On 23 September 2026, OpenAI and Meta made announcements within hours of each other that, taken together, signal a shift in how AI agents will be accessed in the workplace. OpenAI brought voice-triggered agentic workflows to its mobile app. Meta expanded its Muse personal agent with Mac computer control, an email address, and integrations with enterprise tools including GitHub, Notion and Shopify. Neither announcement was framed as enterprise-specific. Both have immediate implications for how businesses evaluate and deploy agentic AI.
OpenAI: voice becomes an agent interface
OpenAI's update to the ChatGPT mobile app allows Plus and Pro subscribers to trigger multi-step workflows using voice commands. Users can draft documents, summarise emails, build simple websites and create presentations from their phone. The feature extends the Work tab functionality that launched on desktop in July 2026, allowing conversations started on mobile to be resumed on desktop.
The technical detail that matters for enterprises is not the voice recognition itself. It is the fact that agents are now expected to operate across devices with state persistence. An employee can start a research task on their commute, refine it at their desk, and review the output in a meeting. The agent maintains context across sessions. That is a meaningful architectural shift from the siloed chatbot model that most enterprises are still piloting.
OpenAI also updated its voice output to include richer text formatting, which matters for document generation and email drafting — the kind of routine knowledge work that makes up the bulk of enterprise agent use cases. Free and Go tier users get access to plugins and connected apps, which broadens the surface area for third-party integrations.
Meta Muse: the agent gets a desk, an inbox and a tool stack
Meta's Connect event on 23 September 2026 positioned Muse as the company's central AI bet. The headline consumer features — a real-time avatar, smart glasses integration and shopping connectors — attracted the most coverage. The enterprise-relevant details were quieter but more consequential.
Muse is gaining computer-use capabilities on macOS, allowing it to operate any application on a user's desktop. Meta's chief AI officer, Alexandr Wang, described the use case explicitly in business terms: "You can have it help you run your small business, or just get work done for you. You can walk away from your computer and it keeps working for you on all the jobs you lined up."
The agent will also receive its own email address, enabling it to be added to threads and delegated tasks via forward. Partnerships with GitHub, Notion, Granola, Shopify, Stripe and PayPal mean Muse can already interact with tools that sit at the centre of many small and mid-sized business workflows. Meta says it received more than 1,500 developer applications for custom connectors in under a week.
Why these two announcements belong in the same conversation
Individually, neither announcement is transformational. OpenAI already had desktop agents. Meta already had a personal assistant. Taken together, they illustrate a convergence that enterprise buyers should track closely.
First, the interface layer is fragmenting. Agents are no longer confined to chat windows. They are accessed by voice, by email delegation, by background desktop automation and by wearable devices. Enterprise IT departments that are still standardising on a single chatbot interface are already behind the curve.
Second, the boundary between personal and work agents is blurring. Muse is explicitly positioned as a personal agent, yet it connects to business tools. OpenAI's Work tab is explicitly for productivity, yet it runs on the same app as general-purpose ChatGPT. For enterprises, this creates a shadow IT problem before shadow agents have even been defined. If employees can delegate work tasks to a personal agent that stores data on Meta's infrastructure, the compliance implications are immediate.
Third, the connector economy is accelerating. Meta's 1,500 connector applications in a week suggest that developers are betting heavily on the agent-as-platform model. For enterprises, this means the tool stack around agents will evolve faster than internal procurement cycles can keep up. The risk of picking a platform that loses the connector race is real.
What this means for enterprise buyers
Three practical questions arise from these developments.
Device and access policy: If voice-activated agents can draft emails and access cloud documents from a mobile device, what is the access policy? Does the enterprise allow agents on personal phones? Does it require managed devices? Does it monitor voice-originated agent sessions with the same rigour as keyboard-originated ones?
Data residency and connector governance: When an agent has its own email address and can read GitHub repositories, Notion workspaces and Shopify catalogues, the data perimeter expands dramatically. Enterprises need to know which connectors are approved, which data categories each connector can access, and how to revoke access when an employee leaves.
Shadow agent detection: Most enterprises are still struggling to discover which SaaS tools employees have signed up for. Agents that operate across devices, tools and platforms make that problem exponentially harder. The first generation of agent governance tools is only just emerging. The gap between agent capability and agent visibility is widening.
The Agentic Expo angle
The agent market is moving from "can it work?" to "where is it allowed to work?" OpenAI and Meta are answering the first question with increasingly capable demonstrations. Enterprise buyers still need to answer the second.
Agentic Expo exists because the buyers who will deploy these tools at scale need more than product announcements. They need to evaluate orchestration platforms, security vendors, integration specialists and governance frameworks side by side. They need to ask hard questions about data residency, connector approval and cross-device policy in rooms where the vendors are present and accountable.
Between now and March 2027, the convergence of voice, desktop and mobile agents will move from beta to production in enough organisations that it becomes a standard procurement consideration. The buyers who understand that shift now will be the ones writing the policies that everyone else copies later.
Sources: TechCrunch, "ChatGPT mobile app gets voice-based agentic features," 23 September 2026; TechCrunch, "Everything new coming to Meta's AI agent Muse," 23 September 2026; About Facebook Newsroom, "Introducing Muse: Personal AI Agent," September 2026.