Exploratory Research of Integrating Large Language Models and Grounded Theory: Scientific and Technological Intelligence Case Study

Abstract

This study addresses the need to transform enterprise scientific and technological intelligence (STI) services from a discrete resource-supply model toward a more systematic value-creation approach, an important challenge in the digital transformation of knowledge-intensive industries. As an exploratory qualitative inquiry, this work combines large language model-assisted analysis with grounded theory to examine the construction logic and operational mechanisms of an embedded intelligent STI service system. Drawing on in-depth interviews with STI professionals, a qualitative corpus was analyzed using human–machine collaborative coding to systematically derive and organize key constructs. The findings yield a preliminary three-layer conceptual framework: “supply-demand interactive matching, organizational embedded services, and digital-intelligent platform support.” Specifically, the supply–demand matching layer facilitates targeted alignment through demand insight, dynamic response, and quality closed-loop management; the organizational embedded service layer delivers intelligence through scenario integration, process integration, and responsibility–authority integration; and the digital-intelligent platform support layer enables core capabilities via data element induction, intelligent diffusion, and tacit knowledge conversion. The proposed framework offers an initial, structured perspective on how embedded intelligent STI services may operate, providing a foundational reference for both research and practice in this emerging domain.

1. Introduction

Against the backdrop of increasingly fierce global sci-tech competition and an accelerating pace of industrial innovation, scientific and technological intelligence (STI) has evolved into a core strategic resource for enterprises to grasp technological trends, mitigate R&D risks, and seize market opportunities. Scientific and Technological Intelligence (STI) refers to the systematic collection, analysis, interpretation, and application of technical information, competitive data, and industrial dynamics to support organizational innovation and decision-making. As a research field centered on information processing and knowledge transformation, STI is inherently rooted in information organization, methodological innovation, and the core disciplines of information science. However, traditional STI services are predominantly resource-supply oriented, plagued by inefficient demand transmission, fragmented service scenarios and disjointed business processes. STI products often remain at the level of information collection, literature compilation and data statistics, failing to deeply integrate into key corporate links such as R&D decision-making and market expansion, which hinders the full release of their intrinsic value. In this context, embedded STI services have emerged as particularly crucial. Their core lies in the deep integration of intelligence capabilities into organizational business scenarios and decision-making processes, enabling a shift from passive supply to proactive adaptation, and thus bearing important practical significance for enhancing the accuracy and effectiveness of STI services.

The booming development of big data and artificial intelligence has opened up new development boundaries for the transformation and upgrading of modern STI services. In the intelligent era, STI work urgently needs to realize in-depth integration with digital and intelligent technologies, so as to break through the limitations of manual work, unlock multi-source data value, and further amplify the strategic supporting value of intelligence services for industrial innovation and enterprise development. Nevertheless, existing research and practical application still lack in-depth integration between intelligent technology and embedded STI scenarios, and fail to form a matched theoretical paradigm and service logic. Against this gap, this study attempts to combine emerging large language model technology with qualitative research methods, and explore a feasible human–machine collaborative analysis idea suitable for STI research. Taking grounded theory as the basic research paradigm and focusing on expert interview data, this study adopts a constrained LLM-assisted open coding scheme under human dominance. Through standardized qualitative sorting and analysis, it preliminarily constructs a three-layer operational framework of the embedded intelligent STI service system, so as to clarify the internal operation logic of embedded intelligence services under intelligent empowerment.

This study delivers tentative progress in both methodological exploration and practical model construction. In terms of research methods, the human–machine collaborative coding mode constructed in this research balances analysis efficiency and research rigor, which can provide a referable exploratory idea for the application of LLM technology in grounded theory and other qualitative studies. In terms of practical value, the initially proposed framework of the embedded intelligent STI service system sorts out the complete operating logic of demand docking, organizational embedding and platform support for intelligence services. It can offer basic reference and directional enlightenment for the optimization of STI service models in industries such as energy and advanced manufacturing. As an exploratory research, this study still has certain limitations in theoretical depth and scenario verification, and the constructed framework remains to be further optimized and supplemented in subsequent empirical research.

This article is organized as follows. Section 1 introduces the research background, motivation, and top-level conceptual framework of the embedded intelligent STI service system. Section 2 reviews the related work and existing literature. Section 3 describes the materials and research design, including data collection, expert selection, and the integrated methodology of large language models and grounded theory. Section 4 presents the main results, including the three-layer architecture and key categories derived from coding analysis. Section 5 elaborates the theoretical and practical implications of each dimension in real enterprise scenarios. Section 6 concludes the paper with a summary, limitations, and future research directions.

