Understanding the Uses of AI in Business to Better Support Them

• AI
• Digital
• Sciences humaines
• Work

Share on Twitter : Understanding the Uses of AI in Business to Better Support Them (New window) Share on Linkedin : Understanding the Uses of AI in Business to Better Support Them (New window)

This article examines the evolution of usage patterns through the lens of users’ ability to adopt AI in professionalcontext, drawing on both previous research regarding the adoption of digital tools in professional contexts and the initial findings of two studies. The first study focuses on the professional use of generative AI solutions by developers on a portal providing access to AI models (GPT, Mistral, Gemini, Codestral, Claude Sonnet, etc.) for professional use by Orange employees. Employees can use this portal to discover and create personalized assistants. Its unique feature is that it is secure: employee data does not leave the company and is not used for retraining the models. The second study was conducted during the launch of a pilot for a symbolic AI solution, Live Wubble (LW), at one of Orange’s partners, a French software publisher. LW is a proactive assistant, in the form of a widget, which, in our case, is designed to help sales administrators (ADV) process requests regarding invoices or purchase orders. LW facilitates access to these documents from the application currently on the screen (Outlook, Teams, Word, etc.), provided the user has authorized it to monitor the screen. More specifically, the LW engine scans the PC screen, searching for references to invoices and purchase orders that it has learned to recognize, using a list of regular expressions. When it finds them, it queries the platform via an API to verify their existence in the document database. Depending on the user’s access rights, LW displays a list of files containing that reference, with links that the user can click directly, without having to perform a search.

The adoption of these tools depends on employees’ ability to take action to adapt them and the way their work is organized, so that they can tailor these tools to their specific tasks and make better use of them

The adoption of AI in light of the capabilities approach

The arrival of AI in the workplace is part of a long history of working with digital technologies, marked by strong, prophetic, and normative discourse on the “power of digital technology”; however, these anticipated “effects” of digital technologies in a professional context cannot materialize without employees taking ownership of them. In this sense, the sociology of digital technologies in the workplace has consistently emphasized the importance of individual, collective, and organizational dimensions in the adoption of digital technologies [1]. The development of these uses depends on employees’ agency to evolve these digital technologies and the organization of their work, in order to adapt these digital technologies to their activities and rely on them more effectively. These two facets of their agency—over their digital technologies and over their work—are inseparable and necessary for employees to improve their control over their work, their environment, and their future. Agency implies freedom of choice (choices that are genuinely accessible), the means, and a set of favorable factors to evolve their tools and the organization of their work.

In this sense, Amartya Sen’s capabilities approach is useful to us, as it emphasizes the distinction between capacity and capability [2], in other words, between “knowing how to do something” and “being able to do something.” Just because employees have access to AI and know how to use it does not mean they will develop ways to apply it in their work. The concept of capability emphasizes the power to be and to do. The transition from capacity to capability is driven by conversion factors – individual, collective, organizational, and institutional. Research has enriched this approach by illustrating the variety of these factors [3, 4, 5, 6].

The specific characteristics of AI-powered tools and users’ ability to adapt them to their work

Compared to symbolic AI (based on decision trees), the new types of AI—connectionist and generative—which often overlap, rely on probabilistic computation. The quality of the results produced by these AI systems depends both on the size of the databases made available to them and on the quality of that data. On the one hand, the data may contain biases, which the algorithm will subsequently reproduce (by reinforcing, for example, existing social inequalities); on the other hand, the data needs to be adapted to the task for which the employee intends to use it [7, 8]. Furthermore, new learning methods based on neural networks can quickly become incomprehensible to end users (and sometimes even to their designers) [9]. But this opacity of the algorithm can also be organizational: it is not solely due to the overly complex functioning of AI and users’ lack of familiarity with the statistical techniques that would allow them to understand it at least in part, but also to the veil with which designers shroud the functioning of their algorithm, considered a competitive advantage not to be disclosed [10].

Thus, the heterogeneity of algorithms, their incomprehensibility, their opacity, and the rapid evolution of LLMs through machine and deep learning mechanisms make them difficult for users to understand. Consequently, end users’ ability to negotiate the modifications to the tool necessary for their work (e.g., accounting for a specific category or mitigating a particular bias) with other employees involved in the design or training of generative AI algorithms is limited. This complexity of the tools that utilize these new types of AI thus undermines end users’ ability to influence these tools.

In the case of the Live Wubble experiment, the designers—who were closely involved in developing the use cases—incorporated the modifications requested by the sales representatives, who wanted to use LW not only to identify files related to purchase orders and invoices but also for license keys. This required modifications to the algorithm (in the construction of the list of regular expressions) and the implementation of a database for searching these license keys. In other words, adapting the tool requires the designers to be able to integrate these modifications and the sales representatives to build and populate this database.

The ability to adapt one’s work to a tool that uses AI

How to integrate this tool into one’s business also depends on a number of factors.

Thus, at the institutional level (both national and international), the entry into force of the European regulation, the AI Act, encourages companies, for example, to consider the responsible development of AI, taking into account social and ethical implications. In the case of our experiment, the development of the LW agent takes place in a context where the law requires companies to digitize their invoices, and these companies are also seeking to optimize the management of these e-invoices.

