Artificial intelligence (AI) is currently a major topic in both society and research, not least since the dawn of end-user generative AI in 2022. AI has the potential to disrupt human life, e.g., how we work, live, communicate, and get medical help (Elliott, 2019). One task that can be handled by AI is classifying text by means of machine learning, which can take the form of being either supervised (Kadhim, 2019) or unsupervised (Thangaraj & Sivakami, 2018), or a combination of the two (Wu et al., 2025).

Search and metadata assignment are among the scientific areas that have explored the potential of AI for various text-based tasks (Corrado, 2021; Golub, 2021; Jha, 2023; Mwantimwa & Msoffe, 2025; Vrindha & Syamili, 2025), and within different contexts, such as libraries (Asula et al., 2021), scientific databases (Moulaison-Sandy et al., 2021), academic papers (Yang et al., 2023), university repositories (Lowe et al., 2021), datasets (Chen et al., 2020), and tweets (Samuel et al., 2020).

However, little is known about how AI can assist in text-based enterprise search. Contrary to the examples given above, text-based enterprise information is characterised by being heterogeneous, existing in a plethora of formats, and being stored across a diversity of information systems (Gunadi & Albayrak, 2015; Kruschwitz & Hull, 2017). At the same time, enterprise search differs from information searching in established databases, such as library databases, and on the Internet by having a larger share of people search, shorter queries, and few advanced queries (Domke et al., 2018; Svarre, Lykke, & Bygholm, 2024).

Enterprise search has been associated with various challenges in previous research. Language use is a significant issue. In corporate settings, there are often multiple professional languages involved, and since information for problem-solving is often combined with information from multiple professional groups and corporate information sources, it can be difficult to find a comprehensive set of search terms, due to domain-specific language use, geographical variations, and changes over time (Cleverley & Burnett, 2015; Furnas et al., 1987; Nielsen, 2005; Schuff et al., 2016). Metadata quality is another frequent problem. Stocker et al. (2015) identified barriers related to adequacy and completeness of metadata. Other challenges are related to the information searcher. Freund (2015) identified challenges, such as searchers’ awareness of information and sources, the abundance of information available, access control, and the amount of time available to search and use information.

The purpose of this study was to get insight into how enterprise searchers used and assessed the support provided by a search tool that is based on AI-generated metadata and contextual information from a diverse set of corporate sources linked to the metadata. The study took place in an international biotechnology company that conducts research, development, and production of industrial biotechnological products. Through several user studies, a need was identified for a tool to support searchers with domain-specific, contextual metadata that could guide searchers to relevant information and extend the search across time and academic domains (Lykke, Bygholm, Søndergaard, & Byström, 2022; Svarre, Lykke, & Bygholm, 2024). Using Scibite, an ontology-based text-mining search engine, a set of twelve company-specific, as well as general biotechnological metadata types, was added to the display of retrieved documents in the form of clickable metadata linking to internal and external information sources. The study examined the functionality of (a) the AI-generated metadata and (b) the information that the metadata linked to in internal and external resources.

Specifically, the aim was to get insight into

Here, we first outline the importance of metadata in enterprise search to elaborate on the specific role of metadata in this particular context. Next, we present AI-based solutions for handling metadata to provide a framework for the case presented later. We focus on AI solutions based on machine learning, since LLM-based solutions represent a very different approach in working with metadata. Lastly, we provide studies that have presented evaluations or user studies to set the stage for the potential of AI-based metadata from an end-user perspective.

Because AI applications gain traction in many areas, the interest in AI literacy, i.e. what is required to use these applications in a sensible and productive way, also increases. Within recent years, a number of papers and reviews have been published with the aim of identifying what the concept of AI literacy might involve and how best to afford the competencies concerned. In this paper, we drew on different conceptualisations of AI literacy to understand the challenges employees experience when using AI metadata in their search practice. Focus was on AI literacy for non-AI professionals in a workplace setting.

The concept of literacy, traditionally understood as the ability to read and write, also includes numerical literacy, health literacy, financial literacy, media literacy, and digital literacy, to name but a few. An often-used reference to the notion of AI literacy is Long & Magerko (2020), who mention the need to investigate “what new competencies will be necessary in a future in which AI transforms the way we communicate, work, and live with each other and with machines” ( p. 1), and they refer to this set of competencies as AI literacy. In their review of the AI literacy of employees at digital workplaces, Cetidamar et.al. (2024) distinguish between studies focusing on AI literacy in education and AI literacy at workplaces, and, furthermore, between AI literacy at the individual level and AI literacy at the organisational level. They note that most studies of AI literacy have been conducted at the individual level and around education. In our attempt to understand employees’ challenges and then discuss what can be done to support employees, we need both the individual and the organisational level. The individual level is about the competencies needed for individual members of the workforce to use AI applications in a productive way and the organisational level is about what it takes for an organisation to benefit from applying AI in their business.