2. Related Work

The in-depth iteration of digital and artificial intelligence technologies is driving the profound transformation of STI services toward intelligence, contextualization, and value orientation. How to break the efficiency bottleneck of traditional STI services through technological empowerment and achieve precise matching between STI resources and innovation demands through system optimization has become a research focus in the current STI field. Existing studies have carried out diversified explorations around three core directions: the application of intelligent technologies in the STI field, the optimization and reconstruction of STI service models, and the value transformation of STI resources, laying an important theoretical and practical foundation for the construction of an embedded intelligent STI service system. However, they also expose problems such as scattered research perspectives and insufficient integration of technology and services.

In the research on intelligent technology empowering STI processing, the intelligent analysis and knowledge mining of multi-source heterogeneous data have become the core exploration direction. Aiming at the problems of low accuracy and high manual participation in knowledge entity extraction from STI resources, Liao et al. proposed an integrated method based on BiLSTM and Conditional Random Field, which can realize efficient extraction of knowledge entities without manual feature construction. This method can also construct knowledge networks and explore knowledge correlations based on the extraction results, providing a technical path for improving the quality of STI services [1]. To meet the demand for integrated analysis of multi-source data, Tang et al. constructed a multimodal BERT-KG model that integrates natural language processing, computer vision, and knowledge graph technologies. The model has shown good performance in tasks such as technological hotspot prediction and false intelligence identification, verifying the application potential of multimodal technologies in the collection and analysis of STI [2]. As a cutting-edge intelligent technology, the adaptability and application boundary of LLMs in the STI field have become research priorities. Considering the characteristics of limited open-source training data and high professional requirements in the STI field, Li et al. constructed a dedicated evaluation benchmark dataset. The study found that existing Chinese LLMs have basic domain knowledge but still have significant shortcomings in dynamic information tracking and in-depth thematic research, pointing out the optimization direction for the subsequent integration of LLMs with the STI field [3].

The optimization and reconstruction of STI service models focus on breaking the linear supply logic of traditional services and realizing a paradigm shift toward data-driven, demand-oriented, and scenario-embedded services. Based on the big data background, Liu et al. pointed out the limitations of traditional library and information services in data processing and demand response, and proposed in-depth transformation strategies from document management to data governance, from information services to knowledge services, and from universal provision to scenario integration. The study emphasized promoting the transformation of library and information institutions into knowledge ecosystem enablers by building a smart resource foundation, designing personalized service paths, and creating an embedded collaboration mechanism [4]. Taking the medical device field as a specific scenario, Ren et al. designed and implemented an efficient and scalable MCG information service system. Through the design of a networked and intelligent architecture, the system realizes cross-institutional data sharing and information exchange, providing a contextual practical example for the construction of industry-specific STI service systems [5]. Against the backdrop of global digital transformation, the technology adoption and system upgrading of intelligence services have also become an international research hotspot. Coulthart pointed out that disruptive technologies such as artificial intelligence and quantum computing bring transformation opportunities for intelligence services, but also face adoption challenges such as technological illiteracy and system incompatibility. A framework centered on “collaboration and connection” was proposed, providing an international reference for the path selection of digital transformation of intelligence services [6]. From the perspective of trust, Fernández et al. constructed a trust evaluation framework for intelligence services and found that service effectiveness and compliance are key factors affecting public trust, providing a diverse perspective for the optimization and improvement of intelligence service systems [7].

The value transformation of STI resources focuses on exploring the internal correlation between STI resource utilization and enterprise innovation development, providing empirical support for the value positioning of STI services. Based on massive enterprise patent data, Hou et al. conducted an empirical study and verified the positive promoting effect of the intensity and imbalance of STI resource utilization on enterprises’ breakthrough innovation. The study also found the moderating effect of strategic aggressiveness and the heterogeneous impact of regions and ownership types, and constructed a “strategy-capability-performance” theoretical framework. This research clarifies the mechanism of STI resources empowering enterprise innovation and also provides a practical basis for STI services to align with enterprise innovation demands and achieve precise empowerment [8].

In summary, existing studies have carried out multi-dimensional explorations around technological empowerment, model optimization, and value transformation, verifying the application potential of intelligent technologies in the STI field and clarifying the core direction of the transformation of service models toward contextualization and embedding. However, existing studies still have obvious limitations: first, technological research mostly focuses on a single link of intelligence processing, lacking systematic integration of the whole process of “data collection-analysis and mining-service output”; second, research on service models either focuses on macro strategies or single scenarios, and has not yet constructed a systematic STI service system adapted to the digital-intelligent era; third, the integration of cutting-edge technologies such as LLMs with STI services is still in the exploratory stage, lacking a theoretical framework and practical path for the deep coupling of technology and services. In this context, constructing an embedded intelligent STI service system that integrates digital-intelligent technologies, covers the entire service process, and adapts to diverse innovation scenarios has become a key direction to fill the gaps in existing research and promote the paradigm upgrade of STI services.

