Integrating artificial intelligence with organisational knowledge management: a grounded theory study Open Access

Artificial intelligence (AI) is reshaping global industries through unprecedented widespread adoption (Lee et al., 2023). AI purports to possess immense potential to enhance productivity and foster innovation (Al-Dmour et al., 2021; Brynjolfsson and McAfee, 2014; Davenport, 2018). The strategic importance of AI is reflected in its rapidly escalating prioritisation among CEOs, moving from the fifth priority in early 2019 to the second by the third quarter of 2024 (Paraskevopoulos, 2024). Despite its potential value, the strategic integration of AI into an organisation is a complex endeavour; most corporate AI implementations take years to complete, with costs often reaching millions of dollars (Lee et al., 2023). While personal use of AI can be intuitive, AI as a corporate technology solution is not a mere plug-and-play technology that functions instantly upon installation (Fountaine et al., 2019; Lee et al., 2023). When integrating AI into an organisation, it often faces a complex set of technological options and strategic directions, which can result in fragmented implementations that neither create lasting value nor align with the organisation’s core objectives (Westenberger et al., 2022). Such challenges are further underscored by the alarming failure rate of AI projects, with estimates as high as 87%, highlighting the urgent requirement for a robust guiding framework (Weiner, 2025). As a fundamental principle for fulfilling the organisation’s unique requirements, many companies incorporate internal data to refine AI tools, a process commonly referred to as domain adaptation or fine-tuning (Alavi et al., 2024, p. 4). Therein, considerable attention should be paid to the organisation’s existing knowledge. It is noted that since the 1990s, organisations have invested substantial resources in knowledge management (KM) initiatives to support knowledge-based decision-making (Smith and Farquhar, 2000), achieving notable benefits to organisations (Alavi et al., 2024). Integrating this existing ‘knowledge’ of the organisation with ‘artificial intelligence’ is an integral component of domain adaptation in AI. While much of the recent literature focuses on AI capabilities or KM principles (Davenport and Kirby, 2016), the benefits of AI to organisations (Arakpogun et al., 2021; Balage and Sedera, 2024, 2025b; Huang and Rust, 2018; Merhi, 2023; Olan et al., 2022) and the challenges of introducing AI to organisations (Dwivedi et al., 2021), there is a lack of empirical guidance on how AI should be integrated into existing knowledge. This paper aims to bridge that gap by addressing the research question: How do organisations meaningfully integrate AI into operational practices in light of the existing organisational knowledge management?

The research question highlights the objectives of the study, where our emphasis is on the manner in which AI is integrated to organisational knowledge ^[1]. To answer this, we use a rigorous grounded theory approach (Strauss and Corbin, 1990, 1998). Our data comprises a large set of 100 publicly available AI implementation case studies, drawn from diverse, knowledge-intensive industries, including health care, financial services, manufacturing, government and education. Through a systematic, multi-level coding process, we generated open codes, which were abstracted into coherent axial codes, ultimately informing final selective codes. The case organisations primarily adopted existing, pre-built AI platforms, which they then customised using their unique internal knowledge base through processes like domain adaptation.

This framework posits that useful AI integration manifests in four distinct strategic archetypes, determined by three primary organisational decisions: the scope of the knowledge application (horizontal vs vertical), the level of human–AI intervention (human-verified vs non-human verified) and the workflow configuration (automation vs augmentation). This paper’s primary contribution is an empirically grounded typology that extends classic KM and IS theories by providing a nuanced framework for understanding KM–AI integration. It offers actionable insights for leaders to strategically assess and implement AI solutions for meaningful organisational transformation.

AI has undergone significant advancements over the years (Luitse and Denkena, 2021), profoundly expanding its influence across various fields of study and industry sectors (Balage and Sedera, 2026; Davenport, 2018; Kar et al., 2023; Weill and Woerner, 2018). McCarthy et al. (2006) defined AI as “the science and engineering of creating intelligent computer programs” in 1956 (Cave and Dihal, 2023; Madan and Ashok, 2023; Uren and Edwards, 2023, p. 1). AI, as a distinct branch of computing (Mao et al., 2016; Sharma et al., 2023), has evolved from being primarily confined to a select group of tech-savvy experts to becoming an accessible resource for non-technical individuals, utilised for both personal and organisational purposes (McKinsey Global Institute, 2017). The recent adoption of GPT-4 has significantly impacted organisations, leading to a 12.5% increase in consultant productivity and a 40% improvement in the quality of their work (Dell’Acqua et al., 2023). This impact is particularly evident in the use by non-technical personnel across tasks ranging from creativity to analysis, writing, marketing and persuasion. Furthermore, AI was found to equalise skill levels (Mollick, 2023).

From a business perspective, AI can be understood not merely as a technological artefact, but as a capability for augmenting and automating organisational decision-making and knowledge work (Davenport, 2018; Huang and Rust, 2018). Rather than replacing traditional IT systems, AI extends them by enabling organisations to interpret large volumes of structured and unstructured data, generate predictions and support or execute decisions. In this sense, AI represents a shift from information processing to intelligence generation, where systems actively contribute to organisational outcomes rather than passively storing or transmitting data (Leoni et al., 2024).

