Generative AI at work: mundane and productive articulation

Source: Ida Larsen-Ledet (University College Cork), CHI 2026 · 2026-04

This paper proposes a distinction between productive and mundane articulation in people’s interactions with generative AI. While generative AI produces useful results for certain tasks and domains, for many types of work, experiences with AI are still dominated by generic and superficially relevant responses. People are often required to provide context and clarifications through several iterations to get usable output. This paper argues that generative AI will not become a true asset in the workplace unless it allows people to focus on productive articulation that helps them reflect on and refine the objectives of their work, by reducing the need for mundane articulation that makes up the rote work of setting up fundamental context for generative AI to deliver useful results.

Introduction. 1 Generative AI at Work Generative AI (genAI) seems to be the new medium for knowledge consumption and use. With generative AI, tools that draw information from multiple sources and combine it on the spot to present to a user are becoming commonplace [43]. So far, however, the user experiences they offer remain fairly rudimentary. They are good, or at least reasonable, at providing well-scoped responses to well-specified questions or prompts. But the more subjective or specialized the request, the more evident the generic nature of genAI becomes. At first glance, the value of genAI comes from its ability to summarize, to generate, and to fill in gaps1. But the real value and, in particular, the novelty of genAI comes from it being responsive – that is, generating things that are suitable for a purpose or a situation [43, ch. 4]. Without the summarization, generation, and gap-filling being relevant, they would not be as useful, at least not for most work activities.

1.1 AI as Part of Knowledge Infrastructures Recent years have seen the uptake of organizational knowledge bases that are automatically constructed and managed with the aid of machine learning and AI [e.g., 18, 19, 45], and the introduction of generative AI to workplace software is providing new ways of utilizing these knowledge infrastructures. Both of these developments have enabled leaps from passive knowledge stores to active surfacing of knowledge [44], bringing us closer to the goal of “connect[ing] knowledge workers with the right set of knowledge resources or people, at the right time” (O’Dell and Davenport as quoted by Jarrahi et al.) [18, p. 97]. GenAI, in particular, is changing the use and reuse of knowledge in organizations by changing the way it is surfaced [43]. The vision seems to be that genAI will be integrated in all manner of workplace applications, or at least office applications ranging from data analysis to word processing and meeting software. With this, knowledge that previously sat much more passively in knowledge bases, mainly to be consumed, may now more actively influence people’s work as it happens. By (organizational) knowledge infrastructure, I refer to knowledge storage, management, sharing, and (re)use of knowledge through knowledge bases, information retrieval, and expertise finding. I use the word infrastructure to denote the interplay between multiple tools, technologies, and practices [23, 33] and how they enable knowing (the reuse and application of knowledge) [1, 3, 28]. I use the term knowledge infrastructure rather than organizational memory [4] to describe the combined technological, practical, and social means for organizational remembering because technology in itself is not memory but an infrastructural enablement of memory as something active and constructive [5]. That is, I am not talking just about (passive or active) knowledge bases but about technology that facilitates the human acting involved in remembering [4] and using knowledge. This is what good genAI should do – not replace human memory, interpretation, or knowledge use but facilitate them as action. Information in an organization often has multiple potential use scenarios, and there is lots of it. Employees constantly assess what information to pay attention to, what to make active use of, and what to disregard – in other words, what is relevant to them and to what they are currently doing. As genAI selectively synthesizes information for people, how can those people be confident that they are actually getting to (all of) the knowledge they need and want for what they are doing? Even beyond information finding specifically, genAI’s connection with and place(s) within knowledge infrastructures are important, as all of genAI’s capabilities are tied in with information through regurgitation and remixing of it. This paper discusses what it means to situate AI in people’s work, as genAI applications promise to weave knowledge surfacing and remixing into the fabric of knowledge work.

Related work. People take to generative AI and its ability to select and summarize as a way to reduce the overhead of finding or consuming information [35, 47]. But it often misses the mark because of mistaken or lacking assumptions [e.g. 22]. This shows that being over-confident in assumptions about the user leads to poor or less-than-useful user experiences. Although a frequent assumption is that genAI enables people to spend more time on “core” work activities and less time on articulation work and tedious rote work [e.g. 17, 35, 46] (which, arguably, might depend on how closely your work lends itself to what current genAIs do well), being productive with genAI requires people to skillfully articulate themselves in the right way, for instance specifying details that would be implicitly understood by a coworker [35, 36] (e.g., that a slide deck should follow the institutional template), and proposed solutions have so far hinged on people doing this explicitly [e.g. 20, 21].