3. Materials and Methods

3.1. LLM-Assisted Collaborative Research Framework Grounded in Grounded Theory

This study establishes an LLM-assisted collaborative research framework with grounded theory as the core methodological approach. The detailed human–machine collaborative grounded theory coding workflow is presented in Figure 1.

This framework employs LLMs as an intelligent collaborative tool exclusively in the open coding phase of grounded theory analysis, which is widely recognized as the most labor-intensive and time-consuming stage of the entire analytical process. Specifically, the LLM functions as a supportive assistant to reduce researchers’ repetitive workload in initial concept extraction and preliminary coding tasks, rather than undertaking independent high-level interpretation or analytical judgments. Axial coding and selective coding are entirely conducted by the research team to maintain analytical rigor and interpretive control.

3.2. Grounded Theory

Grounded Theory, a qualitative research methodology proposed by Glaser and Strauss in 1967, is centered on the proposition that theories are constructed from empirical data through systematic induction, rather than being derived from pre-existing theoretical presuppositions [9]. At its core, it follows a rigorous, iterative process of “data first, theory later”: researchers immerse themselves in raw qualitative data (such as interviews or observations), systematically identifying patterns and emerging concepts through constant comparison, rather than testing pre-conceived hypotheses. Adhering to the principle of constant comparison, this method emphasizes the iterative cycle of data collection and analysis to extract concepts, identify categories, and establish their interrelationships until theoretical saturation is achieved. This “bottom-up” approach ensures that the resulting theory is closely grounded in real-world data, making it particularly suitable for exploring complex, understudied phenomena.

Grounded Theory was selected as the methodological foundation for this study due to two inherent characteristics of research on embedded STI systems. First, the embedded STI system represents an emerging STI service model, whose operational mechanisms and internal logic remain under-explored. Conventional library and information science theories are insufficient to explain the innovation of service models driven by digital-intelligent platforms, and Grounded Theory enables researchers to extract meaningful categories and develop theories from empirical data without relying on a fixed theoretical framework. Second, embedded STI systems exhibit dynamic and complex features, which are difficult to capture using traditional quantitative methods due to their nonlinearity and context dependence. The data-driven and iterative analytical process of Grounded Theory provides an effective approach for the exploratory construction of the dynamic operational mechanisms of embedded STI systems [10].

In the specific coding implementation phase, LLMs were integrated with Grounded Theory to adopt a human–machine collaborative coding approach, drawing on the excellent performance of LLMs in theme induction from semi-structured interview texts [11], as well as their 100% recall rate in open coding and full consistency with manual coding results in open coding [12]. The quality of model-generated coding outputs was controlled through prompt engineering, and the reliability of results was ensured via expert supervision.

3.2.1. Human–Machine Collaborative Coding Process

This study adopts DeepSeek-V3 as the auxiliary coding tool. Notably, the large language model is only applied in the open coding stage, and does not participate in axial coding and selective coding processes. Detailed structured prompt schemes targeting open coding tasks have been formulated to standardize coding granularity and domain semantic boundaries.

Prompt for Open Coding

Role: You are a professional academic researcher specializing in grounded theory analysis. Your role is to assist in open coding strictly based on the original empirical interview text.

Task: Extract initial concepts and initial categories from the given text accurately. Follow the original meaning of the text strictly without subjective extension, generalization, or addition of unsubstantiated content. Maintain consistent academic terminology, ensure clear coding granularity, and avoid duplicate concepts or ambiguous expressions. All outputs must be derived entirely from the provided corpus and conform to the research context of embedded scientific and technological intelligence (STI) services.

3.2.2. Bias Control and Coding Reliability Assurance

To mitigate output biases of LLMs during the coding process, a dual-dimensional bias control strategy was implemented in this study, covering source control and process intervention. For model selection, an LLM with semantic comprehension capabilities adapted to the textual analysis of qualitative research was adopted to reduce the impact of inherent semantic comprehension biases on coding results. For corpus processing, the original corpus was standardized through preprocessing, eliminating ambiguous expressions and standardizing professional terminology to minimize coding errors caused by miscomprehension. During coding implementation, targeted interventions were conducted by optimizing prompt instructions for biased content, ensuring alignment with research requirements.

Multi-dimensional reliability control strategies were further adopted to ensure coding stability. First, standardized prompt constraints were used to limit model coding scope and reduce generalization deviation. Second, full-process manual review was implemented, and all preliminary coding labels generated by LLM were checked and revised item by item by researchers and domain experts. When experts hold different views on the LLM’s preliminary concepts and coding results, all divergent items will be re-examined based on the original interview statements. Combined with unified coding specifications and contextual background of STI services, cross discussion and collective deliberation will be conducted to revise, screen or unify disputed concepts, so as to ensure the objectivity and rationality of final coding results. Third, internal consistency tests were performed by repeatedly inputting the same dataset under identical conditions, so as to verify the stability and reproducibility of the core category outputs.