The impact of AI on business manifests through several interrelated mechanisms. Firstly, AI enables the automation of routine cognitive tasks, reducing reliance on manual processing and increasing operational efficiency (Davenport, 2018). Secondly, it facilitates augmentation of human decision-making, where AI systems provide recommendations, predictions or insights that enhance human judgement rather than replacing it (Jarrahi, 2018). Thirdly, AI supports knowledge synthesis at scale, identifying patterns, correlations and anomalies across data sets that exceed human cognitive limits (Leoni et al., 2024). Finally, AI enables personalisation and real-time responsiveness, allowing organisations to tailor services, interactions and decisions based on continuously updated data inputs (Gelashvili-Luik et al., 2025).

This expanding accessibility is underpinned by significant advances in AI affordability. Although the development of advanced AI systems, such as the GPT-4 model, can incur substantial pre-training costs exceeding US$100m (Xia et al., 2024) and trillions of parameterisations (Luitse and Denkena, 2021; Xia et al., 2024), AI is increasingly integrated into many open-source and licenced products (Cheliotis, 2009). For instance, ChatGPT set a historic milestone as the fastest-growing application, reaching 100 million active users within two months of its launch, making the technology accessible to a diverse range of individuals and businesses (Hu, 2023).

This trend is further fuelled by tech giants such as Google and Microsoft, who are actively focusing on affordable and accessible AI models, thereby reducing dependency on highly skilled AI professionals and enabling users to personalise AI without extensive coding (Metz, 2017; Stewart, 2017; Wilson and Daugherty, 2018). The accessibility of AI for non-technical employees is known as ‘democratisation of AI’ (Wilson and Daugherty, 2018). The growing number of AI service providers has intensified competition, leading to lower prices and making AI more affordable for businesses (Crawford et al., 2024). This transition enables small companies to leverage AI, which was previously reserved for large companies.

As industries globally transition from manufacturing towards knowledge-based economies (Birzniece, 2011; Rothberg and Erickson, 2005), knowledge has solidified its position as an increasingly valuable asset for organisations (Alavi et al., 2024; Durst and Edvardsson, 2012; Walczak, 2005). It is widely regarded as both knowledge capital (Liao, 2003; Rinta-Kahila et al., 2023; Someh et al., 2023) and a crucial source of strategic advantage (Gonçalves et al., 2024; Nakash and Bouhnik, 2021; Nonaka and Toyama, 2003; Younis and Adel, 2020). Organisations that effectively leverage their knowledge are positioned to lead by developing the most innovative products (Akram et al., 2011; Mishra and Pani, 2021).

Polanyi (1958) categorised knowledge into explicit and tacit forms. Explicit knowledge is a theoretical, documented and initial step towards generating tacit knowledge (Hislop et al., 2018; Kovačić et al., 2022; Polanyi, 1958; Sedera et al., 2003; Subasinghage et al., 2014). Conversely, difficulty in observing, articulating or learning through experience or subconsciously is known as tacit knowledge (Baskerville and Dulipovici, 2006; Ebrahimi et al., 2013; Nonaka and Konno, 1998; Polanyi, 1958; Sedera et al., 2003). Once tacit knowledge is converted into explicit knowledge by a knowledge worker, it is made available when and where others need it (Brahami and Matta, 2018; Nemati et al., 2002). Knowledge workers operate at operational, management and strategic levels (Sedera, 2007; Zakaria et al., 2010).

Effective KM is, therefore, a strategic approach designed to address several key organisational challenges, including improving employee knowledge, reducing knowledge gaps and minimising knowledge loss (Davenport and Prusak, 1998; Rothberg and Erickson, 2005). The traditional KM process encompasses capturing, distributing and utilising knowledge (Davenport and Prusak, 1998), with platforms often serving as repositories where employees refer to obtain necessary information to perform tasks.

Beyond defining knowledge as a resource, a substantial body of work grounded in the knowledge-based view (KBV) has sought to explain how knowledge is converted into organisational value (Grant, 1996). A consistent finding across this literature is that competitive advantage flows not from the mere possession of knowledge but from an organisation’s capability to manage it. This capability is typically decomposed into a set of knowledge processes – variously described as acquisition, conversion, application and protection (Naqshbandi and Jasimuddin, 2018) or as acquisition, dissemination and responsiveness (Bashir et al., 2024) – that rest on a supporting infrastructure of technology, organisational structure and culture (Naqshbandi and Jasimuddin, 2018). Two patterns recur. Firstly, this capability is repeatedly positioned as an antecedent of innovation: KM capability has been shown to drive open innovation in multinational subsidiaries (Naqshbandi and Jasimuddin, 2018) and to be the primary driver of business model innovation in small and medium enterprises, which in turn yields competitive advantage and superior firm performance (Bashir et al., 2024). Secondly, in both accounts, the conversion of knowledge into innovation is mediated by human agency – by knowledge-oriented leaders who shape culture, structure and technology and motivate knowledge processes (Naqshbandi and Jasimuddin, 2018) and by managerial cognition that recognises and recombines knowledge into novel models (Bashir et al., 2024). Technology, in this established view, functions as passive infrastructure: a repository or conduit that stores and transmits knowledge for human actors to interpret and apply. This shared logic, where knowledge is a strategic resource, made valuable through human-led processes and innovation outcomes, provides the theoretical baseline against which the arrival of AI must be assessed.

