Agents → operations; governance gap

Office software is turning into operating software: AI agents now search company data and carry out work inside Slack, Google Workspace, and Salesforce. (salesforce.com) On April 22, Salesforce and Google Cloud said customers can deploy AI agents in Slack and Google Workspace, with Agentforce and Gemini Enterprise sharing context across both systems. Salesforce said the setup is meant to cut “context switching” and reduce data movement between tools. (salesforce.com) Google said Workspace Studio reached general availability in late 2025 as a place to design and manage agents that automate “complex workflows” inside Docs, Gmail, Sheets, and other Workspace apps. Slack’s developer docs say its MCP server lets AI apps search channels, send messages, manage canvases, and take other Slack actions. (workspace.google.com) (docs.slack.dev) OpenAI has pushed the same direction inside ChatGPT. Its company knowledge feature for Business, Enterprise, and Edu connects workplace apps including Slack and Google Drive, and OpenAI says admins must enable those apps before teams can use them. (help.openai.com) The change is that these systems are no longer pitched only as drafting tools. Slack says agents can become “teammates” in channels, and Salesforce says Slack is becoming an “agentic OS” where employees and agents search, collaborate, and act across company apps and data. (slack.com) (salesforce.com) That shift is colliding with a second trend: employers collecting more detailed worker telemetry. Reuters reported on April 21 that Meta told U.S.-based employees it would install software to capture mouse movements, clicks, and keystrokes for AI training tied to agents that can perform work tasks autonomously. (thestar.com.my) Regulators have already sketched the legal perimeter for that kind of monitoring. In Memorandum GC 23-02, issued on October 31, 2022, National Labor Relations Board General Counsel Jennifer Abruzzo said electronic monitoring and algorithmic management can interfere with workers’ Section 7 rights under U.S. labor law. (govinfo.gov) Privacy regulators in Europe and the United Kingdom have made similar points in data-protection terms. The U.K. Information Commissioner’s Office published worker-monitoring guidance on October 3, 2023, and its broader AI guidance says organizations must review lawfulness, fairness, and transparency when personal data is used in AI systems. (ico.org.uk) Vendors are responding by selling controls alongside capability. OpenAI says company knowledge includes citations, security, privacy, and admin controls; its help pages say connections can be created, deleted, or modified, and some apps with sync have region-specific availability rules. (openai.com) (help.openai.com 1) (help.openai.com 2) The governance problem sits at the workflow level, not the chatbot window. Once an agent can read Slack, pull a file from Google Drive, update Salesforce, and trigger a message or approval, the audit trail has to cover each handoff, each permission, and each source of training data. (docs.slack.dev) (salesforce.com) (help.openai.com) The near-term test for companies is no longer whether an agent can write a memo. It is whether the company can show who let it act, what data it touched, and whether workers were told how their own digital exhaust would be used. (ftc.gov) (govinfo.gov)