3.3. Data Sources

The core empirical data of this study were collected through semi-structured in-depth contextual interviews with human experts. The study invited six human experts as formal interviewees, including four embedded STI service practitioners from universities (who provide intelligence support for research team construction and discipline development) and two experts from enterprise intelligence departments (responsible for strategic intelligence support for enterprise technology innovation and decision-making). With frontline practical experience in STI services and organizational application scenarios, these experts provided high-density and high-quality empirical data for the grounded theory analysis.

The entire interview process was conducted under a unified and standardized protocol. All interviews focused on the full life cycle of STI services guided by the core question bank (

Table 1. Core Question Bank.

No.	Core Question
1	How do you understand the embedded STI system, especially in comparison with traditional intelligence service models?
2	What do you consider to be an effective collaboration model between intelligence professionals and business units?
3	How do you think we should identify and translate business department needs into intelligence services? What challenges might be encountered in this process?
4	In the embedded STI system, how does knowledge flow and be shared? What impact do you think this knowledge flow has on decision support?
5	In your opinion, how important is it to establish an effective feedback mechanism during intelligence services? How does feedback information influence the optimization of subsequent services?
6	What are your expectations and suggestions for the future development of the embedded STI system? How can innovation be carried out in the context of technological advancement?

). Interviews with human experts were carried out via face-to-face voice communication, and all narratives were accurately recorded and transcribed into text using speech-to-text technology. All interview texts were subjected to standardized preprocessing (including deduplication, noise removal, and segment annotation), and a high-quality original corpus with a total of 25,697 Chinese characters was finally constructed for subsequent grounded theory coding analysis.

The core question bank covers six key dimensions related to the connotation, collaboration, demand translation, knowledge flow, feedback mechanism, and future development of embedded STI services. These questions are designed to fully capture experts’ practical experience, cognitive perspectives, and operational insights, ensuring that the collected data cover the whole life cycle of STI service design, implementation, and optimization. The semi-structured design allows flexible exploration while maintaining a consistent analytical focus.

During the interview process, when divergent opinions emerge among experts, we will first conduct targeted follow-up questions to clarify the contextual backgrounds, application scenarios, and underlying rationales behind different viewpoints. All divergent perspectives will be fully and objectively recorded without subjective filtering or premature judgment. In the subsequent grounded theory coding phase, these differing opinions will be systematically compared and analyzed to identify both core consensus and meaningful contextual variations. This approach ensures that the constructed theory can accommodate diverse practical experiences and reflect the inherent complexity of real-world STI services, providing a solid and comprehensive data foundation for subsequent grounded theory coding.

4. Results

4.1. Open Coding

Open coding is a critical step in grounded theory, which involves systematically extracting concepts and categories from raw data to gain an in-depth understanding of the core phenomena under investigation [13]. In this study, large language models (LLMs) were employed to assist researchers in the initial extraction and induction of concepts during open coding, while subsequent screening, refinement, and constant comparison procedures were completed by the researchers to ensure analytical rigor. Through analyzing, filtering, refining, and inducting initial concepts to derive initial categories, and eliminating contradictory concepts through constant comparison and analogy [14], this study ultimately extracted 34 initial concepts. Based on the logical relationships between these initial concepts, those with the same essential attributes were categorized, resulting in the integration of 9 basic categories, as detailed in

Table 2. Open Coding.

Initial Concept Categories	Conceptualization	Corresponding Raw Statement (Illustrative Example)
Demand insight	Demand foresight and prediction	“We need to proactively predict user intelligence needs, such as target market policies and consumer preferences.”
	In-depth demand deconstruction	“Through one-on-one interviews with business managers, we explore potential technical intelligence requirements.”
	Implicit demand analysis	“We identify cutting-edge technical intelligence needs by reviewing R&D plans.”
	Demand hierarchy configuration	“Prioritize intelligence needs by urgency: competitive intelligence first, new market information later.”
Dynamic response	Instantaneous response reach	“In case of emergencies, hold ad hoc meetings to quickly obtain targeted intelligence.”
	Agile intelligence transmission	“We set up real-time communication groups for competitor dynamics to release emergency intelligence instantly.”
	Real-time interactive support	“Before sales negotiations, we provide competitor price intelligence and share daily market updates.”
	Precision time-effect compression	“Adjust information collection workflows to accelerate delivery speed and meet tight deadlines.”
Quality closed-loop	Closed-loop quality management	“If business feedback indicates insufficient depth, we trigger further investigations.”
	Value evaluation closed-loop	“We quantify the timeliness and accuracy of intelligence in a closed-loop improvement cycle.”
	Service effectiveness release	“Reduce the time users spend filtering intelligence to improve decision-making efficiency.”
	Resource effectiveness evaluation	“Optimize input-output ratios based on feedback data and reallocate resources to high-demand areas.”
Scenario embedding	Multi-dimensional scenario analysis	“Map intelligence needs across all supply chain stages, from supplier selection to contract execution.”
	Scenario immersive insight	“By accompanying sales teams on client visits, we discover key product function concerns.”
	Customized intelligence products	“Design mobile-friendly intelligence reports for users who frequently access information on the go.”
	Scenario-based intelligence push	“Deliver marketing campaign reports to the marketing team and technical application cases to the R&D team.”
Process embedding	In-depth service integration	“Embed intelligence services deeply into users’ daily workflows and decision-making processes.”
	Dynamic process coupling	“Intelligence services run throughout the entire project lifecycle, from initiation to execution.”
	Precision node positioning	“Anchor intelligence support at key stages of marketing, including strategy, execution, and evaluation.”
Responsibility and authority embedding	Co-construction of trust and responsibility	“Hold regular feedback meetings to align goals and build mutual trust between teams.”
	Clear role division	“Business units clarify objectives; intelligence teams develop collection plans and convert knowledge.”
	Flexible standard interaction	“Establish norms for intelligence exchange while offering modular service components.”
	Dynamic balance of rights and responsibilities	“Define clear data sharing rights and responsibilities between different enterprises.”

