Dialect LLC v. Salesforce Inc

I. Executive Summary and Procedural Information

• Parties & Counsel:
  • Plaintiff: Dialect, LLC (Texas)
  • Defendant: Salesforce, Inc. (Delaware)
  • Plaintiff’s Counsel: Blue Peak Law Group LLP; Scott Douglass McConnico LLP
• Case Identification: 7:25-cv-00061, W.D. Tex., 05/06/2025
• Venue Allegations: Venue is alleged to be proper in the Western District of Texas because Defendant Salesforce maintains a regular and established physical place of business at an office in Austin, Texas.
• Core Dispute: Plaintiff alleges that Defendant’s Salesforce Voice AI Agent infringes three patents related to conversational voice interfaces and natural language understanding technology.
• Technical Context: The technology at issue involves systems that enable more natural, human-like spoken interaction with computers by interpreting context, user history, and domain-specific knowledge.
• Key Procedural History: The patents-in-suit originated with VoiceBox Technologies, Inc., a company described as a pioneer in natural language understanding. The complaint notes that in April 2024, Google filed petitions for inter partes review (IPR) against two of the asserted patents ('006 and '209). In October 2024, the Patent Trial and Appeal Board (PTAB) denied institution of those IPRs, a procedural outcome that prevents a formal trial on the patents' validity before that tribunal.

Case Timeline

Date	Event
2002-06-03	Priority Date for ’209 Patent
2002-06-03	Priority Date for ’006 Patent
2005-08-10	Priority Date for ’659 Patent
2008-07-08	’209 Patent Issued
2011-09-06	’006 Patent Issued
2013-12-31	’659 Patent Issued
2024-04-XX	Google files petition for inter partes review of the ’006 Patent
2024-04-XX	Google files petition for inter partes review of the ’209 Patent
2024-09-13	Salesforce acquires Tenyx
2024-10-XX	PTAB denies institution of inter partes review of the ’006 Patent
2024-10-XX	PTAB denies institution of inter partes review of the ’209 Patent
2024-XX-XX	Salesforce Voice AI Agent introduced
2025-05-06	Complaint Filed

II. Technology and Patent(s)-in-Suit Analysis

U.S. Patent No. 8,015,006 - "Systems And Methods For Processing Natural Language Speech Utterances With Context-Specific Domain Agents"

• Patent Identification: U.S. Patent No. 8,015,006, "Systems And Methods For Processing Natural Language Speech Utterances With Context-Specific Domain Agents," issued on September 6, 2011 (Compl. ¶29).

The Invention Explained

• Problem Addressed: The patent’s background describes the difficulty of creating natural human-machine communication because human language is often ambiguous and reliant on context, whereas traditional machine queries were highly structured and rigid (’006 Patent, col. 1:33-41). Existing systems were described as "generally unable to provide a complete environment for users to make natural language speech queries" (’006 Patent, col. 1:50-53).
• The Patented Solution: The invention proposes a system that makes "significant use of context, prior information, domain knowledge, and user specific profile data" to better understand a user's intent (’006 Patent, Abstract). A key architectural element is the use of "domain agents," which are software modules that organize behavior and information for specific topics (e.g., weather, directions) (’006 Patent, col. 2:54-3:7). The complaint reproduces the patent's Figure 1, a diagram showing how components like a speech recognition engine, a parser, and various databases interact with these agents to process a user's request (Compl. ¶35).
• Technical Importance: This system architecture represented an approach to move beyond simple speech-to-text transcription toward true conversational understanding, a foundational concept for later digital assistants (Compl. ¶¶19, 33).

Key Claims at a Glance

• The complaint asserts one or more claims, including at least independent Claim 5 (Compl. ¶56).
• The essential elements of Claim 5 include:
  • receiving a natural language speech utterance containing a request;
  • recognizing words or phrases in the utterance;
  • parsing the information to determine a meaning and a context;
  • formulating the request in accordance with a grammar used by a domain agent associated with the determined context, which includes extracting, inferring, and transforming criteria and parameters;
  • processing the formulated request with the domain agent to generate a response; and
  • presenting the generated response (Compl. ¶58).

U.S. Patent No. 8,620,659 - "Systems And Methods Of Supporting Adaptive Misrecognition in Conversational Speech"

• Patent Identification: U.S. Patent No. 8,620,659, "Systems And Methods Of Supporting Adaptive Misrecognition in Conversational Speech," issued on December 31, 2013 (Compl. ¶39).

