Software development organizations (SDOs) are increasingly working to adopt artificial intelligence (AI) tools, like GitHub Copilot, to meet varied expectations. Nevertheless, we know little about how SDOs manage these expectations. This paper investigates how different SDOs expect AI tools to impact software development, and how these expectations change after a period of considering and evaluating AI tools. We conducted a multiple-case study involving three SDOs. To elicit initial expectations towards AI tools, we collected data using semi-structured interviews and field visits. To assess the persistence of expectations towards AI tools, we collected data from meetings, a debriefing, and retrospectives on AI tools. We found three expectations particular to one SDO; four shared between two SDOs; and six pervasive across all SDOs. Five expectations did not persist after experiential learning with AI tools, due to platform- and SDO-related factors. SDOs must carefully manage their expectations towards AI tools due to the variety and complexity of expectations. Some expectations are niche-specific based on their compatibility with the unique SDOs' people- and structure-related aspects, while others are becoming mainstream for a broader array of SDOs. Recognizing factors that affect the persistence of expectations and how they manifest in the individual SDO will enable SDOs to form their initial expectations and understand how these might change during adoption of AI tools, supporting expectation management.

软件开发组织 （SDO） 越来越多地努力采用人工智能 （AI） 工具，例如 GitHub Copilot，以满足不同的期望。然而，我们对 SDO 如何管理这些期望知之甚少。本文研究了不同的 SDO 如何期望 AI 工具影响软件开发，以及这些期望在考虑和评估 AI 工具一段时间后如何变化。我们进行了一项涉及三个 SDO 的多案例研究。为了引起对 AI 工具的初步期望，我们使用半结构化访谈和实地考察收集了数据。为了评估对 AI 工具的期望持久性，我们从会议、汇报和 AI 工具回顾中收集了数据。我们发现一个 SDO 有三个特定的期望;4 个由 2 个 SDO 共享;6 个在所有 SDO 中普遍存在。由于平台和 SDO 相关因素，在使用 AI 工具进行体验式学习后，五个期望并未持续。由于期望的多样性和复杂性，SDO 必须仔细管理他们对 AI 工具的期望。一些期望是基于它们与独特 SDO 的人员和结构相关方面的兼容性而特定于利基市场的，而另一些期望正在成为更广泛的 SDO 的主流。识别影响预期持久性的因素以及它们如何在单个 SDO 中表现出来，将使 SDO 能够形成他们的初始期望，并了解这些期望在采用 AI 工具期间可能如何变化，从而支持期望管理。