. The integration of concepts into categories adheres to the principles of logical consistency and internal relevance, where each category represents a set of concepts with shared characteristics or similar attributes.

As shown in Table 2, the initial concepts and categories derived from open coding reflect the key elements of embedded STI services. Demand insight, dynamic response, and quality closed-loop jointly reflect the supply–demand interaction logic. Scenario embedding, process embedding, and responsibility–authority embedding represent the organizational integration path of intelligence services. Each category is formed by merging concepts with consistent attributes and logical relevance, ensuring that the coding structure is clear, hierarchical, and theoretically meaningful.

4.2. Axial Coding

Axial coding is the process of reconfiguring data by linking the conditions, contexts, action strategies, and consequences of the analyzed phenomena across different categories. This process involves systematically examining the relationships between basic categories to identify higher-level abstract concepts. Through in-depth analysis and integration of the 9 basic categories identified in open coding, this study ultimately distilled three main categories: supply-demand interactive matching, organizational embedded service, and digital-intelligent platform support, as shown in

Table 3. Axial Coding.

Main Categories	Initial Concept Categories	Core Connotation
Supply–demand interactive matching	Demand insight	A systematic framework for identifying explicit and implicit intelligence needs, dynamically tracking changes, and prioritizing resource allocation.
	Dynamic response	Real-time adjustment of service strategies based on environmental changes, ensuring the timeliness and accuracy of intelligence delivery.
	Quality closed-loop	A feedback-driven mechanism to continuously evaluate and improve service quality, quantifying the impact of intelligence on business decisions.
Organizational embedded service	Scenario embedding	Designing tailored intelligence products and delivery strategies for specific business contexts to increase relevance and effectiveness.
	Process embedding	Integrating intelligence support directly into key workflow steps to ensure timely, context-aware decision support.
	Responsibility and authority embedding	Embedding intelligence services into organizational decision chains by clarifying roles, rights, and responsibilities for a responsive service system.
Digital-intelligent platform support	Data element induction	Aggregating multi-source data and performing governance tasks (cleaning, classification) while ensuring data security.
	Intelligent technology penetration	Leveraging AI and big data to automate collection and enhance analysis efficiency and accuracy.
	Tacit knowledge transformation	Converting employees’ tacit experience into shared organizational knowledge, driving its application and iteration in business scenarios.

.

Table 3 presents the hierarchical relationship between main categories and basic categories. Supply–demand interactive matching serves as the driving force of the system. Organizational embedded service acts as the practical execution path. Digital-intelligent platform support provides the underlying technical and data conditions. This hierarchical structure reveals the internal logic of the embedded intelligent STI service system and lays a foundation for constructing the overall theoretical framework.

4.3. Selective Coding and Theoretical Saturation Test

The purpose of selective coding is to extract the core category from the main categories. This process includes analyzing the relationships between the core category, main categories, and other relevant categories, developing a storyline that runs through the entire study, conceptualizing the forms of association between categories, and thereby evolving a new theoretical framework [15].

The embedded intelligent STI service system constructed in this study aims to break the traditional separation between intelligence and business operations. Through the dynamic collaboration of supply-demand interactive matching, organizational embedded service, and digital-intelligent platform support, it promotes the upgrading of intelligence from a mere auxiliary tool to a strategic productive force. In this system:

Supply–demand interactive matching: accurately identifies and responds to business needs, providing the “demand-driven impetus” for system operation.

Digital-intelligent platform: support provides technical and data backing through data integration, intelligent technology, and knowledge transformation.