2.2 Relevance The notion of relevance is essential to organizational knowledge infrastructures, as information only becomes usable knowledge when that information is meaningful to a person and a situation [8]. When viewing organizational knowing as emergent from the situated skillful acting of people [28], the connection to person and situation also depends on what that person is doing with the information. As such, information’s status as knowledge depends on how relevant the information is to the person and their activity. In the fields most often interested in studying information relevance, such as information retrieval, relevance was originally conceived as topicality: in essence, the match between terms in a search query and terms in a search result [30]. It has long since become clear that relevance is contextual [29], and that not all factors pertinent to relevance can be codified [29, 30]. With this in mind, when we consider genAI as machine that relies on textual semantics, the experience of surface-level relevance becomes less surprising. Its textual operations may be quite advanced, but the effects of primarily topical relevance are still felt [21, 35].

2.3 Personalization and Context Personalization for information consumption and knowledge use can consist both in adapting the information that is provided to people and in adapting the way that information is presented and the available modes for interacting with it [9, 27]. This adaptation will be based on information about the person and/or their situation [e.g., 9, 14]. Underlying many such proposals is the premise that what is explicit – often a search query – is not enough for determining what a person needs (see also contextual relevance above) [40]. Teevan et al.

Method. 3 Situating AI in People’s Work So, context and clarification help genAI produce better results. But so far, attempts to build clarification into genAI applications through explicit input fail to distinguish between things that one would expect to need to specify and things that should be implicitly understood by an “intelligent” agent. For genAI’s output to feel not only surface-level relevant but also personally and situationally relevant, we need to distinguish between articulation that is a productive part of work and articulation that is simply mundane. This is not just about who the person is or what information they need but also what needs to happen. Usefulness depends on how fit for purpose the possibilities for action are. Current modes of (inter)action in most genAI tools are relatively basic, as are the places they appear in: Most of them are restricted to conversational interfaces [43, ch. 4] or, at best, confined to making things happen in a single app. This does not match how people work [6], and it effectively decontextualizes use by siloing the AI. Interaction with genAI should be contextualized in relation to people, activities, and collaborative settings, not just in relation to whatever app is currently in use.

3.1 Mundane and Productive Articulation To deliver the most useful response, the system or tool should not only understand the explicit content of a request or statement but also the intent behind it [40, 43]. A request such as: “who should I ask about the Planning department’s procedure for new build surveys?” should produce different responses depending on who is asking and what situation they find themselves in. However, many current genAI tools seem to operate with a superficial notion of context that relies on explicit information such as the user’s organizational or physical location, or what software application the given instantiation of genAI “lives” in (or simply the preceding utterances in a conversational interaction [43, ch. 3]). If genAI systems do not become capable of genuine contextualization, people will continue to be forced to articulate the most mundane aspects of their situation before getting to anything useful. Many readers have probably tried the back-and-forth it takes to get genAI tools to carry out a request in a way that is right for them. A lot of this involves getting the tool to “understand” you and any relevant circumstances [36]. We can draw a parallel to the articulation work that is required in collaborative work: For multiple actors to work together effectively, they need to establish and maintain common ground, and to understand responsibilities and division of work, etc. [31]. This articulation work is not by definition good or bad; it is simply a part of collaborative work. Similarly, genAI requires some articulation of intent and context on the user’s part, which requires the person to figure out both what to tell the system and how to tell it. This articulation can be productive, just as it can be productive when we explain a problem to a coworker (or a rubber duck on our desk) only to come up with an idea for a solution ourselves without the coworker needing to speak a word – but articulating mundane aspects (such as the fact that if I need ideas for how to structure an article, I mean an academic article and not a news article), especially repeatedly, is usually not. If genAI requires people to spend their energy on mundane articulation, it is likely to become an additional cognitive burden rather than fulfilling the promise of freeing up people’s time to let them focus on what matters to them. If, instead, genAI’s responses and suggestions are contextually anchored, the demand on people to articulate contextual aspects will be reduced, allowing them to turn their focus to productive articulation that helps them reflect on and develop their conceptualization of the task they want to do or the problem they are trying to tackle. In sum: (Appropriately) Anchoring people’s use of AI in their context (person and situation) can allow them to focus on productive articulation rather than mundane articulation.

