DETAILED ACTION
Introduction
Claims 1-7, 9-15, and 17-20 are pending. Claims 1, 9, and 17 are amended. Claims 8 and 16 are cancelled. No claims are added. This Office action is in response to Applicant’s request for continued examination (RCE) filed on 2/20/2026.
Other Relevant Prior Art
Kumar (US 9,313,162) teaches predicting a user action based on content of one or more email messages.
Moon (US 2021/0157618) teaches predicting a user action based on messages exchanged with a chatbot.
Martel (US 2019/0220245) teaches most of the features of claims 1, 9, and 17, as discussed below.
Response to Arguments
Examiner discusses the arguments of Applicant’s representative below.
Rejection of claims 1, 9, and 17 under 35 U.S.C. 103
Applicant’s representative has amended claims 1, 9, and 17 to recite steps of constructing a prompt containing the content of the message and metadata of the message, and inputting the prompt into the LLM. Applicant’s representative now argues that the combination of Jothilingam and Dan/Desmond does not teach the system of claims 1, 9, and 17, as amended. In support of this argument, Applicant’s representative raises four additional arguments.
First Argument
Applicant’s representative argues that “Dan’s prompt construction, at best, merely enables an AI participant to respond to queries within a group messaging thread, not, as claimed, analyzing a single user’s electronic message and its associated metadata to predict and suggest a real-world action that the system then performs on behalf of the user.” However, Examiner respectfully disagrees because this argument attacks Dan in isolation, but the rejection relies on the combination, with Jothilingam as the primary reference and Dan only as the secondary reference for the LLM/prompt-construction aspect. Jothilingam already teaches analyzing an individual electronic communication, augmenting it with associated information such as requester/requestee, location, and time/date, inferring the semantics of the request or commitment, generating task-oriented actions, presenting those actions to the user, receiving the user’s edits or selections, and then performing the selected task-oriented processes. See Jothilingam, par. 17, 21-22, 58, 60, 73, and 79-80.
Thus, the premise that Dan does not “analyze a single user’s electronic message and its associated metadata to predict and suggest a real-world action that the system then performs on behalf of the user” does not rebut the combination, because Dan is not relied on for that entire chain. Jothilingam supplies the single-message/task-management workflow. For example, the Jothilingam discloses extracting tasks from messages or email exchanges, augmenting extracted task content with requester/requestee, locations, and times/dates, analyzing the communication to infer semantics, and using those inferences to drive reminders, to-do-list revisions, appointments, meeting requests, and other time-management activities. See Jothilingam, par. 17, 21-22, 58, and 60. It also discloses presenting a list of task-oriented actions to the user, allowing the user to select, refine, delete, or add actions, and then performing the approved task-oriented processes. See Jothilingam, par. 73 and 79-80.
The argument also understates the teachings of Dan because Dan is not limited to merely answering queries within a group thread. Dan expressly says the generative model can generate output based on the messages, that the messaging application constructs a prompt including message content and identities of users, and that the generative model can do more than answer questions—for example, it can propose follow-up actions, assist with scheduling events, and assist with coordinating schedules. See Dan, par. 7-9. Dan likewise teaches that the model is a transformer-based GLM that receives text input and generates text output. See Dan, par. 3, 25. That is enough to teach the secondary point for which Dan is being used: packaging message content and associated information into a prompt for a generative model that returns natural-language output.
Accordingly, a person of ordinary skill would have understood Dan’s prompt-construction technique to be readily applicable to Jothilingam’s single-message setting. Jothilingam already has the relevant input materials—email/message text plus associated fields such as requester/requestee, location, and times/dates—and already uses those materials to infer task semantics and generate task-oriented actions. Dan simply provides a known generative-model implementation for doing so via a prompt and GLM output. Using Dan’s prompting framework in the system of Jothilingam would therefore have been a predictable substitution of one known analysis/generation mechanism for another, not an improper attempt to convert Dan’s entire group-conversation system into the claimed invention.
