Detailed Analysis
A user's Reddit post detailing frustration with uninstalling Claude for Excel and PowerPoint highlights a recurring pain point in Anthropic's push to embed Claude directly into Microsoft Office applications. The poster describes a month-long struggle: the add-ins appear in the ribbon and function normally, yet they are invisible in the standard Office "Add-ins" management pane, do not respond to right-click removal attempts, and persist even after a full uninstall and reinstall of PowerPoint. As a Microsoft 365 admin, the user also checked the admin center's "Integrated apps" settings—both for their own account and the tenant-wide admin view—and found no trace of Claude listed there either, despite having installed the integrations directly through their Claude account rather than via Microsoft's app marketplace.
This scenario points to a structural ambiguity in how Anthropic has deployed these productivity integrations. Claude for Excel and Claude for PowerPoint are relatively new offerings, part of Anthropic's broader strategy to compete with Microsoft Copilot and Google's Gemini integrations by embedding AI assistance directly into the tools knowledge workers already use daily. But because the install originated from the Claude.ai account side rather than through Microsoft's official app registration and consent flow, it seems to sit outside the normal Office add-in lifecycle that admins and users expect to control. This is precisely the kind of gap that erodes trust: a tool that grants itself ribbon-level access to desktop applications but doesn't expose itself through any of the conventional uninstall pathways—Add-ins menus, right-click context actions, or admin center integrations—leaves users feeling like they've lost visibility into, and control over, software running against their own documents.
The stakes here go beyond mere inconvenience. The user explicitly frames the issue as a security and autonomy concern: "I don't want Claude having any power to touch my desktop apps." For an AI company whose products are increasingly granted agentic capabilities—reading, writing, and modifying files, executing multi-step tasks, and now operating inside core productivity software—an inability to cleanly revoke access undermines the trust model that enterprise and individual users need to feel comfortable granting such permissions in the first place. The fact that even Claude itself, when asked, could not walk the user through a working removal process is notable, since it suggests either a genuine product gap in Anthropic's support tooling or documentation, or an installation architecture that wasn't designed with straightforward deprovisioning in mind.
More broadly, this incident reflects growing pains common to the current wave of AI agents moving from browser-based chat interfaces into deeper OS- and application-level integration. As Anthropic, Microsoft, Google, and OpenAI race to make AI assistants persistent fixtures inside email, spreadsheets, presentations, and code editors, the lifecycle management of these integrations—installation, permissions, updates, and especially removal—has often lagged behind the pace of feature rollout. Clean, discoverable uninstall paths are a basic expectation for any software with system-level hooks, yet as AI copilots proliferate across office suites, users are increasingly encountering tools that are easy to add but hard to fully excise, a friction point that could shape enterprise IT policies and user willingness to adopt these integrations going forward.
Read original article →