Analysing the AI and KM literatures reveals a point of convergence and a point of tension that jointly motivate this study. The convergence is theoretical: studies of AI in organisations (Davenport, 2018; Leoni et al., 2024), and the KBV tradition that frames KM as a driver of innovation (Bashir et al., 2024; Naqshbandi and Jasimuddin, 2018) agree that value is realised only when knowledge is actively processed, rather than merely stored. Both traditions, moreover, decompose that processing into a recognisable sequence, for example, sourcing knowledge, converting or recombining it, applying it and embedding it for reuse. In this respect, AI does not introduce a new vocabulary of knowledge work so much as it intervenes in an existing one. The tension lies in who, or what, performs the mediation. In the established literature, the link from knowledge to innovation is carried by human actors. Naqshbandi and Jasimuddin (2018) demonstrate that KM capability fully mediates the relationship between knowledge-oriented leadership and both inbound and outbound open innovation – that is, the capability that converts knowledge into innovation is exercised through leaders who build infrastructure and steward knowledge processes. Bashir et al. (2024) similarly position KM as the antecedent that managers translate into competitive advantage, which in turn drives firm performance. In both, technology is infrastructure, and the intelligence doing the converting is human. Contemporary AI disturbs this settlement: machine learning and generative systems are data-driven, probabilistic and adaptive, generating, transforming and recombining knowledge through pattern recognition rather than executing pre-defined human rules (Nguyen and Vuong, 2025; Tshitoyan et al., 2019). AI, therefore, migrates from passive infrastructure to an active participant in the knowledge process. This is a shift that the human-mediated models of KM capability were not designed to accommodate.

This repositioning surface three unresolved issues that prior work, precisely because it predates pervasive AI, leaves open. Firstly, where earlier KM capability frameworks treat the technology infrastructure dimension as an enabler of human knowledge work (Naqshbandi and Jasimuddin, 2018), they do not specify how an infrastructure that itself generates knowledge should be governed, trusted or integrated into the process. Secondly, the established view privileges stable, codified repositories optimised for reuse and consistency (Davenport, 1998), whereas AI pushes organisations towards dynamic, continuously recombined knowledge – a movement from static repositories to evolving knowledge ecosystems. Thirdly, although the locus of mediation shifts towards the algorithm, the literature converges on the enduring importance of the human and social dimension: just as leadership and culture remain decisive in classic KM capability accounts (Naqshbandi and Jasimuddin, 2018), the question of human oversight, verification and trust becomes more, not less, salient as systems become more autonomous.

Taken together, these studies establish a robust account of human-mediated KM but leave algorithm-mediated KM under-theorised, which AI now makes possible. Specifically, the literature offers limited empirical insight into how organisational knowledge is selected and adapted for AI systems, how AI-generated outputs are incorporated into KM practices and how the balance between algorithmic autonomy and human verification is configured in practice. Table 1 consolidates this synthesis, and our study addresses these questions directly by tracing how organisations successfully convert latent human expertise into stable, trusted knowledge assets through AI.

The authors have chosen a qualitative, interpretive case study method to reflect the exploratory nature of the study and to examine the research question (Ponelis, 2015; Tan et al., 2018; Walsham, 1995). Multiple interpretive case studies were used to enable in-depth analysis and to address the “how” aspect (Tan et al., 2018; Tan et al., 2017; Tan et al., 2016; Yin, 2009) of the research question: How do organisations meaningfully integrate AI into operational practices in light of the existing organisational knowledge management? The use of secondary data within Information Systems (IS) has historically been limited (Balage and Sedera, 2025a; Rinta-Kahila et al., 2022). However, the proliferation of AI has created a rich repository of publicly available data sufficient for this study, as evidenced by the success stories of organisations that have implemented AI using Microsoft or Amazon Web Services (AWS) solutions. This sample of cases provided the richness of information needed for the study (Ebneyamini and Moghadam, 2018; Eisenhardt, 1989; Mazurova et al., 2022; Patton, 2002). Several criteria were employed to select cases. The case organisation must demonstrate a proven success of AI integration, and by focusing on implementation narratives published by industry-leading platform providers, Microsoft and AWS, the study ensured access to a repository of cases that had already undergone internal corporate validation. This strategy mitigated the risk of analysing speculative or abandoned projects. The final sample consists of 100 case studies published between 2017 and early 2025 (see Table 2).