At the individual level, a person’s background, the training they’ve completed, their skills, their domain expertise, and the tools they’ve already used all contribute to shaping their ability to code, formulate prompts for AI, spot biases in algorithms, identify and correct errors in results, or even rephrase queries. Domain expertise therefore plays a significant role in this learning process.

“So, it’s not the AI part that’s hard; it’s the part about becoming a senior developer, mastering clean code practices, katas, domains… well, all the things you learn through practice […] With AI, it really highlights your own skills.” (senior developer)

When companies implement AI-powered tools, organizational factors are crucial, as building users’ capacity begins with the discussions they have locally, within their work context—both regarding the added value of these tools for their specific tasks (e.g., choosing the right LLM for a given task) and the impact these tools have on their work. These discussions, whether at the team level, within project groups, or in other more ad-hoc work groups, allow members to reflect on how their work aligns with AI, outlining potential courses of action to better understand and utilize it.

These discussions may simply involve sharing tips and highlighting the limitations of available tools, especially in a context where LLMs evolve very quickly (e.g., no longer using a particular LLM after a certain stage of the process). The developers interviewed explain, for example, that these discussions are all the more useful because they sometimes feel overwhelmed by the mass of information generated by AI. Business experts organize these types of exchanges within broader cross-functional networks. To this end, they prepare comparisons between LLMs to provide a clearer picture of their performance and resource consumption relative to the tasks at hand; however, as they explain, this task is complex.

“Now, if smaller models start producing good results too, it gets complicated […] Then, since new ones come out every month or every two months… in the end, we never stop benchmarking […] and it’s also complicated because depending on the task at hand, one model or another might be the right fit, so benchmarks can quickly become very complex.” (business expert)

Managers have an important role to play in facilitating these discussions within their local context, as close as possible to the business operations. They are also the ones who can support experimentation within their scope, encourage participation, and open up discussions about the challenges associated with these tools and the solutions to overcome them. These discussions may, for example, focus on mentoring newcomers to the team (young people or those who have undergone an upskilling training), who tend to rely solely on AI out of fear of disturbing their colleagues or being judged by them, without being able to assess either the errors that creep into the AI’s responses or the risk of isolation they face.

These discussions may lead to a redefinition of roles and responsibilities, the need for mentoring or work to ensure compatibility with existing tools, the identification of a loss or gain in skills regarding certain aspects, etc.

In the case of the license keys from the Live Wubble study, sales representatives point out, for example, that building the database for these licenses requires them to agree on the content to be included and the time to be devoted to populating the database. This requires prior collaboration with a member of the company’s IT department, who helps them develop a classification plan. In the case of invoices and purchase orders, where the database is already in place, they highlight the issue of the quality of the files returned by the tool. Even if the tool returns the correct file, they do not always find the information they are looking for. In fact, sales representatives, often in a hurry to get a contract signed, do not always fill out the purchase orders in the database correctly. The speed with which they can process their requests depends on the quality of the files that the sales representatives have uploaded. The quality of this business process is the subject of discussions between the sales administration team and the sales representatives.

Organizational factors refer to the framework established by the company to facilitate this alignment between AI and work at the local level. This framework takes the form of company policies or agreements, which must address not only employment issues but also the transformations in work brought about by AI [11]. For example, these policies can:

• raise awareness among employees and managers regarding the importance of these “local readjustment” discussions concerning both the tool and the organization of work; in the case of LW, the significance of the partner’s strategic and commercial challenges led managers to encourage their team members—who did not see the value of this solution—to test it;
• facilitate management’s consideration of requests arising from these local AI-work adjustments, as well as the provision of resources for these changes identified at the team level;
• provide for the appointment of key employees (ambassadors, champions, etc.) to facilitate understanding of how these AI-powered tools work, as well as discussions and transformations related to adapting these tools to local contexts;
• take into account, in terms of budgeting and contract terms, the need to regularly adapt algorithms to the business (e.g., categorizations performed by the algorithm that are closely aligned with the specific characteristics of the business), database updates, and the development of prompt libraries, so that business units (e.g., HR, Legal, etc.) have AI tools that reflect the social and organizational reality of their company.

Conclusion

This article emphasizes that the future of working with AI hinges on the practical ability of employees to adapt AI-powered tools to their specific activities and to evolve their work in line with the possibilities offered by these tools.

With these new types of AI, users’ ability to act involves more complex tools (opacity, incomprehensibility, heterogeneity, the constant evolution of LLMs, the need to build databases tailored to their specific field, etc.) and involves collaboration with more employees than before (e.g., data scientists, IT system personnel assisting in the creation of a classification scheme for building a database, etc.).

Support, which has also become more complex, is built by relying on the local identification, at the team level, of these individual, collective, organizational, and institutional factors. Consequently, supporting these uses means ensuring the conditions for their development at all these levels.

While this article helps identify factors specific to local contexts, new studies on the use of AI in professional context will provide a clearer picture of their full diversity.

This text has been translated by an artificial intelligence.