An example of a conceptualisation of AI literacy at the individual level is Ng et al. (2021) which, based on a literature review and inspired by Bloom’s taxonomy, proposed four types of competencies to characterise different kinds of capabilities required in interaction with AI applications. Bloom’s taxonomy (Bloom, 1956) classified cognitive skills into six hierarchical levels, with each level building upon the previous one, and is a widely used framework within the educational area to identify learning objectives that target different levels of complexity independently of specific learning domains. The taxonomy has been revised and expanded several times (Krathwohl 2002; Huitt 2011). The first three aspects of the AI literacy concept proposed by Ng et al. (2021) cluster the six levels from Bloom into three, that is 1) know and understand, 2) use and apply, and 3) evaluate and create. With the fourth aspect, ethical issues, they add an ethical dimension to the concept and include issues, such as transparency, accountability, and fairness.

Based on a revised version of the conceptual framework, Ng et al. (2024) also developed a questionnaire emphasising not only cognitive but also behavioural and attitudinal issues to measure student learning outcomes in AI learning programmes. Other authors working within the educational and individual field have also developed questionnaires and instruments to measure AI literacy. What they had in common was that they were based on some core constructs identified through literature review, sometimes supported by various theoretical constructs, expert testimony, and testing with students or citizens. For example, Laupiciler et al. (2023) based their scale for the assessment of non-experts in AI literacy on “technical understanding,” “critical appraisal,” and “practical application.” Wang et al. (2023) used “awareness,” “use,” “evaluation,” and “ethics” as basic concepts in a scale for measuring AI literacy. Pinski and Benlian (2023) aimed to develop a scale for measuring general AI literacy, and, based on concepts drawn from the information science domain, they suggested seven dimensions grouped in three categories, that is AI actor knowledge, AI process knowledge, and AI experience. There were then a number of suggestions as to what should be included in the education of students and citizens in the field of AI and how this knowledge could be measured. Common to these was that knowledge of the technology was not enough; you also need practical experience and skills that allow you to assess the consequences of using the AI application critically.

Moving to the organisational level, the findings may be considered somewhat similar. That is, acquiring AI technology in itself does not necessarily make a difference to an organisation. In spite of the potential that AI technologies offer, organisations are faced with a number of challenges that prevent them from realising the benefits. AI is far from being a plug-and-play technology, as Fountaine et al. (2019) put it. They emphasised that it is equally important to align culture, structure, and ways of working to support AI adoption. Thus, there is a need for understanding what such an alignment might involve, that is how organisations build AI capability. Mikalef & Gupta (2021) aimed at identifying the necessary organisational resources that will enable a company to build their AI capabilities. They drew on resource-based theory, the basic idea being that “the building of resources facilitates the formation of organisational capabilities, which, in turn, drive performance gains” (Mikalef & Gupta, 2021, p. 2). Building on theoretical insight from resource-based theory, as well as studies outlining the challenges related to AI adoption, Mikalef & Gupta (2021) proposed eight resources that allow organisations and companies to obtain AI capabilities. Using a classification scheme from the resources-based theory domain, the eight resources were grouped in three main categories: 1) Tangible resources, including data, technology, and basic resources, 2) Human resources, including technical skills and business skills, and 3) Intangible resources, including fostering inter-departmental coordination, organisational change capacity, and risk proclivity. The authors underscored that although both tangible and human resources are necessary and a prerequisite, the intangible resources are extremely important for obtaining AI capabilities. These are characterised as more difficult to identify and also more unique because they are developed through a mix of history, people, and processes that characterise the specific organisation.

Cetidamar et al. (2024) focused on employees and aimed at developing a model of the competences needed to utilise AI technologies for achieving company goals. The idea was that strengthening employees’ AI literacy would enable them to participate in the design and use of AI technologies. Furthermore, they stated that AI literacy is “...an organisational-level capability where individual capabilities add up to an organisational level strength.” (Cetidamar et al., 2024, p. 810). Thus, they argued that employees’ collective AI literacy builds an organisational capacity. Based on a literature search and a bibliometric analysis, they suggested that AI literacy for employees should consist of four capabilities, that is technology-related, work-related, humanmachine related, and learning-related.

