AI agents at the office: what really changes, and what doesn't

Contents
An AI agent no longer just answers a question: it chains steps, opens your files, connects to your apps and hands back a finished deliverable. OpenAI launched ChatGPT Work on July 9, 2026, Mistral renamed its assistant Vibe in late May and added a Work mode, and both Anthropic and Microsoft now sell a tool called Cowork. For employees, the question has shifted from whether to use AI to what can safely be handed over.
What is an AI agent, in practice?
A classic chatbot writes text in a chat window. An agent gets a goal and plans its own way there. According to OpenAI, ChatGPT Work gathers context from your apps and files to deliver finished documents, spreadsheets, presentations, reports or websites.
Mistral's Vibe Work mode connects to Google Workspace, Outlook, SharePoint, Slack or GitHub. Heise reports that it asks for your consent before starting, logs every step and can repeat a task daily, weekly or monthly. OpenAI goes further with Dots, always-on agents with their own cloud computer and browser, governed by rules the user writes: what they may do alone, what needs approval and what is off-limits.
Which tools are available in late 2026?
| Tool | Maker | What it does | Who gets it |
|---|---|---|---|
| ChatGPT Work | OpenAI | Documents, spreadsheets, decks, reports and sites from your apps and files; scheduled tasks | Launched July 9, 2026 for Pro, Enterprise and Edu, then Plus and Business |
| Dots | OpenAI | Always-on agents, 4,000+ connectable apps, autonomy rules | Pro and Business Premium subscribers in eligible markets |
| Vibe Work mode | Mistral AI | Multi-step tasks, upfront consent, recurring tasks | Vibe assistant (formerly Le Chat), Free to Enterprise |
| Cowork | Anthropic | Delegated work by Claude across your files and tools | Claude Team and Enterprise plans |
| Copilot Cowork | Microsoft | Delegated tasks in Microsoft 365: inbox workflows, research, documents, reusable skills | Microsoft 365 Copilot business customers, via the Frontier early-access program |
To compare the subscriptions themselves, see our guide ChatGPT, Claude, Gemini, Mistral or Copilot: which one to pick.
What really changes
Adoption is real. Microsoft's 2026 Work Trend Index, published May 5 and based on 20,000 AI-using workers in ten countries including the US and UK, found that active agents in Microsoft 365 grew 15-fold in a year, and 18-fold in large enterprises.
The kind of work handed over is changing too. People no longer ask for a paragraph but for a summary of twenty documents, a tracker refreshed every Monday or a first draft of a slide deck. The gains come mostly from repetitive, well-defined work.
And what doesn't change?
Reliability on real projects is still limited. The Remote Labor Index, built by Scale AI and the Center for AI Safety, gives agents 240 real freelance jobs. As of November 17, 2026, the top model, Claude Fable 5.2, produced work a client would accept in 41.25% of cases; the runner-up reached 21.25%. In other words, most deliverables still fail.
Accountability can't be delegated either. In Microsoft's survey, 86% of workers treat AI output as a starting point rather than a final answer, and the human skill they cite most is quality control of results (50%). Back in June 2025, Gartner predicted that over 40% of agentic AI projects would be canceled by the end of 2027 over unclear value or weak risk controls. For the jobs question, read what the studies really say.
What are the security and data risks?
An agent that reads your email and acts inside your software widens the attack surface. On May 1, 2026, the US Cybersecurity and Infrastructure Security Agency (CISA) and partners including the Australian Cyber Security Centre published guidance on careful adoption of agents. Their advice: fold these tools into existing cybersecurity rules, never grant broad or unrestricted access, strengthen identity management and monitor continuously. In Europe, these uses also fall under the AI Act.
How to get started without nasty surprises
- Start with a repetitive, low-risk task: meeting notes, a tracking sheet, a research roundup.
- Describe the expected result, format and sources precisely; our ready-to-use prompts can serve as templates.
- Limit connections: an agent doesn't need your whole inbox to summarize one file.
- Keep a human approval step for anything that leaves the company.
- After a month, measure the time actually saved, review time included.
Frequently asked questions
Can an AI agent send emails on my behalf?
Technically yes, if you connect it to your mailbox. Tools such as Vibe or OpenAI's Dots let you set rules that require your approval before that kind of action. Turn them on.
Do I need a paid plan?
Not always to try it: OpenAI's pricing grid lists limited ChatGPT Work access on desktop even on the free tier. Regular use means paying, though: Dots are for Pro and Business Premium, and Anthropic's Cowork for Team and Enterprise plans.
Are agents reliable?
Not yet for whole projects: on the Remote Labor Index, the best model completes 41.25% of the real jobs tested. They are more useful on short, well-scoped tasks.
Can my employer ban these tools?
Yes. Employers decide which tools are allowed and set confidentiality rules. If in doubt, ask your IT team before connecting work data.
Sources
- BNN Bloomberg – OpenAI launches ChatGPT Work (July 9, 2026)
- The Batch (DeepLearning.AI) – Sol Nearly Eclipses Astra: GPT-6.1 Sol, Dots (September 29, 2026)
- heise online – Mistral's chatbot is now called Vibe (May 29, 2026)
- Anthropic – Plans & Pricing (Claude) (November 17, 2026)
- Microsoft WorkLab – 2026 Work Trend Index: Agents, human agency and the opportunity for every organization (May 5, 2026)
- Scale Labs / Center for AI Safety – Remote Labor Index, leaderboard (November 17, 2026)
- Microsoft 365 Blog – Copilot Cowork: from conversation to action across skills, integrations and devices (May 5, 2026)
- OpenAI – ChatGPT Plans & Pricing (November 17, 2026)
- Gartner – Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (June 25, 2025)
- CISA – Careful Adoption of Agentic AI Services (May 1, 2026)
- CISA – CISA and Partners Release Guidance on Adopting Agentic AI Services (bulletin) (May 1, 2026)