Organizational embedded: service embeds intelligence generated by the digital-intelligent platform into business scenarios, processes, and responsibilities, ensuring that intelligence effectively impacts business operations.

These three components rely on three interactive spaces: demand-technology linkage, technology-service transformation, and service-demand feedback, forming a cycle of “demand driving technology, technology transforming into services, and services feeding back to optimize demand”. This cycle breaks through the traditional one-way supply model of intelligence, reveals the synergistic effect of demand insight accuracy, organizational embedding depth, and digital-intelligent technology penetration intensity, and provides a theoretical framework and practical path for building an intelligence service system with dynamic adaptation between demand and service. As illustrated in Figure 2, the functional model clearly presents the cyclic interaction among the three core dimensions: supply–demand interactive matching acts as the demand-oriented driving force, digital-intelligent platform support provides data and technical empowerment, and organizational embedded service converts intelligence capability into practical business value.

This study conducted a theoretical saturation test using reports related to the construction of enterprise STI systems as supplementary raw data. After performing open coding, axial coding, and selective coding on a total of 14,253 Chinese characters of these reports, theoretical saturation was fully verified based on four rigorous criteria.

First, no new information, concepts, or attributes emerged from the incremental supplementary data, and all data content only supported the existing coding results without generating new analytical elements.

Second, the entire category system was sufficiently developed with complete dimensions and clear internal boundaries, and all core categories and subcategories had reached adequate connotation expansion.

Third, the logical relationships and connection paths among main categories remained stable and consistent, and all inter-category correlations were fully supported by empirical evidence without ambiguity or contradiction.

Fourth, the three-layer theoretical framework exhibited comprehensive explanatory power for the formation mechanism, operational logic, and practical value of the embedded STI intelligent service system, with sound internal consistency and external applicability. Therefore, the theoretical model was confirmed to be saturated and validated.

5. Embedded STI Intelligent Service System (E-STI-ISS)

The embedded STI system adopts a macro-strategic perspective to drive the deep integration of STI into the enterprise innovation ecosystem. By establishing professional teams with forward-looking insight and dynamic response capabilities, it embeds the full cycle of intelligence services into the enterprise innovation value chain, proactively perceives and anticipates intelligence needs, and provides timely intelligence products and services to maximize innovation value, thereby fulfilling the important leading role of STI [16]. This study further refines the positioning and functions of “intelligence demand traction”, “organizational embedded service”, and “digital-intelligent platform support” by constructing a supply-demand matching layer, an organizational embedding layer, and a platform support layer within the intelligence service system. The model of the embedded STI service system is illustrated in Figure 3.

Figure 3 further presents the three-layer structural framework of the embedded intelligent STI service system. The driving layer focuses on supply–demand matching to ensure accurate demand response. The execution layer emphasizes organizational embedding to realize in-depth integration of intelligence and business. The support layer relies on digital-intelligent platforms to provide stable data and technical support. The three layers cooperate with each other to form a systematic, complete, and operable intelligent service system.

5.1. Supply–Demand Interactive Matching

5.1.1. Demand Insight

Demand insight serves as the cognitive core of intelligence-driven demand traction, aiming to build comprehensive perception and standardized management capabilities for enterprise intelligence needs. Faced with massive and fragmented external market information, enterprises commonly suffer from information overload, in which redundant data hinders the screening and precise analysis of valuable information [17]. By integrating multi-source data covering market dynamics, user behavior and competitive intelligence, combined with semantic network analysis, this system can capture both explicit demands (e.g., targeted acquisition of competitor operational data) and implicit potential needs (e.g., long-term technical trend forecasting). Furthermore, a dynamic demand list and lifecycle tracking mechanism unify scattered business requirements into a standardized management framework. This ensures intelligence services are closely aligned with strategic goals, shifting service logic from passive response to proactive demand prediction.

5.1.2. Dynamic Response

Dynamic response acts as the operational core of demand traction, focusing on balancing fluctuating business demands and flexible service capabilities. The core advantage of embedded intelligence services lies in data insight, which helps identify weak early warning signals from complex operational data, including abnormal market fluctuations, subtle consumer preference changes and differentiated industry development trends [18]. Supported by streaming computing architecture and elastic resource scheduling, emergent business demands can be perceived and responded to in a timely manner. Meanwhile, modular decomposition of intelligence products realizes rapid combination and iterative optimization, breaking the constraints of the traditional linear service delivery model. Facing the dual pressure of diversified demands and timeliness requirements, it achieves efficient coordination between intelligence supply and enterprise business rhythm.

5.1.3. Quality Closed-Loop

Quality closed-loop management constitutes the performance guarantee of demand traction, forming a cyclic optimization mechanism of “intelligence output-business feedback-service iteration”. First, real-time monitoring of full-link indicators such as data completeness, logical rationality and delivery efficiency, combined with anomaly diagnosis models, enables timely identification and rectification of service defects. Second, multi-dimensional evaluation indicators, including strategic matching degree, decision-making support effect and input-output ratio, are adopted to assess the dual value of intelligence services. It systematically quantifies tangible benefits such as improved decision-making efficiency and intangible gains such as R&D risk reduction, so as to continuously polish and upgrade service quality.