To sum up, AI literacy can be considered from an individual as well as an organisational perspective. The individual view focuses on employees’ capabilities in using and interacting with AI-based systems. Although there may be overlaps, the corporate view differs in including how organisations build capabilities and company goals, and how employees can be included in the design and use of AI. Together, the two perspectives offer a complete view on human AI capabilities in workplace contexts.

The study took place in an international biotechnology company that conducts research, development, and production of industrial biotechnological products and biopharmaceutical ingredients. The search tool was developed to support search for internal information available in a corporate research archive. The digital research archive was established in the 1980s and contained internal documentation about the company’s products and R&D work, e.g., lab test reports, memos, guidelines, and project reports. In general, the research archive was used to find information about previous R&D work and historical information about, for example, products, projects, techniques, and industry applications. Specifically, it was used by the searchers to find new application areas for specific products; look up previous results for certain test techniques or materials; or experience with the application of a product in a specific industry.

The documents were uploaded to the research archive, and metadata were registered by the authors. The registration was based manually assigned metadata: document type, title, author(s), project number, industry, product name, external relations, and internal source. It was also possible to attach related documents in the form of PDF files. Access restrictions were specified by indicating allowed readers or groups. The metadata document category (for example, author, project number, industry, and product names) were registered using controlled drop-down lists. The research archive was accessible by the enterprise search system, which was based on SharePoint and, in addition to the research archive, provided access to the corporate information sources intranet, people directory, news, quality documentation database, laboratory system, and social media. The research archive could be searched as part of the enterprise search interface, and via an independent search interface. The search interface facilitated commandbased keyword searching and filtering by metadata.

The development of the search tool was initiated in 2020, where a local interview study with eight informants showed that the employees from R&D had difficulty both searching for and assessing the relevance of retrieved documents. The metadata registration was optional and, therefore, all metadata categories were not necessarily linked to all documents. The document titles were not always designed appropriately and contained neither relevant search terms nor words that could be used to assess the content and relevance of the document.

The findings of the interview study were supported by a larger study of enterprise searching practices across different work areas and work tasks in the enterprise search system (Lykke, Bygholm, Søndergaard, & Byström, 2022). The main finding of the study was the identification of an explorative, tracing search technique, where searchers, by use of historical and contextual knowledge, interactively tried out different search paths through the enterprise search system and its sources to put together a set of information that jointly provided the needed information. The needed contextual and historical information could, for example, consist of information about previous research projects or laboratory results or about earlier business application areas where an enzyme has been used or tested. The searchers did not always possess the necessary historical or contextual knowledge and often had to take a detour by consulting a colleague to get the needed information to continue the search and find the desired information for the search task (Svarre, Lykke, & Bygholm, 2024). Thus, an important finding was the need for a tool to support searchers with specific contextual and historical information about, for example, research projects, lab results, and business application areas to support the tracing search strategy.

Based on these results, the company chose to use machine-learning categorisation to add highly contextual metadata to retrieved information, to make it easy for the searchers to orient themselves quickly in document content and information relevance and to guide the searcher through the added metadata to the needed contextual and historical information to extend the search strategy. By using Scibite, an ontology-based text-mining search engine, a set of twelve company-specific, as well as general biotechnological, metadata types were added to the display of retrieved documents in the form of clickable metadata links (see Figure 1).

The twelve metadata categories were selected by information specialists in collaboration with the vendor (Scibite). Their selection was informed not only by their perceived importance for internal search processes, but also by considerations of practical applicability. One information specialist characterised this approach as “the art of the possible”. Consequently, certain types of information recognised as valuable for contextual searching, such as personal names, were excluded from the metadata schema. Within the organisation, individuals are identified primarily by their initials, and the inclusion of such identifiers would have generated an excessive number of non-discriminatory results. By contrast, entities, such as product names and enzymes, proved more amenable to systematic organisation and were, therefore, more readily incorporated into the metadata categories.