Second Argument
Applicant’s representative argues that “Dan does not describe or suggest identifying a single electronic message authored by a user, packaging the content of that message together with its metadata into a prompt, and providing that prompt to an LLM so that the LLM can determine semantic meaning and predict a potential action for the user, as recited in the amended claims.” However, Examiner respectfully disagrees for reasons similar to those presented in the preceding paragraphs. Specifically, this argument is not persuasive because it attacks Dan alone rather than the teachings of the combination. Jothilingam already teaches identifying task content from an individual electronic communication such as an email or message, augmenting that content with associated information such as requester/requestee, locations, and times/dates, analyzing the communication to infer semantics of the request or commitment, generating task-oriented actions, presenting those actions to the user, receiving user edits or selections, and performing the approved task-oriented processes. See Jothilingam, par. 17, 21-22, 58, 60, 73, and 79-80. Dan is relied on for the additional teaching of constructing a prompt for a generative model from message content and associated information, including message content, timestamps, user identities, requesting-user identity, and optional location, and providing that prompt to the generative model to generate output. See Dan, par. 8-9, 22, 33-35, 38 and 58. Thus, the combination teaches packaging Jothilingam’s single-message content and associated metadata/information into a prompt for a generative model so that the model can be used in the Jothilingam’s task-analysis and action-suggestion workflow.
Third Argument
Applicant’s representative argues that “Desmond does not describe, teach or suggest constructing a prompt that includes the content of an electronic message together with its metadata fields and providing that prompt to an LLM to predict potential user actions based on the semantic meaning of that message. Desmond’s system moderates when an AI agent should speak in a group dialog; it does not analyze a user’s outgoing or incoming electronic message to suggest and then perform real-world actions on behalf of that user.” However, Examiner no longer relies on Desmond to reject any of the claims. Therefore, this argument is moot.
Fourth Argument
Lastly, “The proposed combination of Jothilingam with either Dan or Desmond lacks a proper motivation.” However, Examiner respectfully disagrees. Nonetheless, Examiner would like to take this opportunity to provide Applicant’s representative with clarifications regarding the motivation for modifying the system of Jothilingam to incorporate the teachings of Dan. Specifically, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the system of Jothilingam so that the step of inputting the content and the at least one metadata field into the artificial intelligence component comprises constructing a prompt for a generative model (the prompt including content of the message and metadata) and inputting the prompt into the generative model, and so that the description of the task-oriented actions comprises a natural language description of the task-oriented actions output by the generative model, because replacing Jothilingam’s learned/predictive tasks-analysis with the prompt-based LLM framework of Dan would provide a more flexible and context-aware way to analyze message content, extract task-oriented actions, and present those actions back to the user. For instance:
Jothilingam already seeks to determine the meaning of message content and metadata, and to infer desired actions/tasks from that information. See 22, 58, and 60. One benefit of modifying the system of Jothilingam to use the LLM approach of Dan is that the modified system could perform the primary reference’s semantic analysis using a prompt-driven LLM, which would be expected to handle a wider variety of natural-language phrasings and less structured message content.
Jothilingam already teaches presenting suggestions or lists of task-oriented actions to the user for review. See par. 73, 79. Dan teach that the corresponding responses generated by the LLM are concise natural-language outputs such as a single insightful sentence or other appropriate responses. A clear benefit, then, is that the modified system could present Jothilingam’s task-oriented actions to the user in better natural language form, rather than as merely extracted labels or rigid suggestions that might be more difficult to understand by the user.
Claim Rejections: 35 U.S.C. 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1-3, 5, 9-11, 13, and 17-19 are rejected under 35 U.S.C. 103 because they are unpatentable over Jothilingam (US 2017/0193349) in view of Dan (US 2024/0305589).1
Regarding claims 1, 9, and 17, Jothilingam teaches a method comprising: identifying, by a processor, an electronic message that comprises content written by a user and at least one metadata field (The system identifies a task in a message composed by a user, which presupposes a step of identifying the message. See par. 77. The message may include metadata such as a subject header, length of the message, a date/time the message was sent, and information on the sender and recipients of the message. See par. 45); predicting, by the processor, a potential action for the user by inputting the content and the at least one metadata field into an artificial intelligence component, wherein the artificial intelligence component analyzes the content and the at least one metadata field together to determine semantic meaning of the electronic message(The system analyzes content of the message and the metadata to determine one or more “meanings” of the content, where the term “meaning” is defined as “how one would interpret the content in a natural language.” See par. 58. A predictive model may be used to infer that an outgoing or incoming email contains a task. See par. 60), wherein the artificial intelligence component is configured to output one or more specific actions for the user to take based on the electronic message and the at least one metadata field (The system extracts one or more tasks from the message based on the analysis. See par. 58. The system derives one or more task-oriented actions based on the tasks. See par. 78);suggesting, by the processor, the potential action to the user by displaying a description of the potential action in a user interface (The system outputs the one or more task-oriented actions to a user interface for inspection or review by the user. See par. 79); receiving, by the processor, user input related to the potential action (The system receives a user input related to the one or more task-oriented actions, such as a user input selecting a task-oriented action to be performed by the system on behalf of the user, a user input refining one or more task-oriented actions, a user input deleting one or more task-oriented actions, or a user input manually adding one or more task-oriented actions. See par. 79); and performing, by the processor, on behalf of the user, a subsequent action conforming to the user input (The system performs an action corresponding to the user input, such as performing a task-oriented action selected by the user. See par. 79).