The selection of sectors followed a theoretical sampling logic (Eisenhardt, 1989), aimed at capturing diversity in AI implementation contexts rather than statistical representativeness. The chosen sectors of health care, financial services, education, government and manufacturing are all knowledge-intensive domains with increasing levels of AI adoption but differ significantly in terms of knowledge structure, regulatory constraints and operational processes. This variation enabled the identification of cross-contextual patterns while ensuring that findings were not confined to a single industry setting. The complexity of knowledge processes created the homogeneity across the heterogeneous contexts, making them particularly relevant for examining the integration of AI into KM practices.

Furthermore, the sample includes organisations of varying scales, based on employee numbers (see Table 3). The number of employees was used as a proxy for organisational scale, capturing variation in resources, process complexity and AI adoption capacity. Including organisations of different sizes allowed us to observe AI–KM integration across diverse contexts, from smaller, less formalised environments to large organisations with structured knowledge and data infrastructures. This variation in scale is particularly relevant, as it influences the extent to which organisations can support automation, human verification and knowledge stabilisation mechanisms identified in our analysis.

Since the question required a series of data points, we needed as many examples as possible to extract patterns. We employed an inductive, grounded theory approach to our analysis method, allowing theoretical constructs to emerge directly from the data (Kaluarachchi et al., 2025; Rinta-Kahila et al., 2022). To ensure this was conducted rigorously and systematically (Mazurova et al., 2022), the document analysis was conducted in three phases (see Figure 1).

The first phase was the data preparation stage. We imported our selected cases into NVivo, utilising optical character recognition documents, and numbered each case accordingly. As the initial step, we used file classification as organisation and included attributes such as sector, number of employees and country in NVivo. We then created memos for each case, including important notes and details.

The second phase began with code derivation. We used descriptive coding by assigning a concise phrase to summarise the data, which is considered a suitable method for document analysis (Saldaña, 2021; von Richthofen et al., 2022). During open coding, abstract conceptual labels were assigned to common data that capture the characteristics of AI, KM and the organisation. The open codes (Strauss and Corbin, 1990, 1998) are an informant-centric approach in which codes are built directly from the data or through a ground-up process. The objective of this stage is to develop a comprehensive set of codes that facilitates the description of the data (Vollstedt and Rezat, 2019), ranging from single terms to multiword expressions (Flick, 2009). The line-by-line coding method was employed to remain close to the data during analysis (Charmaz, 1996; Gibbs, 2007) and to minimise coding biases (Bytheway, 2018) associated with researcher assumptions and subjective perspectives (Gibbs, 2007).

To avoid the inherent negativity of single-researcher bias in qualitative research, we used inter-coder reliability (Krippendorff, 2018). We began with 20 cases, conducted individual coding by two independent researchers, and continued to 45 cases until we achieved inter-coder reliability greater than 75%. The resulting codes were then compared, and any discrepancies were resolved through discussion and deliberation. This process not only validated the coding structure but also led to the refinement of code definitions, ensuring a high degree of interpretive consistency throughout the analysis. The insights from this reliability check were then applied to the coding of the full data set.

The third phase involves continuing to create open codes for the remaining cases while continually comparing them against the already created codes. Once the open codes were completed, we then created the axial codes by identifying emerging themes (Strauss and Corbin, 1990, 1998). The axial codes focus on examining and developing relationships within and across the categories (Creswell et al., 2007; Strauss and Corbin, 1998; Vollstedt and Rezat, 2019). This stage is driven by the researcher’s interpretation (Gioia et al., 2012; Mazurova et al., 2022), which enables us to connect the data to our research question. We focus on the data through a knowledge lens as we create axial codes.

Finally, we derived the selective codes (Strauss and Corbin, 1990, 1998). This stage helps to formulate a theory (Creswell et al., 2007) and is explained as the process of synthesising and further developing categories (Strauss and Corbin, 1998; Williams and Moser, 2019). This stage will detect the core category, thereby allowing the research question to be answered (Vollstedt and Rezat, 2019). To create the aggregated dimension, we focus on axial codes as a knowledge-AI integration process. Phase three was a non-linear process (Williams and Moser, 2019), where repetition of mapping allowed stable selective, third-level codes.

See  Appendix for samples of open codes, mapped into their corresponding axial codes. We then interrogated the axial codes, keeping our primary objective of this research in mind to understand and identify how organisations meaningfully integrate AI into operational practices in light of existing organisational KM. Therein, using KM theory as the sensitising device (Walsham, 1995), we developed four axial codes, akin to the KM process described in Alavi and Leidner (2001).

We observed that opportunities for AI were not framed as a way to replace human experts. Rather, they were consistently identified in contexts where deep human expertise was already present but hindered by significant operational frictions that prevented its effective application.

Firstly, we observed that human expertise is represented by diverse professionals, ranging from clinicians to manufacturing workers who identified the opportunity of AI integration. In case 001, “That led us to explore how technology can give us a better indication of how patients will fare so we can have more meaningful conversations with them and better plan for their surgery.”