In practical enterprise operation, the supply-demand interactive matching layer can be implemented in daily innovation management and market layout. For technology-oriented manufacturing enterprises, this layer can accurately capture the intelligence demands of R&D, marketing and strategic departments, rapidly respond to sudden industry policy adjustments and competitor dynamic changes, and optimize intelligence service efficiency through whole-process quality management. In practice, it helps enterprises avoid information deviation and demand lag, reduce invalid intelligence output, and provide targeted, high-value decision-making references for market competition and technological research and development.

5.2. Organizational Embedded Service

5.2.1. Scenario Embedding

Scenario embedding defines the value orientation of organizational embedded services, delivering customized intelligence support for diversified business scenarios. According to industrial attributes, decision-making subjects and business cycles, differentiated service strategies are formulated to match scenario characteristics. A multi-modal intelligence product system is constructed, covering in-depth technical analysis reports and professional data sorting, as well as lightweight decision briefs and interactive dynamic tools. Diversified content forms and flexible delivery modes can fully adapt to the personalized needs of core scenarios such as technological R&D, market expansion and supply chain management.

5.2.2. Process Embedding

Process embedding reflects the core execution logic of embedded services, realizing deep integration of intelligence support into key decision-making links of business processes. By disassembling enterprise operational processes, key nodes such as demand initiation, scheme decision and project execution are accurately identified, with clear service trigger conditions and delivery standards. This design enables on-demand invocation and precise matching of intelligence resources in the whole business chain, realizing synchronous coordination between intelligence supply and daily business actions, and providing continuous and effective support for process optimization and project advancement [19].

5.2.3. Responsibility and Authority Embedding

Responsibility and authority embedding provides institutional guarantees for embedded services, integrating intelligence capabilities into enterprise organizational structures and collaborative mechanisms. By optimizing internal right and responsibility division, the functional boundaries of business departments and intelligence teams in demand submission, resource allocation and achievement application are clearly clarified [20]. Standardized collaborative norms further improve the intelligence-driven decision-making system, endowing intelligence services with legitimate organizational participation. Clarified right and responsibility boundaries reduce internal collaboration costs, forming an efficient operation ecosystem with rapid demand response and smooth value transformation.

At the organizational level, the embedded service layer can be directly applied to enterprise daily management, cross-departmental collaboration and business process reengineering. Taking high-tech enterprises as an example, scenario-based embedded intelligence can provide exclusive analysis services for R&D iteration and market expansion; process embedding integrates intelligence results into key production and operation links; responsibility and authority embedding standardizes the collaborative mode between functional departments. In practice, this layer eliminates the separation between intelligence services and actual business, realizes the seamless docking of intelligence achievements and enterprise management, and improves the overall operational collaboration efficiency and organizational decision-making execution.

5.3. Digital-Intelligent Platform Support

5.3.1. Data Element Induction

Data element induction lays the fundamental foundation for digital-intelligent support, with standardized intelligent data governance as the core. It covers the whole lifecycle management of STI data such as collection, storage, processing and sharing. Unified industry specifications and strict quality control standards ensure the accuracy, consistency and integrity of multi-source data [21]. Supported by intelligent algorithms, automatic cleaning, classification and standardization of heterogeneous data are completed, converting scattered raw data into reusable standardized intelligence elements. A traceable and shared intelligent data resource pool is ultimately built, providing stable and high-quality basic data support for subsequent intelligent analysis and knowledge mining.

5.3.2. Intelligent Technology Penetration

Intelligent technology penetration acts as the technical pillar of the platform layer, promoting the deep integration of emerging digital technologies and full-chain STI services. Core technologies including large language models, machine learning and natural language processing reshape the whole logic of intelligence production and output. It realizes multi-source data fusion, semantic content analysis, rule mining and visual result presentation, and empowers intelligent upgrading of traditional manual work links. In-depth technical penetration fully activates data potential, excavates hidden knowledge correlations and industry development laws, and significantly improves the depth, efficiency and professionalism of STI service output [22].

5.3.3. Tacit Knowledge Transformation

Tacit knowledge transformation is the core value engine of digital-intelligent services, breaking the one-way barrier between tacit experience and explicit organizational knowledge. Relying on LLM knowledge modeling, knowledge graph and intelligent case base management, scattered experience and scenario-based professional insights of intelligence practitioners are solidified into shareable and iterable organizational knowledge assets. Meanwhile, context-aware push and human–machine collaborative decision tools promote the internalization of explicit knowledge into business practices, forming a closed loop of two-way knowledge transformation. It accelerates cross-departmental knowledge sharing and continuous value appreciation, and solves the problem of difficult inheritance of professional experience in traditional intelligence work.