However, Jothilingam does not teach that the step of inputting the content and the at least one metadata field into the artificial intelligence component comprises constructing, by the processor, a prompt for a large language model, the prompt comprising the content of the electronic message and the at least one metadata field, and inputting the prompt into the large language model. In addition, Jothilingam does not teach that the description of the potential action is a natural language description of the potential action generated via the large language model. Nonetheless, Dan teaches a method comprising steps of: receiving a message (See par. 32); generating, by a prompt generator module, a prompt comprising content and metadata of a message, such as a timestamp associated with a message and the identity of the user who generated the message (See par. 33, 35, 38; fig. 4); inputting the prompt into a generative model (i.e., LLM. See par. 38); predicting, by the LLM, a potential action by analyzing the semantic meaning of the message and the metadata (See par. 38); and displaying a natural language description of the potential action generated by the LLM in a display of a user interface (See par. 38; fig. 2).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the system of Jothilingam so that the step of inputting the content and the at least one metadata field into the artificial intelligence component comprises constructing a prompt for a generative model (the prompt including content of the message and metadata) and inputting the prompt into the generative model, and so that the description of the task-oriented actions comprises a natural language description of the task-oriented actions output by the generative model, because replacing Jothilingam’s learned/predictive tasks-analysis with the prompt-based LLM framework of Dan would provide a more flexible and context-aware way to analyze message content, extract task-oriented actions, and present those actions back to the user. For instance:
Jothilingam already seeks to determine the meaning of message content and metadata, and to infer desired actions/tasks from that information. See 22, 58, and 60. One benefit of modifying the system of Jothilingam to use the LLM approach of Dan is that the modified system could perform the primary reference’s semantic analysis using a prompt-driven LLM, which would be expected to handle a wider variety of natural-language phrasings and less structured message content.
Jothilingam already teaches presenting suggestions or lists of task-oriented actions to the user for review. See par. 73, 79. Dan teach that the corresponding responses generated by the LLM are concise natural-language outputs such as a single insightful sentence or other appropriate responses. A clear benefit, then, is that the modified system could present Jothilingam’s task-oriented actions to the user in better natural language form, rather than as merely extracted labels or rigid suggestions that might be more difficult to understand by the user.
Regarding claims 2, 10, and 18, Jothilingam and Dan teach the method of claim 1, wherein identifying the electronic messages comprises identifying an electronic message sent by the user (Jothilingam teaches that the system can identify a task in a message that has already been sent by the user, as well as a task in a message that is in the process of being drafted by the user. See par. 22-23).
Regarding claims 3, 11, and 19, Jothilingam and Dan teach the method of claim 1, wherein identifying the electronic messages comprises identifying an electronic message being composed by the user that has not yet been sent (Jothilingam teaches that the system can identify a task in a message that has already been sent by the user, as well as a task in a message that is in the process of being drafted by the user. See par. 22-23).
Regarding claims 5 and 13, Jothilingam and Dan teach the method of claim 1, wherein performing the subsequent action comprises creating a digital calendar event on behalf of the user (Jothilingam teaches that the task-oriented actions may include updating a calendar on behalf of the user. See par. 78).
Claims 4, 12, and 20 are rejected under 35 U.S.C. 103 because they are unpatentable over Jothilingam and Dan, as applied to claims 1, 9, and 17 above, in further view of Menon (US 11,855,934).