The aim was, through machine-learning assigned metadata, to highlight information from the retrieved document that was important both in terms of relevance assessment and formulation of search queries. When the searcher clicked on the assigned metadata, different types of either internal or external information were displayed depending on the metadata category. The added metadata led, for instance, to internal lab data reports and results for lab study identification numbers appearing in the retrieved document, to taxonomic information about organisms mentioned in the retrieved document from the external NCBI National Center for Biotechnology Information database, to project information and application areas from the internal R&D project database when a project name and/or number was mentioned, or to predefined Google searches for companies whose names were found in the retrieved document. When a searcher clicked on the added metadata for, for example, the ELN lab test database, first a list of related tests was displayed. When the searcher clicked on one of the test numbers, information about the test was shown. The type of information shown was different for the twelve metadata categories (see Table 1). Some information came from internal sources, i.e., the lab database or the corporate product database, and from external sources, i.e., well-known nomenclatures or taxonomies, such as the NCBI. The choice of associated sources and information was decided by the corporate information specialists.

The development of the tool was carried out in collaboration with Scibite (https://www.scibite.com/), which also provided several of the taxonomies that formed the basis for the identification and assignment of metadata. In practice, however, it turned out that these external generic taxonomies, developed to represent a professional domain in general, were not sufficiently specific in relation to the company’s use of and approach to categories. Therefore, internally developed taxonomies were primarily used, and even these had to be developed to achieve a sufficient quality in assignment of metadata.

We used an exploratory sequential mixed-methods design (Clark et al., 2021) to study the use of and challenges related to the search tool. First, we reviewed the AI-based enterprise search tool with the purpose of gaining insight into the twelve types of AI metadata, with particular emphasis on the kinds of information they lead to. We also focused on interface design and functionality in interaction with the AI metadata, specifically investigating the visibility, accessibility, understanding, and usability of the metadata types and their associated information. The study was conducted between 15 January 2022 and 31 December 2022.

This review was followed by an interview with the information specialist responsible for the development of the AI metadata tool. The interview was semi-structured and aimed to illuminate the background and purpose of the AI metadata, the development process, and existing user experiences with the tool. In addition, during the development process, the development team held informal conversations with eight stakeholders, consisting of a mix of internal employees working with digitalization, enterprise systems, and R&D tasks.

Next, we collected qualitative data on concrete end-user engagement and experiences with the AI interface, metadata, and associated information. Based on the findings from the qualitative interview study, we then designed a questionnaire study to assess the range and generalisability of the qualitative findings.

The end-user interviews were also semi-structured interviews. They opened with questions about participants’ jobs and organisational positions, their work tasks, and their use of the research archive in their work. Participants were then asked to carry out two recent searches, while explaining in detail their search moves and considerations. We were particularly interested in how they used and assessed the information provided by the AI metadata and the related information sources. Questions concerning the interviewees’ affiliations, work practices, and search tasks were designed to elucidate the context in which the AI search tool was employed. The subsequent structured questions were informed by established insights into challenges associated with enterprise search, as well as by the AI literacy literature concerning key competencies in the use of AI technologies. Open-ended questions emerged from observations of, and discussions about, the interactions that occurred during the searches undertaken by the interviewees as part of the interview. The interviews concluded with a general discussion of the role and usefulness of the semantic AI tool.

Interviews took place via MS Teams, with screen sharing enabled while participants carried out the two search tasks. Interviewees were located in their offices at the case company, while the three researchers were co-located in a conference room at the university. The interviews were audio-recorded and subsequently transcribed.

Nine people participated in the study (Table 2). The interviewees were sampled on the basis of two criteria. First, they were required to have used the newly developed AI search tool and the twelve AI metadata elements. Second, the sample was constructed to ensure collective representation of the organisational units and search tasks for which the tool was designed. Sampling and data collection were discontinued at the point at which additional interviews no longer yielded novel insights or emergent themes. Overall, participants worked with research and development of enzyme products (six participants), product safety (two participants), and product quality (one participant). It is important to note that the interviewees possessed limited experience with the newly introduced AI tool, since it had been released only one month prior to the commencement of the study.

The transcriptions were read and coded inductively by the three researchers. Initial open coding was conducted individually, after which the researchers discussed and consolidated the identified themes. In addition to identifying themes, the discussion aimed to determine which themes should be investigated further through the questionnaire.

The questionnaire was distributed to 297 research archive users. Participants were selected based on search frequency, with employees who had conducted the highest number of searches during the previous three months invited to respond. The questionnaire consisted of thirty-eight questions, including opinion and assessment questions using ordinal scales, as well as background and factual questions concerning, for example, types of searches conducted in the archive, search frequency, years of employment, and organisational division and function.

Prior to answering the questionnaire, respondents were asked to view a one-minute video introducing the functionality of the AI metadata. This introduction was included, because the research archive search interface displayed two types of metadata. On the left side of the interface, metadata used to filter search results was displayed, while the machine-learning-based metadata was shown on the right side. During the interviews, it became clear that several participants did not distinguish between these two types of metadata. The introductory video was, therefore, included at the beginning of the questionnaire to ensure that respondents understood what was meant by AI metadata.