The digital-intelligent platform support layer is the basic technical guarantee for the long-term operation of embedded intelligent STI services, which is widely applicable to digital transformation and knowledge management of large and medium-sized enterprises. In practice, the platform can unify the management of internal and external intelligence data of enterprises, rely on AI technologies to realize automatic analysis and efficient output of massive information, and solidify the tacit experience of core employees into institutional knowledge assets. It effectively reduces manual data processing costs, realizes sustainable accumulation and inheritance of enterprise intelligence resources, and provides solid technical and data support for the long-term innovation and high-quality development of enterprises.

6. Conclusions

6.1. Theoretical Implications

This study responds to the evolving demands of scientific and technological intelligence (STI) services in the digital-intelligent era by constructing a preliminary theoretical framework for embedded STI intelligent service systems. It integrates the perspectives of embedded services and digital intelligence, unpacking the basic internal logic and potential operational mechanisms of this emerging service paradigm.

Methodologically, the explored human–machine collaborative coding approach provides tentative insights and a referable exploratory path for qualitative analysis in the AI era. By limiting LLM involvement to the labor-intensive open coding stage and maintaining researcher-led control over axial and selective coding, this approach balances analysis efficiency and basic research rigor. Beyond STI research, it offers heuristic references for qualitative studies in management, information science and related social science fields, providing a viable exploratory thought for combining generative AI with grounded theory workflows.

Furthermore, the initially constructed three-layer theoretical model of the embedded STI intelligent service system (E-STI-ISS) supplements and enriches partial perspectives in the theoretical research of STI service innovation. The framework clarifies the basic multi-dimensional cooperative mechanism of demand matching, organizational embedding and platform support, and expands the tentative application boundary of embedded service theory in typical emerging STI scenarios, including power grid technology intelligence, new energy industry intelligence, and advanced manufacturing innovation intelligence. It offers a conditional analytical perspective to assist in understanding the integration of STI services with industry-specific innovation ecosystems under specific contextual conditions.

6.2. Practical Implications

The supply–demand interactive matching layer provides tentative practical references and localized optimization ideas for alleviating the common disconnection between intelligence supply and innovation demand within enterprises. In the typical context of power grid technology intelligence services, for reference, this layer can help intelligence teams accurately capture partial dynamic R&D demands of new energy grid projects through continuous demand insight, flexibly adjust intelligence themes and scopes according to project iteration, and form a basic quality closed-loop via follow-up feedback. This model is conducive to promoting the transformation of partial STI services from passive information supply to limited proactive embedded support, so as to better fit the incremental innovation demands of enterprises.

The organizational embedded service layer helps alleviate the separation between intelligence work and daily business processes to a certain extent. Taking the new energy industry intelligence scenario as a typical example for discussion, this layer provides actionable ideas for embedding intelligence links into key innovation activities such as new material R&D and strategic layout. By integrating partial intelligence work into daily business links and clarifying matched division of responsibilities, it enables intelligence services to better connect with business scenarios, and improves the pertinence and timeliness of intelligence support within limited application scopes.

The digital intelligence platform support layer offers basic technical thinking and developmental references for the digital and intelligent upgrading of STI services. In the scenario of advanced manufacturing innovation intelligence, this framework provides a feasible developmental direction for the standardized collection of multi-source data, classified dissemination of intelligence achievements, and inheritance of empirical experience of analysts. It is helpful to improve the operational efficiency of intelligence teams within a certain range and promote the incremental iterative optimization of embedded STI service modes.

6.3. Limitations

Sample and contextual constraints: This research draws empirical materials from a small group of professional experts, which guarantees targeted insight for the selected research context. However, due to the limited sample scale, the current findings and framework are context-specific. Further cross-industry and large-sample verification is required before extended generalization to diverse organizational scenarios.

Framework application boundary: The three-layer framework is summarized and constructed based on the current research context, which can effectively explain the operational logic of embedded intelligent STI services in this study. Still, its long-term adaptability and scalable application in more complex industrial environments need continuous improvement and empirical testing in future practice.

Method optimization space: The LLM-assisted human–machine collaborative coding scheme is a feasible exploratory attempt for qualitative research. Nevertheless, to form a more standardized and universal methodological paradigm, more repeated tests and iterative optimization are needed in subsequent research.

6.4. Future Research Directions

Future work could (1) expand the sample scope to include larger, cross-industrial and cross-regional datasets to test and refine the framework’s generalizability; (2) conduct longitudinal case studies to examine the practical implementation and dynamic evolution of embedded intelligent STI services in real-world enterprises; and (3) further optimize the human–machine collaborative coding workflow to enhance its reliability, transparency, and replicability for qualitative research applications.