Regarding claims 4, 12, and 20, Jothilingam and Dan teach the method of claim 1, wherein receiving user input related to the potential action comprises receiving a modification of the potential action provided by the user and generating a new potential action, wherein performing the subsequent action comprises performing the new potential action (Jothilingam teaches that the user can modify the suggested actions, which causes the system to regenerate the task-oriented processes to potentially include new suggested actions. See par. 79). However, Jothilingam and Dan do not teach that the modification is received via a chat interface. Nonetheless, Menon teaches a method of training an LLM whereby the LLM receives a prompt from a user via a chat interface, whereby the LLM generates a first recommendation for the user based on the user prompt, whereby the LLM receives feedback from the user via the chat interface, and whereby the LLM provides a second recommendation for the user based on the feedback. See col. 14, ln. 24-67.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the system of Jothilingam and Dan so that the LLM receives a modification to the one or more task-oriented actions via a chat interface with the LLM because doing so allows the user to train the LLM with feedback regarding the one or more task-oriented actions generated by the LLM and thereby receive more accurate task-oriented actions in the future.
Claims 6 and 14 are rejected under 35 U.S.C. 103 because they are unpatentable over Jothilingam and Dan, as applied to claims 1 and 9 above, in further view of Willet (US 2017/0149703).
Regarding claims 6 and 14, Jothilingam and Dan does not teach wherein performing the subsequent action comprises sending an additional electronic message on behalf of the user. However, Willet teaches a messaging system whereby the system suggests sending a message on behalf of the user in response to natural language processing of a message, and sends the message upon receiving approval from the user. See par. 47; fig. 9, item 906.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the system of Jothilingam and Dan so that the system sends a message on behalf of the user because doing so allows the system to assist the user with sending messages.
Claims 7 and 15 are rejected under 35 U.S.C. 103 because they are unpatentable over Jothilingam and Dan, as applied to claims 1 and 9 above, in further view of Fenton (US 2021/0073293).
Regarding claim 7 and 15, Jothilingam and Dan do not teach the method of claim 1, wherein performing the subsequent action comprises creating additional content and adding the additional content to the electronic message. However, Fenton teaches a messaging system whereby the system analyzes the content of a message that a user is drafting to predict the intent of the user, whereby the system suggests a task to be performed based on the analysis, whereby the system performs the task on behalf of the user, and whereby the task may include adding content to the electronic message on behalf of the user. See par. 40, 159, 164; fig. 5D.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the system of Jothilingam and Dan so that the system creates supplemental content and adds the supplemental content to the message, because doing so allows the system to assist the user with sending messages.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Andrew Georgandellis whose telephone number is 571-270-3991. The examiner can normally be reached on Monday through Friday, 7:30-5:00 PM EST. If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Tonia Dollinger, can be reached on 571-272-4170. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of an application may be obtained from the Patent Application Information Retrieval (PAIR) system. Status information for published applications may be obtained from either Private PAIR or Public PAIR. Status information for unpublished applications is available through Private PAIR only. For more information about the PAIR system, see http://pair-direct.uspto.gov. Should you have questions on access to the Private PAIR system, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative or access to the automated information system, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/ANDREW C GEORGANDELLIS/Primary Examiner, Art Unit 2459
1 Martel (US 2019/0220245) teaches the same features of claims 1, 9, and 17 that Jothilingam teaches. For instance, Martel teaches identifying, by a processor, an electronic message that comprises content written by a user and at least one metadata field (The system receives “dialog content” from a user and contextual information associated with the dialog content. See fig. 8A, step 804; par. 75); predicting, by the processor, a potential action for the user by inputting the content and the at least one metadata field into a processing component, wherein the processing component is configured to output one or more specific actions for the user to take based on the electronic message and the at least one metadata field (An assistance engine uses the dialog content and the context information to determine actions to perform for the user. See fig. 8A, step 814); suggesting, by the processor, the potential action to the user (The system presents indicia of the actions. See par. 267, 269); receiving, by the processor, user input related to the potential action (A user may select an action. See par. 271); and performing, by the processor, on behalf of the user, a subsequent action conforming to the user input (After the user selects an action, the system performs the action. See par. 271).
Prosecution Insights