The Invention Explained

• Problem Addressed: The invention addresses the challenge of making conversational systems adaptive, enabling them to learn from user interactions and predict subsequent actions to improve conversational flow, particularly in cases of speech misrecognition.
• The Patented Solution: The claimed method uses two distinct learning models: a "personalized cognitive model" and a "generalized cognitive model" (’659 Patent, Claim 42). The personalized model is generated by tracking interaction patterns of a specific user, while the generalized model is based on patterns from a plurality of users (Compl. ¶78). The system first determines if the personalized model has sufficient information to predict a user's next action; if not, it uses the generalized model as a fallback, allowing the system to be both user-specific and robust (Compl. ¶78).
• Technical Importance: This technology describes a method for creating personalized and predictive conversational AI, allowing a system to anticipate user needs and become more efficient over time (Compl. ¶41).

Key Claims at a Glance

• The complaint asserts one or more claims, including at least independent Claim 42 (Compl. ¶76).
• The essential elements of Claim 42 include:
  • receiving a first input from a user comprising a natural language utterance;
  • generating an interpretation and a request, and transmitting it to a domain agent;
  • determining whether a personalized cognitive model (based on the user's interaction patterns) has sufficient information for predicting subsequent actions; and
  • predicting the subsequent actions based on a generalized cognitive model (based on patterns from multiple users) if the personalized model is found to lack sufficient information (Compl. ¶78).

Multi-Patent Capsule: U.S. Patent No. 7,398,209 - "Systems And Methods For Responding To Natural Language Speech Utterance"

• Patent Identification: U.S. Patent No. 7,398,209, "Systems And Methods For Responding To Natural Language Speech Utterance," issued on July 8, 2008 (Compl. ¶42).

Technology Synopsis

This patent, similar to the '006 Patent, addresses the problem that "machine-based queries" are "not inherently natural to the human user" ('209 Patent, col. 1:27-35). The solution involves a system that uses domain agents to organize domain-specific information, parses a query to determine the required domain and context, and invokes the proper resources to formulate and process a response ('209 Patent, col. 2:48-59, 3:53-54). The complaint includes Figure 6 from the patent, a flowchart illustrating the process of parsing a query, determining parameters, selecting agent(s), and generating a formatted command (Compl. ¶46).

Asserted Claims

The complaint asserts at least Claim 1 (Compl. ¶96).

Accused Features

The complaint alleges that the general natural language processing capabilities of the Salesforce Voice AI Agent, which allow it to engage in conversation and perform tasks, infringe the patent (Compl. ¶¶50-52, 99).

III. The Accused Instrumentality

Product Identification

The accused products are Salesforce Voice AI Agents and the associated products and services available through the Salesforce platform (Compl. ¶¶11, 50).

Functionality and Market Context

• The complaint describes the Salesforce Voice AI Agent as a technology "powered by Salesforce's large language model" that uses natural language processing (NLP) and machine learning (Compl. ¶50). Its purpose is to function like a customer service representative by engaging in conversations, answering queries, and performing tasks (Compl. ¶50).
• Users can interact with the agent via voice to perform functions such as obtaining product recommendations, handling returns, and troubleshooting technical issues (Compl. ¶51).
• The complaint alleges that Salesforce is a major technology company with significant market capitalization and revenue, and notes its 2024 acquisition of Tenyx, a developer of AI-powered voice agents, as context for its activities in this market (Compl. ¶¶49-50).

IV. Analysis of Infringement Allegations

The complaint references Appendices A, B, and C, which appear to contain claim charts, but these exhibits were not provided. The infringement theory is therefore summarized from the narrative allegations in the complaint.

’006 Patent Infringement Allegations

The complaint alleges that the Salesforce Voice AI Agent performs the method of Claim 5 by receiving natural language speech from users, using NLP to recognize words and parse them to determine meaning and context, and processing the resulting requests to generate and present a response (Compl. ¶¶50-52, 59). The core of the infringement theory appears to be that the different functionalities of the Voice AI Agent (e.g., for handling returns versus providing recommendations) operate as the claimed "domain agents" using specific "grammars" to process user requests (Compl. ¶¶51, 58).

’659 Patent Infringement Allegations

The complaint alleges that the Accused Products, which use machine learning, practice the method of Claim 42 (Compl. ¶¶50, 79). The infringement theory appears to be that Salesforce's machine learning systems function as the claimed "personalized" and "generalized" cognitive models. It is suggested that Salesforce's AI learns from individual user interaction histories (the "personalized" model) as well as from the aggregate data of many users (the "generalized" model) to predict user actions and improve responses (Compl. ¶¶78, 50).

