Two AI agents able to handle an office task alone, from brief to finished document, just arrived on the market two days apart. Anthropic extended Claude Cowork to mobile and web on July 7, 2026, followed by OpenAI's ChatGPT Work on July 9, 2026. For a small business leader, this race is not just a communications duel: it marks the moment the office AI agent moves from experimental stage to mainstream product.
In brief
- Claude Cowork (Anthropic) moved from desktop to mobile and web on July 7, 2026, in beta for Max plan subscribers.
- ChatGPT Work (OpenAI) launched on July 9, 2026, powered by GPT-5.6, with a web, mobile, and desktop rollout.
- Both agents can chain several steps without constant supervision: read files, draft documents, keep working even with the device off.
- Human approval is still required before sensitive actions, a safeguard every small business should check before delegating a task.
- The office-agent market is accelerating: Anthropic reports annualized revenue growing from around 9 billion dollars at the end of 2025 to more than 30 billion dollars in 2026.
Two launches, three days apart
Claude Cowork was not born on July 7, 2026. Anthropic first unveiled it as a desktop preview on January 12, 2026, then expanded it in February with connectors for Google Drive, Gmail, Docusign and FactSet. July's new feature is the arrival on mobile (iOS, Android) and web: a session started on a computer keeps running in the cloud even after the computer is closed, and the user can track progress and approve sensitive steps from their phone.
OpenAI answered two days later with ChatGPT Work, an agent that gathers context from several apps, files, and workflows to produce finished documents, spreadsheets, presentations, or reports. The agent can stay active on a complex project for hours by breaking it into smaller sub-tasks. It connects to Google Drive, Slack, Salesforce, and other tools through a plugins directory. Its rollout started with Pro, Enterprise, and Edu plans, before expanding to Plus and Business; the desktop app now merges Chat, Work, and Codex and becomes available on every plan, including Free, on Windows and Mac.
January 12, 2026
Claude Cowork is born
February 2026
Enterprise expansion
July 7, 2026
Cowork reaches mobile and web
July 9, 2026
ChatGPT Work launches
What these agents can (and cannot) do
Both tools share the same promise: hand over a goal rather than a list of instructions. In practice, a business owner can ask "prepare the client meeting summary from these notes and the latest quote", and the agent fetches the files, cross-references the information, and delivers a structured document without every step being spelled out.
Both still keep important limits, though. Neither Cowork nor ChatGPT Work acts indefinitely behind the user's back: when a step involves a decision (sending an email, approving an expense, publishing content), the agent stops and asks for human confirmation. On mobile, that confirmation arrives as a notification: tracking happens from the phone, but execution stays anchored on the computer or in the provider's cloud.
| Claude Cowork (Anthropic) | ChatGPT Work (OpenAI) | |
|---|---|---|
| Broad mobile/web launch | July 7, 2026 | July 9, 2026 |
| Underlying model | Claude family (incl. Sonnet 5) | GPT-5.6 |
| Initial access | Max plan, gradual beta | Pro, Enterprise, Edu, then Plus and Business |
| Connectors cited | Google Drive, Gmail, Docusign, FactSet | Google Drive, Slack, Salesforce, plugins directory |
| How it works | Persistent cloud sessions, resumes after the computer closes | Splits a complex project into sub-tasks, works for hours at a time |
| Human approval | Requested on mobile before sensitive steps | Reports progress, stops on high-stakes decisions |
Why this launch matters for small businesses
A small business rarely has the time or the team to keep watching a multi-step AI project continuously. That is exactly what these agents aim to fix: delegate an entire task (preparing a file, summarizing customer feedback, drafting a first version of a presentation) rather than a single isolated prompt. Anthropic notes that real-world Cowork usage goes well beyond code: most sessions involve ordinary office work, not software development.
Without an office agent
An employee copies and pastes between tools, chases updates by email, assembles the document by hand. Time spent on coordination often exceeds actual production time.
With an office agent
The agent gathers the sources, produces a first deliverable, and only escalates the decisions that matter. The employee keeps control of judgment, not of manual handling.
This shift fits a clear market trend. Anthropic reported that its annualized revenue grew from around 9 billion dollars at the end of 2025 to more than 30 billion dollars in 2026, driven in part by enterprise demand for its agentic tools.
Precautions before delegating an agent
An agent acting across several applications at once also multiplies the entry points to your data. Before opening Claude Cowork or ChatGPT Work to a team, three checks are essential: which tools the agent can actually touch (limit connectors to what's strictly necessary), which actions require mandatory human approval (never let an agent send an email or approve a payment without confirmation), and who in the company can review the history of completed tasks.
Key takeaway
An agent connected to your Drive, your inbox, or your CRM inherits the same access rights as the user who configured it. Before any rollout, limit connectors to what the task strictly requires and require human approval for irreversible actions (sending, paying, publishing).
This matches what LUWAI already recommends for any agent connected to business tools: a narrow, well-monitored scope beats broad, loosely defined access.
FAQ
Are Claude Cowork and ChatGPT Work available outside the US?
Yes, both tools are accessible internationally through an eligible subscription (Max plan for Cowork, Pro, Plus, Business, Enterprise, or Edu plans for ChatGPT Work). The rollout happens gradually by subscription tier.
Does a small business need to choose between Claude Cowork and ChatGPT Work?
Not necessarily. Both tools address the same need (delegating a multi-step task) with different connector ecosystems. The choice often depends on the tools already in place at the company (Google Workspace, Slack, Salesforce) and the AI model teams are already comfortable with.
Do these agents replace an employee?
No. Both tools are designed to stop and ask for human approval before any high-stakes decision. They cut down time spent on coordination and formatting, but business judgment stays with the team.
What are the main risks to watch?
The main risk is poorly scoped connectors: an agent with access to too many tools can act beyond what was intended. Best practice is to limit connectors to what's strictly necessary and require human approval for irreversible actions.
To go further on AI agent governance in business, read our guide Agents IA en entreprise : ce qui change en 2026 and our article on securing AI connectors. For hands-on support, explore our success stories.
Sources: Anthropic Newsroom, TechCrunch, OpenAI News.
