014] As shown in FIG. 1A, and by reference number 115, the security system 105 may receive input data identifying input features associated with a plurality of users of a network. For example, a network (e.g., a telecommunications network, an Internet service provider network, and/or the like) may provide a variety of features for a plurality of users of the network, such as text features, email features, call features, identification proof features, call or chat transcripts, transaction features, device features, account features, and/or the like. The network may generate the variety of features and may continuously and/or periodically store the variety of features in the data structure 110 as the input data identifying the input features associated with the plurality of users of the network.” (Emphasis Added)
It is submitted that a person of ordinary skill in the art would readily recognize that the text and other text-based features received could also be the analyst notes as recited in the Claim. Furthermore, the “initial model” is used to generate a fraud-based analysis of these features:
“[0018] As further shown in FIG. 1A, and by reference number 120, the security system 105 may process the input data for a user, with an initial model, to predict one or more fraud stages for the user. For example, the security system 105 may analyze the input data for a user of the plurality of users to determine whether the user is associated with fraudulent activities. The security system 105 may be associated with an initial model that is a machine learning model, such as a neural network model with an embedding layer, long short-term memory (LSTM) layers, a dense layer, and/or the like. The security system 105 may utilize the initial model to determine one or more fraud stages for the user based on the input data. For example, the user may be associated with a first fraud stage indicating identity theft of the user; a second fraud stage indicating an account takeover for the user (e.g., based on updated account information); a third fraud stage indicating fraud associated with adding and/or updating services and/or accounts of the user, subscriber identity module (SIM) swapping, unapproved purchase of devices, adding a new customer without user authorization, illegitimate accessing of user accounts or performance of transactions, the user being unable to access an account, etc.; and/or the like. In some implementations, the user may be associated with one or more of the first fraud stage, the second fraud stage, the third fraud stage, and/or the like.” (Emphasis Added)
In Singh, one of a plurality of subsequent models is then applied to sequentially process the input data:
“[0020] As shown in FIG. 1C, and by reference number 130, the security system 105 may process the input data for the user, with the one or more sets of models, to determine one or more fraud parameters associated with the user. For example, the security system 105 may utilize the one or more sets of models to determine the one more fraud patterns associated with the user and based on the input data. In some implementations, when processing the input data, with the one or more sets of models, to determine the one or more fraud parameters, the security system 105 may sequentially process the input data, with the one or more sets of models, to determine the one or more fraud parameters associated with the user, or may process the input data, in parallel with the one or more sets of models, to determine the one or more fraud parameters associated with the user.” (Emphasis Added)
This sequence or ensemble of models is illustrated in Fig. 1B:
PNG
media_image3.png
396
780
media_image3.png
Greyscale
Finally, a person of ordinary skill in the art would readily recognize from Singh that an ensemble of LLM’s is used for this purpose, with the well-known output being narrative data in natural language format:
“[0024] As shown in FIG. 1E, and by reference number 140, the security system 105 may utilize a large language model to generate steps to resolve the fraudulent activity based on historical fraud resolutions. For example, the security system 105 may be associated with a large language model. The large language model may have access to a data structure that includes historical fraud resolutions generated for historical fraudulent activities. The large language model may receive the historical fraud resolutions from the data structure, and may process the fraudulent activity and the historical fraud resolutions to generate the steps to resolve the fraudulent activity. In some implementations, the large language model may compare the fraudulent activity and the historical fraudulent activities to determine one of the historical fraudulent activities that is most similar to the fraudulent activity. The large language model may identify one of the historical fraud resolutions that corresponds to the one of the historical fraudulent activities. The large language model utilizes the historical steps utilizing in the one of the historical fraud resolutions to generate the steps to resolve the fraudulent activity. In one example, the steps to resolve the fraudulent activity may include updating credentials and access of the user (e.g., when the fraudulent activity is associated with the first fraud stage); updating credentials and access of the user and reviewing an account of the user (e.g., when the fraudulent activity is associated with the second fraud stage); reviewing an order of the user, a pick up location of the user, and device options of the user (e.g., when the fraudulent activity is associated with the second fraud stage); causing the user to immediately contact a fraud department (e.g., when the fraudulent activity is associated with the third fraud stage); and/or the like.” (Emphasis Added)
Therefore, it would have been obvious to one of ordinary skill in the relevant art at the time of filing the claimed invention to have modified the combined ML fraud detection system of Edwards in view of Araujo and further in view of ChatGPT, which teaches an LLM narrative output, to add the sequential model teachings of Singh. The motivation to do so comes from Edwards. As quoted above and in the Final Rejection, Edwards teaches the use of neural networks. An LLM is a form of neural network. It would greatly enhance the efficiency and accuracy of the system of Edwards to use the LLM teachings of Singh.