Secondly, organisations faced bottlenecks from tedious manual data handling. For instance, health care providers were overwhelmed with processing “With around 13 million faxes still exchanged each year in medical facilities” (Case 002) and “the knowledge needed to properly carry out these critical functions was scattered across physical documents, digital files and the minds of experienced employees.” (Case 058), making critical knowledge slow and difficult to access. This manual work was not only inefficient but also was a primary driver of professional burnout, as witnessed through “In many healthcare practices and facilities, physicians and staff spend hours each day sorting through paper documentation related to patient health records. This keeps them from spending as much time with patients as they would like and leads to feelings of burnout” (Case 002).

Thirdly, we identified a pattern of cognitive overload, where the sheer volume or complexity of information made it difficult for experts to perform their core tasks. This was evident in clinicians, “Studies suggest that family physicians spend over 17 h a week on administrative tasks such as reviewing records and taking notes. That is equivalent to two full days a week spent on paperwork, rather than patient care” (Case 043). Furthermore, in Case 026, legal teams are facing “time-consuming guesswork.” In these situations, the cognitive load of finding and synthesising knowledge was detracting from the expert’s ability to apply their judgement.

Finally, knowledge friction manifested as limited access to specialised expertise. This was a critical issue in contexts such as providing “Access to mental health is exceptionally low in India. The gap between people needing care and available treatment is one of the highest in the world, estimated to be as high as 80%” (Case 004) or securing a timely “specialist review” for a complex cancer diagnosis (Case 005). In each case, the core opportunity for AI was not to automate expertise, but to build bridges across these points of friction, allowing latent human knowledge to be surfaced, structured and scaled.

We identified that both tacit and explicit knowledge have been selectively externalised into machine-readable form, which we categorised as knowledge extraction. Our analysis consistently showed that organisations did not simply connect AI to their entire data estate. Instead, they made deliberate, expert-led decisions to codify only knowledge relevant to decisions. This selective approach was crucial for training and localising pre-built AI models using high-quality, valid data.

The first category consists of the Domain rules we coded. This includes the formal, often documented, rules and procedures that govern specific tasks. For example, to automate a claims process, one organisation extracted its “policy number recognition feature”, which contained “There are around 40 different templates here, which translate into approximately 80,000 policy numbers” (Case 017). In a health-care setting, this involved codifying clinical guidelines, such as building a system where a “bedside nurse gets an alert when a patient is meeting or triggering those criteria for early sepsis” (Case 075), to ensure the AI’s logic was grounded in established medical protocols.

The second category is Contextual knowledge. This refers to the specific, nuanced information surrounding a particular case or individual, which is often required for personalisation. We saw this in the extraction of a patient’s “account history, including notes, emails and conversations” (Case 027) to inform a customer service bot, or the use of “conversation, medical records” (Case 043) to generate personalised care plans. This knowledge was often less structured and required capturing practitioners’ implicit understanding.

The third category we identified is visual or narrative knowledge. This encompasses unstructured data that requires advanced processing to be made usable as input. We consistently saw this in health-care settings with the use of “medical images such as ultrasounds, CT or MRI scans” (Case 044) and unstructured “physician’s notes” (Case 013). In other sectors, this included large volumes of textual data, such as “decades of existing research” (Case 035), which were digitised and prepared for AI analysis.

The third identified axial code is knowledge expansion, in which AI is used to synthesise extracted knowledge in novel ways, generating new insights and artefacts. This is not about the AI inventing knowledge from nothing; rather, it is about recombining existing human knowledge at a scale and speed that humans cannot.

The first, and perhaps most powerful, pattern is that AI detects anomalies or correlations unseen by individuals. We observed cases where AI identified significant correlations in data that were not previously known or officially recognised by human experts. The most striking example of this occurred during the COVID-19 pandemic, where “The model flagged the strong correlation between positive COVID tests and patients who had lost their sense of smell and taste. These symptoms were integrated into the MUSC screening algorithm before they were officially recognised as symptoms by the Centres for Disease Control and Prevention (CDC)” (Case 075). This demonstrates a clear case of AI expanding organisational knowledge by detecting a pattern that was not yet part of the established human domain rules.

The second pattern we coded was cross-model knowledge translation. This involved using AI to transform knowledge from one format into another, more usable form. This was most commonly seen in the automated creation of meeting summaries. For instance, in Case 100, “broke the meeting down into sections was really good because the forms that we need to put them on to our system are all sectioned.” This process resulted in “massive time savings when taking notes in meetings, summarising points and actions and distilling large documents into sensible summaries” (Case 012).

The third pattern was AI-driven diagnostic highlighting. In these cases, AI models were used to rapidly scan complex medical imagery and highlight areas of clinical significance for human review. For instance, an AI model was trained to “scan magnetic resonance imaging (MRI) images and quickly locate the prostate”, to assist physicians (Case 005). The AI was also used to “measure critical markers – such as the presence and intensity of receptors within tissue samples to classify based on Gleason grading system – to aid in cancer diagnosis”, thereby highlighting the key data points needed to aid in cancer diagnosis. In these examples, the AI was not making the diagnosis itself, but was instead synthesising complex visual information to direct the attention of the human expert to the most critical areas, “assists physicians in identifying regions of interest” (Case 005), for rapid, accurate analysis.