Discussion. 4 So, what is the mundane?

... or: what is it mundane to have to articulate (over and over)? To paraphrase the above, the mundanely articulated is that which adds no value beyond ensuring that outputs are suitable for the person and the situation, while the productively articulated is that which progresses not only the generation of output but also how the work is conceptualized. A parallel can be drawn to Shah and Bender’s notion of “beneficial friction”, which acknowledges that information access is not just about getting an answer but also about sense-making [32]. My purpose here is not provide a concrete list of what is mundane and what is productive, partly as the answers will likely be contextually dependent themselves. My purpose is to point to the crucial distinction between productive and mundane articulation, begin a dialogue, and propose a direction of inquiry. In essence, I propose mundane and productive articulation as lenses to examine (gen)AI systems with, for purposes of both understanding and design. A possible line of inquiry in this regard is whether it is possible at a general level to map different aspects of a situation to a mundane/productive articulation distinction or axis, and if (likely) not, how dependent such a mapping is on the setting, the domain, the application area, and so on. It may (or may not) be that it is always less mundane to have to state what you are working on right now than to reiterate that you are a researcher and not a marketing director or a student. Another part of such an inquiry is to decide if mundane and productive articulation are most usefully conceived of as discrete categories or as ends of a spectrum. This, too, may prove to be dependent on other factors. Figure 1 summarizes this mapping challenge. To have some scaffolding for this inquiry, it would be helpful to have a properly nuanced understanding of context. Context is a contentious notion in HCI, and one definition or model of it may be suitable for one use case and inappropriate for another, making it impossible to definitively say what “context” is comprised of [7, 12, 16]. Abstracting away from an enumeration of concrete components of context, with the intention of focusing on the purpose of context information in this inquiry, one trap we might fall into is to assume that some aspects of context are completely situational – such as who else is in the room or what they are trying to achieve together – and others are a parcel of information that always travels with the person and is always present, such as a person’s job or educational background. It would be easy, then, to assume that it is mundane to tell the system what your job is and productive to tell it what you are trying to achieve. But these things may not actually be so straightforward to separate. For example: Articulating what or how much you already know about a subject, to avoid the AI overexplaining, can be both difficult and tedious – on the other hand, it can (as mentioned) be very productive to voice your thoughts and assumptions out loud, including what you already know about a problem and what knowledge you might apply to solve it. To tackle this, we should factor in that circumstances and characteristics can be more or less defining for a situation. Even relatively constant or slow-evolving characteristics come into play in a situational manner.

Conclusion. 5 GenAI is an Answer, Not the Answer That I discuss how to make genAI better should not be taken as an argument for genAI to be used everywhere or all the time. GenAI now seems to be part of the future of work and, as such, should be considered in research on work, particularly given its links with organizational knowledge use (Microsoft’s M365 Copilot, for example, uses an organization’s files to answer questions and surface potentially useful files to the user [26]). However, it is both worthwhile and important to consider that the way to create more relevant knowledge experiences at work can be better design, with or without (better) AI.

5.1 Is Generative AI Worth It?

Any suggestions for improving AI also become suggestions for advancing AI’s role in the world, and I am ambivalent about contributing to this. If the choice is between better AI and what we currently have, better is the obvious choice. But Sturdee et al. rightfully argue that “the potential negative consequences of research should be questioned, critiqued, and discussed as part of the publication and peer review process” [34, p. 1]. (Generative) AI comes with a range of potential and certain costs. There are consequences for people using it or working in an organization where it is used. There are also consequences for other people, both before that use and as a consequence of it. Consequences before use arise (and have already arisen), for instance, during model training and the creation of datasets for that training, e.g., when data labellers are underpaid yet have little choice but to do the work [15, 38], in many cases without being part of the population that reaps the largest benefits [48]. Consequences of use could be when people’s lives are influenced directly or indirectly by decisions and actions taken in the course of work mediated by AI-powered tools (a very concrete parallel is the use of predictive algorithms in, for example, social services and police work [25]). It could also, more indirectly, be the climate harms and displacement [2] caused by the environmental impact of generative AI.