Therefore, the Rejection of these Claims is also maintained.
Response to Arguments
3. Applicant's arguments set forth in the Remarks section of the Amendment have been fully considered but they are not persuasive.
With regard to section 101 rejection, Applicant argues as follows:
PNG
media_image4.png
617
633
media_image4.png
Greyscale
The shortcoming in this argument is that these features – while important to detecting fraud – are not recited nor reflected in the Claims.
Applicant further argues:
PNG
media_image5.png
412
667
media_image5.png
Greyscale
However, improving manual review for a human – merely recited as a broad concept - is NOT a technical problem. It is important that the Claim reflect the features noted by the Applicant, so long as they are recited with specificity. It is likely that the nuances and specifics of travel should be recited in the Claim.
Therefore, an interview is suggested.
With regard to §103:
Applicant’s arguments with respect to the Claims are moot in view of the new grounds of Rejection.
Conclusion
4. Applicant should carefully consider the following in connection with this Office Action:
A. Search and Prior Art
The search conducted in connection with this Office Action, as well as any previous Actions, encompassed the inventive concepts as defined in the Applicant’s specification. That is, the search(es) included concepts and features which are defined by the pending claims but also pertinent to significant although unclaimed subject matter. Accordingly, such search(es) were directed to the defined invention as well as the general state of the art, including references which are in the same field of endeavor as the present application as well as related fields (e.g. use of machine learning to detect fraud.). Indeed, there is a plethora of prior art in these fields.
Therefore, in addition to prior art references cited and applied in connection with this and any previous Office Actions, the following prior art is also made of record but not relied upon in the current rejection:
U.S. Patent Publication No. 2025/0272477 to Bhan. This reference relates to the concept of ensemble LLM’s for content summarization.
U.S. Patent Publication No. 2025/0390675 to Sharma et al. This reference relates to the concept of cascading ML models.
B. Responding to this Office Action
In view of the foregoing explanation of the scope of searches conducted in connection with the examination of this application, in preparing any response to this Action, Applicant is encouraged to carefully review the entire disclosures of the above-cited, unapplied references, as well as any previously cited references. It is likely that one or more such references disclose or suggest features which Applicant may seek to claim. Moreover, for the same reasons, Applicant is encouraged to review the entire disclosures of the references applied in the foregoing rejections and not just the sections mentioned.
C. Interviews and Compact Prosecution
The Office strongly encourages interviews as an important aspect of compact prosecution. Statistics and studies have shown that prosecution can be greatly advanced by way of interviews. Indeed, in many instances, during the course of one or more interviews, the Examiner and Applicant may reach an agreement on eligible and allowable subject matter that is supported by the specification.
Interviews are especially welcomed by this examiner at any stage of the prosecution process. Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool (e.g. TEAMS). To facilitate the scheduling of an interview, the Examiner requests either a phone call at the number set forth below or the use of the AIR form as follows:
USPTO Automated Interview Request http://www.uspto.gov/interviewpractice.
Other forms of interview requests filed in this application may result in a delay in scheduling the interview because of the time required to appear on the Examiner's docket. Thus, the use of the AIR form is strongly encouraged.
D. Communicating with the Office
Any inquiry concerning this communication or earlier communications from the examiner should be directed to WILLIAM BUNKER whose telephone number is (571)272-0017. The examiner can normally be reached on M - F 8:30AM - 5:30PM, Pacific.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Abhishek Vyas, can be reached at 571-270-1836. Information regarding the status of an application, whether published or unpublished, may be obtained from the “Patent Center” system. For more information about the Patent Center system, see https://patentcenter.uspto.gov/
/William (Bill) Bunker/
U.S. Patent Examiner
AU 3691
(571) 272-0017 - office
william.bunker@uspto.gov
June 13, 2026
/ABHISHEK VYAS/Supervisory Patent Examiner, Art Unit 3691
Prosecution Insights