The final theme that emerged from our data is knowledge stabilisation. We define this as the process by which AI-generated insights are embedded into organisational workflows, creating stable and reusable knowledge assets. In this stage, the AI functions as a form of organisational memory system, reducing the knowledge loss, drift and inconsistency that often occur over time and across different employees. This process ultimately stabilises organisational learning.

Firstly, we observed the creation of structured records that replaced fragmented documents. The AI’s ability to synthesise disparate information created a single, reliable source of truth that could be consistently accessed across the organisation. For example, a legal team moved from “time-consuming guesswork” to “facilitated quick, data-driven decisions” (Case 026). Similarly, a customer service team was able to provide more consistent support by using a virtual agent that offered a “comprehensive, at-a-glance record of a customer’s account history, including notes, emails and conversations” (Case 027).

Secondly, stabilisation occurred through the generation of annotated and summarised histories. The AI created persistent, condensed artefacts that made complex knowledge easier to interpret and reuse consistently. In medical diagnostics, for instance, the AI’s output was not just a temporary analysis but a permanent record of “annotated images” and “summarised cases” (Case 005). This same pattern was evident in Case 012, “massive time savings when taking notes in meetings, summarising points and actions and distilling large documents into sensible summaries.” This process directly reduced the variability in interpretation. By creating a single, authoritative summary, organisations ensured that employees worked from the same understanding. This led to tangible benefits, such as “getting more consistent answers from live agents because they’re using the same tool to research client questions” (Case 022).

The culmination of our analysis is a four-phase process model that explains how organisations meaningfully integrate AI into their operational practices (see Figure 2). This model represents the aggregate dimensions that emerged from our axial codes, the dynamic process through which organisations convert latent human expertise into stable, actionable and trusted organisational knowledge by leveraging AI. The model is not strictly linear; a feedback loop from the final phase illustrates that the process is iterative, driving continuous organisational learning. In the following sections, we detail each of the four phases.

While the process is presented in phases for analytical clarity, in practice these activities are deeply embedded within organisational workflows and often occur in parallel. AI enables knowledge to be continuously evaluated, updated and operationalised in near real-time, thereby integrating KM directly into business processes rather than positioning it as a separate or delayed function. In this sense, AI-enabled KM becomes inseparable from business process management, as knowledge is generated, validated and applied within the flow of work rather than outside of it.

Our analysis reveals that AI-enabled workflows vary along a spectrum from augmentation to full automation. While some contexts enable near-complete autonomous execution, others deliberately retain human oversight due to risk, uncertainty or regulatory requirements. Therefore, autonomy is not treated as a binary capability, but as a context-dependent organisational choice.

The process begins not with a technology push, but with a human-centric recognition of ‘knowledge friction’. This is a critical starting point, as it frames AI adoption as a solution to existing organisational pain points where valuable expert knowledge is trapped, inaccessible or slow to deploy. It begins with identifying a bottleneck, whether this takes the form of cognitive overload, manual processing demands or constrained access.

In response to this friction, organisations engage in a process of ‘selective externalisation’. This phase is crucial as it counters the naive assumption of simply “connecting the AI to the data.” Instead, it is a deliberate, expert-led process of codifying only the decision-relevant domain rules, contextual nuances and trusted narratives. This act of selective curation serves as the critical input that grounds the AI model in the organisation’s unique context, ensuring its outputs are relevant and valid.

The third phase, ‘Synthetically expand knowledge through AI’ acts as a tool for combinative synthesis, amplifying the selective knowledge it was fed. This expands organisational knowledge by identifying novel patterns, translating knowledge across different formats (e.g. from transcripts to summaries), and highlighting critical insights at a scale beyond human cognitive limits. This is a process of knowledge amplification, not replacement.

The newly synthesised knowledge is then consolidated into stable, reusable organisational assets or artefacts during the fourth phase, ‘Consolidate knowledge into stable artefacts’. This is the phase in which AI functions as organisational memory. It transforms outputs into persistent assets such as structured dashboards, annotated records and consistent summaries. This process reduces knowledge drift and interpretation variability across the organisation, effectively stabilising organisational learning. The final stage is the continuous calibration of trust through (human) verification. This is the essential socio-technical interface that makes the entire process sustainable. By embedding human oversight or non-human validation, the organisations actively manage AI’s inherent risks, such as hallucination. Crucially, this phase creates a feedback loop. The insights gained during verification inform the ongoing recognition of new knowledge frictions and refine the selective externalisation process for the next cycle, driving continuous improvement.

While our process model explains the sequence of AI integration, we conducted a further cross-case analysis to understand the different ways organisations integrate AI. To do this, we employed a mixed-strategy approach as described by Miles and Huberman (1994), using a meta-matrix of our cases (rows) and second-order themes (columns).

Firstly, a variable-oriented analysis, which involved reading the matrix vertically, revealed two key dimensions of variation:

1. Consolidate knowledge into stable artefacts: We categorised it as human-verified if human intervention was observed; otherwise, it was categorised as non-human-verified.