Identified Points of Contention

• Scope Questions: A central question may be whether the term "domain agent," as described in the ’006 and ’209 Patents with a seemingly modular architecture, can be construed to read on the functionality of a modern "large language model," which the complaint alleges powers the accused product (Compl. ¶50). The defense may argue that an LLM-based system operates fundamentally differently than the discrete, agent-based system disclosed in the patents.
• Technical Questions: For the ’659 Patent, a key technical question is what evidence the complaint provides that the accused system performs the specific two-step process required by Claim 42: first, "determining whether a personalized cognitive model... includes sufficient information," and second, predicting actions "based on a generalized cognitive model in response to a determination that the personalized cognitive model does not" have sufficient information (Compl. ¶78). The dispute may focus on whether the accused machine learning system contains this specific decision logic or uses a more integrated learning process.

V. Key Claim Terms for Construction

• The Term: "domain agent" (from Claim 5 of the ’006 Patent and Claim 1 of the ’209 Patent)

• Context and Importance: This term is foundational to the architecture of the inventions in the ’006 and ’209 Patents. The outcome of the infringement analysis may depend on whether components of Salesforce's allegedly LLM-based system meet this definition. Practitioners may focus on whether this term requires a structurally distinct software module or can describe a functional capability within a larger, integrated AI.

• Intrinsic Evidence for Interpretation:

  • Evidence for a Broader Interpretation: The specification of the related ’209 Patent describes agents as "autonomous executables that receive, process and respond to user questions, queries and commands" and as "packages or modules of functionality, typically for a specific domain" (’209 Patent, col. 2:50-54). This language could support construing the term to cover logical software functions, regardless of their specific implementation.
  • Evidence for a Narrower Interpretation: The ’209 Patent also describes agents as "complete, convenient and re-distributable packages" that can be added or updated, suggesting discrete, self-contained, and interchangeable software components (’209 Patent, col. 2:52-54, 2:59-62). This could support a narrower construction that does not read on diffuse functionalities within a monolithic AI model.

• The Term: "personalized cognitive model" (from Claim 42 of the ’659 Patent)

• Context and Importance: This term and its counterpart, "generalized cognitive model," are central to the adaptive learning concept of the ’659 Patent. The dispute will likely focus on whether Salesforce's machine learning architecture maps onto this specific two-model structure.

• Intrinsic Evidence for Interpretation:

  • Evidence for a Broader Interpretation: The claim itself defines the term functionally as being "generated based on a tracking of a pattern of interactions between the user and the system" (’659 Patent, Claim 42). This could support a broad reading on any machine learning system that uses an individual's interaction history to personalize its behavior for that user.
  • Evidence for a Narrower Interpretation: The claim requires a specific decision process: "determining whether" the personalized model has "sufficient information" before deciding to use the generalized model (’659 Patent, Claim 42). This language may support a narrower construction requiring two distinct models that are consulted sequentially, which may differ from how modern integrated learning systems operate.

VI. Other Allegations

Indirect Infringement

The complaint alleges inducement of infringement for all asserted patents. The factual basis alleged is that Defendant supplies the Accused Products and encourages their use through its websites and product documentation, which allegedly instruct customers on how to use the infringing features (Compl. ¶¶62, 65, 82, 85, 102, 105).

Willful Infringement

Willfulness is alleged for all asserted patents based on Defendant’s knowledge of the patents and the alleged infringement "at least as of the filing of the Complaint" (Compl. ¶¶69, 89, 109). The complaint also pleads willful blindness (Compl. ¶¶69, 89, 109).

VII. Analyst’s Conclusion: Key Questions for the Case

• A core issue will be one of architectural scope: Can the patents' descriptions of a system built on discrete "domain agents" and distinct "personalized" versus "generalized" learning models be construed to cover the functionality of the accused Salesforce Voice AI Agent, which is allegedly based on a more modern, integrated large language model?
• A central evidentiary question will be one of operational proof: What evidence can Plaintiff uncover in discovery to demonstrate that the accused system performs the specific, multi-step logical processes required by the claims, particularly the ’659 Patent’s requirement to first assess the sufficiency of a personalized model before falling back to a generalized one?
• The case may also present a question of technological evolution: How will the court interpret claim terms drafted in the context of early-2000s technology when applying them to the substantially different architectures of modern AI and machine learning systems developed more than a decade later?