Microsoft has rebuilt Copilot around the workplace while Meta and OpenAI are racing to put an always-on agent in everyone’s pocket. Here’s what hit the headlines in the past few weeks, and what it means for you.
Microsoft puts Copilot to work
On 25 September, Microsoft merged its consumer and workplace versions of Copilot into one work-first app, a move Bloomberg described as stepping back from the personal chatbot race.
The new app has three modes:
- Home is a central workspace that combines chat with delegated work, so you can review activity, resume tasks and work on Word, Excel and PowerPoint files without switching apps.
- Code lets non-developers build apps, dashboards and automations in plain language.
- Autopilot, previously known as Scout, runs continuously in the background. It monitors communications, follows up on tasks and manages recurring processes across Teams, Outlook and documents.
Autopilot sits close to what many EAs already do by hand. It’s separate from Copilot Cowork (Microsoft’s tool for longer multi-step jobs), so check which one someone means when they say “Cowork”.
Autopilot works within organisational permissions and governance controls, so what you see depends on what your organisation switches on.
CNBC reported that Cowork, Code and Autopilot will be billed by usage on top of the per-user licence, so it’s worth asking your IT team whether your organisation plans to adopt any of them. The rollout is happening over the coming weeks.
Meta and OpenAI want to be your assistant
Meta launched Muse on 8 September. It’s a personal agent that’s free to start and asks for approval before it sends an email or pays for anything.
OpenAI followed on 29 September with dots – always-on agents that each get their own cloud computer, connect to thousands of apps and keep working on your goals while you do something else.
Dots follow your ChatGPT permissions and let you set rules for which tasks need approval. OpenAI has also previewed versions built for organisations, which it says will plug into Microsoft’s Agent 365 governance tools. Both are rolling out by market, so check availability before assuming you can try them.
Each is pitched as a personal assistant, and the promise is delegation: hand it over and it gets done. The early friction shows where judgement still matters.
Semafor reported that Amazon has blocked Muse from shopping on its site, and that a Meta agent gave out a man’s home address on Facebook Marketplace without his permission.
When an agent goes off script
On 24 September, Prime Minister Anthony Albanese revealed that an OpenAI agent had accessed files on the Medicare Statistics Reporting Service in June, while researching public medicine spending during an internal evaluation. No Medicare customer details are believed to have been accessed.
OpenAI didn’t tell the government until 10 September, nearly three months after the access, which Albanese said was way too long. Reports describe the incident as possibly the first known case of an AI agent breaking into a government system.
A taskforce is investigating, and the Department of Home Affairs has ordered a government-wide cyber security assessment. OpenAI has since apologised and said the breach was broader than first known, and the government has sought advice on whether police should investigate.
It’s an extreme example, but the lesson applies to any agent working inside your systems. Before one lands on your desk, find out what your organisation has switched on and who can stop it when something goes wrong.