2. Externalise selective expert knowledge: This dimension defines the source of knowledge that the AI system operates on. We distinguish between two processes: vertical and horizontal. Vertical knowledge processes involve adding new knowledge or data sources. For example, in Case 001, the organisation enhanced its surgical predictive process by integrating “blood test results”, a previously unconsidered data source. During our analysis, we categorised any instance involving the integration of new data sources as a ‘vertical knowledge’ process. In contrast, horizontal knowledge processes primarily operate on existing, codified knowledge without introducing new knowledge. For instance, case 002 was designed to classify the same “faxes” the organisation was already processing; no new knowledge or data source has been introduced. In similar instances, we categorised them as ‘horizontal knowledge’ during our analysis.

Next, a case-oriented analysis, which involved reading the matrix horizontally, revealed our third dimension:

(3) The workflow configuration: This describes how AI integration happens as a process. Augmentation describes workflows where AI enhances or modifies an existing process with new capabilities, while automation focuses on automating an existing process.

Through cross-case analysis, we identified four key patterns in which AI has been successfully integrated (see Figure 3).

A process characterised by a human-verified, vertical, augmentation workflow. In this pattern, AI is used to enhance an expert’s decision-making by synthesising new data sources to create novel predictive insights.

A prominent example of this was observed in a health-care setting (Case 001), where an organisation sought to improve surgical outcome predictions. The knowledge process was vertical, as new knowledge in the form of blood test results was integrated into their predictive models. This workflow was one of augmentation, as it fundamentally changed the process from providing general prognoses to offering granular, data-driven risk assessments. The AI revealed that “platelet count was one of the most important predictive factors for the success of an operation,” an insight that was “really surprising” to the clinicians.

Crucially, human intervention was high, following a strict human-in-the-loop model. The AI did not autonomously diagnose patients; rather, it provided a risk assessment that clinicians then had to validate. As noted in the case, the final step was human-led, as clinicians use the tools to “understand, refine and explain” the outcomes. This pattern demonstrates a clear shift in the expert’s role: away from manual data processing and towards high-level contextual validation, where they are responsible for translating statistical correlations into meaningful patient care.

The second pattern we identified is defined by a human-verified, horizontal, augmentation workflow. This pattern uses AI to streamline routine knowledge work by synthesising existing information into new, structured artefacts. This horizontal process focuses on the efficient transformation of available internal knowledge, such as emails or documents.

Case 012 exemplifies this approach, where the AI is deployed “to compile information in my emails into timelines for the situation briefing reports I need to write. I have trained it to format the information into the right structure for an SBAR, saving hours on every report.”

In this workflow, the AI acts as a structural synthesiser, converting unstructured data (a collection of emails) into a codified and usable format (a structured report). However, the human remains the critical “knowledge consolidator.” As the reports are used for high-stakes briefings, the user must verify the AI-generated timeline for accuracy before finalising it. The value of the human-in-the-loop is therefore efficiency with integrity. The human expert shifts their effort from the low-value, laborious task of compilation to the high-value task of validation, ensuring the final output is both generated rapidly and is factually accurate.

The third pattern leverages the organisation’s existing knowledge base to fully automate high-volume routine tasks, while retaining a critical human verification step to ensure data integrity. This approach strikes a balance between the efficiency of automation and the need for oversight, particularly in regulated environments where accuracy is crucial.

Case 002 illustrates this through the automation of manual fax classification, where an AI tool is designed to “classify patients’ health records” and route them to the correct file. The solution, “The tool quickly scans each fax and determines what type of document it is and to which patient it belongs. So far, 427 practices have implemented this solution and scanned nearly 2.2 million faxes, saving up to one minute per incoming fax. The solution accurately identifies and matches patients for 75%–85% of incoming faxes.”

In this workflow, the AI acts as an automated classification engine. Its primary function is not to create new content, but to accurately sort and route existing information based on pre-defined organisational rules. While the AI achieves a high accuracy rate, the remaining cases represent a critical failure point. Therefore, human intervention is structured as an exception-handling and quality assurance. The human role is not to perform the task itself, but to manage the cases and errors that the AI cannot resolve. The value of this pattern lies in achieving massive scale with managed risk. The organisation automates most tasks while applying targeted human expertise to safeguard process integrity, where a single misfiled document could generate significant consequences.

The fourth pattern represents the complete algorithmic execution of established workflows. It occurs in contexts where the organisation’s existing knowledge is well-structured and codified that the AI system can perform the entire process with high accuracy, shifting human involvement from direct verification to periodic monitoring and system oversight.

We observed this pattern in highly structured, high-volume decision-making processes. For example, in the financial sector, a credit approval process that “Previously, an internal team would analyse credit applications within 48 h” was fully automated, with the AI performing the analysis “analyses are performed automatically in seconds, allowing customers to receive immediate credit approval” (Case 021). This was possible because the domain rules for credit assessment were explicit and could be reliably executed by the AI.

