Among the latest developments in welfare services is the rise of generative artificial intelligence (AI), particularly large language models (LLMs) in the form of chatbots, virtual conversational agents designed to simulate human dialogue and assist with tasks, queries, or decision-making. The integration of LLMs into social work organizations presents both opportunities and risks. They may support social workers in reflecting on dilemmas during client assessments or guide citizens through the process of applying for social assistance. They also pose ethical dilemmas such as transparency in decision-making, algorithmic bias, and concerns around data access and privacy.

However, unlike traditional digital technologies such as documentation systems, LLM use is not confined to formal, organizationally sanctioned implementations. Increasingly, chatbots also enter practice informally, as social workers experiment with freely available platforms such as ChatGPT, Copilot, or Gemini. With private accounts, they may consult legislation, reflect on practice dilemmas, or draft case notes, introducing new complexities for organizational governance and professional accountability. In this sense, AI and LLMs are seeping into social work in ways that circumvent formal strategies, complicating efforts to regulate, evaluate, or support their use within welfare services.

By using social assistance as a case, and introducing concepts such as delegation and domestication, this paper discusses the formal and informal uses of LLMs by social workers in Swedish social services and its significance for professional discretion. We define formal use as organizationally integrated AI solutions, often customized or procured systems, while informal use captures the individual and creative ways social workers employ private AI tools in everyday reasoning and client-related tasks. Whereas formal use tends to rely on explicit knowledge, informal use is grounded in tacit knowledge.

A key issue arising from this formal/informal dichotomy is the participation of social workers in the development, adaptation, and evaluation of technologies. Scholars have long cautioned that without practitioner involvement, digital tools risk aligning primarily with organizational needs, efficiency and managerial control, rather than professional practices rooted in tacit, relational, and client-centered approaches. Absent such involvement, social workers may experience technostress, disempowerment, or resistance, as tools designed to support them instead impose additional burdens. Ensuring alignment with professional practices is therefore essential if LLMs are to strengthen rather than undermine the ethical and relational dimensions of social work.

In relation to the conference theme Social Work and Interdisciplinary Research: researching and facilitating evidence-informed practice and policy, this paper demonstrates how attention to dual modes of LLM use can inform strategies for integrating AI into welfare services. It further highlights how LLMs are already shaping practice on the ground in ways that elude formal governance. We argue that interdisciplinary research bridging technological innovation, organizational policy, and social work practice is critical to the development of chatbot applications that are not only efficient but also ethically robust and relationally sensitive.