Seventy users completed the questionnaire in full, corresponding to a response rate of 24%. The questionnaire was prepared and the data collected using the online survey tool SurveyExact. Subsequently, the data were analysed using Excel. Uni- and bivariate statistical analyses were conducted to examine distributions and correlations.

We start the presentation of findings by describing how the searchers assessed the usefulness of the AI metadata and the associated information in their search activities. Hereafter, we zoom in on user engagement and on what characterised the searchers’ interaction with the AI metadata and related information. We present their specific search activities and the challenges that characterised their use of the semantic tool. The study findings are presented in an integrated manner, combining results from the initial qualitative interview study involving nine users and the subsequent quantitative questionnaire study, completed by seventy respondents. Participants in both samples were selected to represent the groups of information searchers for whom the AI tool is intended to provide support in their search.

The aims of the study were to gain insight into how workplace searchers used AI-generated metadata and the information associated with it, and to identify literacy challenges in user–system interaction to explore how searchers can be supported. Cetindamar et al. (2024) distinguishes between AI literacy at the individual and organisational level. However, the principal insight derived from the conceptual framework is that experience in the practical use of technology and knowledge is essential for the development of AI capabilities at both the individual and organisational levels. In the discussion, we focus on users’ experiences with AI-generated metadata, examining both how such metadata supports and enhances search processes, how it presents challenges for information searchers, and how they may best be equipped to utilise AI-generated metadata effectively.

The present study constitutes a case analysis of an international biotechnology company, drawing upon semi-structured interviews with the corporate information specialist responsible for the development of the AI tool under investigation, a questionnaire study with seventy searchers for whom the tool was developed, and nine semi-structured interviews, including demonstrations of two to three searches conducted using the AI tool by searchers with prior experience of it. The AI tool had only been implemented for one month; consequently, users’ experience with the tool was limited. The sampling of interview participants was discontinued at the point at which no further novel insights emerged.

Overall, the study has limited generalisability; however, it provides new and valuable insight into how a group of highly experienced information searchers, possessing substantial domain and search expertise, interpret and utilise highly specific, contextual metadata and information to expand and reformulate their queries in pursuit of information that is at once precise yet previously unknown. The findings may be considered in relation to similar contexts, in which experienced information searchers are introduced to and interact with new search technologies. Future research should, therefore, seek to examine this particular user group in greater depth and across comparable contexts.

The aim of this study was to investigate how workplace searchers understood and used an AI-based search tool consisting of domain-specific metadata assigned to retrieved documents. Users’ assessments and interactions with the tool were examined through interviews with search demonstrations and a questionnaire survey conducted as part of a case study in an international biotechnology company.

The informants found the twelve AI metadata types useful for both exploratory search and specific searches for known information. In terms of user–system interaction, the tool was generally perceived as easy-to-use. However, users also encountered irrelevant and incorrect information, which they found annoying but understandable given the complexity of domainspecific relationships. Several challenges were identified: domain and search knowledge were a prerequisite for effective use; the use of metadata in search was not always readily apparent; there was a need for improved communication of what metadata represent and how they are utilised in search; and new routines were required within established search practices.

During search activities, the interviewees demonstrated the four AI competencies identified by Ng et al. (2021), drawing on existing search experience and understanding of search algorithms, including AI. They were able to understand, apply, and evaluate the AI metadata critically, identify problems, and suggest improvements. The study underscores that users with extensive search experience and deep domain knowledge can use AI-based search systems with the necessary critical approach. Moreover, effective use requires an open-minded, explorative attitude and a willingness to continuously interact with the system to achieve accurate and relevant search results.

The development of AI-related competences can be facilitated through the strategic alignment of tangible, intangible, and human resources, encompassing instructional materials, interdepartmental coordination and organisational change capacity, as well as the ongoing cultivation of advanced search skills and domain-specific knowledge in information-seeking practices.

The insight into how experienced corporate users adopt a novel AI-based search tool—by drawing upon their existing search expertise and substantial domain knowledge, while concurrently extending their overall information-searching literacy to encompass AI-based search technologies—constitutes the principal contribution of the study. The study further provides important insights and practical implications regarding how experienced searchers may be supported, for example, through video-based introductions, explanatory features, and potentially chat-based dialogue integrated within the search interface.