This pattern was also evident in the automated generation of legal documents and the extraction of contract clauses (Case 026). The AI was able to “The solution also revamped the way agreements are managed. Once the required fields are filled, the system auto-generates legal documents. Third-party agreements are easily uploaded, marked up and disseminated to relevant stakeholders. Artificial intelligence (AI) automatically identifies crucial contract clauses, such as termination dates and notice periods, making post-signature management more efficient.”

In this workflow, the system operates autonomously on a case-by-case basis because the organisation has a high degree of trust in the codified knowledge and the AI’s ability to execute it. The value of this pattern lies in its transformational efficiency. It doesn’t just make an existing process faster; it fundamentally changes the operational model, enabling outcomes like immediate credit approval that were previously impossible.

The persistence of human verification in patterns 1 and 2 should not be interpreted as a technological limitation, but rather as a governance mechanism to ensure reliability, accountability and trust in AI-generated outcomes, particularly in high-stakes environments. For example, in domains such as health care, even when AI identifies potential treatment adjustments in real time, implementation may still require human validation due to clinical governance protocols and ethical considerations. Rather than assuming full autonomy as the default end-state, our findings show that organisations actively configure the balance between automation and human oversight as a strategic design choice.

This paper presents empirical findings from a rigorous grounded theory analysis through a systematic, multi-level coding process, we developed two key theoretical contributions. Firstly, we theorised a four-stage process model that explains the sequential journey of AI integration, from recognising knowledge friction to consolidating knowledge into stable artefacts. Secondly, we developed four distinct KM–AI integration patterns. Together, these findings provide a comprehensive answer to our central research question: How, and in what ways, do organisations meaningfully integrate generative AI into their operational practices?

The need to understand AI–KM integration is amplified in increasingly volatile, uncertain, complex and ambiguous environments. In such contexts, organisations face rapidly changing conditions, incomplete information and heightened decision complexity. Traditional KM approaches, which rely on stable knowledge repositories and past experience, are often insufficient for responding to dynamic and unpredictable challenges. AI introduces the capability to continuously update, synthesise and operationalise knowledge in near real-time, enabling organisations to respond more effectively to environmental turbulence.

This research extends the current theoretical understanding of AI integration by repositioning it as a socio-technical knowledge transformation process rather than a discrete technological event. Primarily, it advances the KBV by theorising ‘Knowledge friction’ as the primary catalyst for algorithmic intervention. This suggests that AI integration is most effective when it addresses specific points of informational drag where human expertise is present but functionally constrained. Secondly, the study offers a critical evolution of the traditional SECI model (Nonaka and Takeuchi, 1995) by demonstrating a shift from purely human-driven externalisation to algorithm-mediated synthesis. By identifying how AI codifies unstructured, tacit interactions into stable, explicit artefacts, we provide a baseline for future researchers to examine how “algorithmic knowledge” is integrated into the permanent organisational memory. Crucially, our findings indicate that as AI becomes more autonomous, the theoretical importance of human verification increases, suggesting that epistemic authority in the modern firm is a negotiated, hybrid construct.

Our findings offer a structured diagnostic roadmap for practitioners, providing a guiding framework for managers to move beyond ad-hoc AI initiatives towards strategically aligned solutions that leverage existing KM. This model guides deliberate decision-making on where and how to deploy AI for optimal efficiency, effectiveness and stakeholder value. Beyond the individual organisation, this research holds significant value for policy contexts. Specifically, our findings regarding human verification suggest that regulatory bodies should focus on establishing industry-standard audit protocols for AI-generated knowledge. Furthermore, national policy initiatives should shift from generic “tech training” to supporting “knowledge curation” skills, ensuring that the workforce is capable of acting as the final, critical “knowledge consolidators” in a hybrid environment.

By demonstrating how AI can specifically reduce administrative burdens that lead to professional burnout, this study provides an evidence-based rationale for public-sector investment in AI to sustain critical service delivery in health care and education. The framework highlights the evolving human–AI dynamic, underscoring the shift from human-driven to algorithm-driven knowledge creation. It informs strategic discussions on the future of work by explaining how AI integration can be architected to augment human capabilities and elevate professional roles, rather than merely replacing them. Ultimately, this research provides the conceptual tools to navigate the transition from friction to function, fostering a more symbiotic human-machine relationship within the modern enterprise.

Like any study, our work has limitations that open up avenues for future research. Firstly, our analysis is based on publicly available customer success stories. While these provide rich narratives of implementation, they may be subject to a positive reporting bias. Future research should employ in-depth, ethnographic case studies to gain a more critical and nuanced understanding of the challenges and failures of AI integration.

Furthermore, our four patterns provide a stable framework, but the underlying technology is evolving rapidly. Future research should investigate how the emergence of new AI capabilities might create new, hybrid patterns of integration not captured in our current model. By doing so, scholars can continue to build a rich and dynamic understanding of this profound technological and organisational transformation.

While the initial funding is heartening, we recognise that there are several limitations of this study. Addressing these limitations in future work can provide further insights to this important area of research. Firstly, the focus on KM initiatives in this study pertains only to the organisational level. We recognise that investigating KM initiatives at the process level could provide further insights as to how AI is being integrated into specific processes.