Always-on agents: Google, Meta and OpenAI are selling you a digital colleague
· Tarek Nachnouchi
In three weeks, the three giants unveiled agents that work continuously, plugged into your tools. I read their help pages. Here is what a European company can actually use today, and the checklist I would apply before switching one on.
What was announced?
On 29 September, OpenAI presented Dots, "always-on" agents with their own computer in the cloud. A Dot keeps working on a project between two conversations and runs scheduled tasks with the apps you connect.
The same day, Meta extended its Muse agent to small businesses. According to Meta, it connects to Shopify, Stripe, QuickBooks and Facebook, Instagram and Meta Ads accounts. It is free within a usage limit, then costs $20 or $100 a month.
On 8 October, Google Cloud launched "Gemini agent", built into Gmail, Docs and Sheets and connected to tools such as Salesforce or Shopify. It picks the right model for each task and offers real-time spending caps.
What can a European company use today?
Official pages checked on 19 October:
- Dots (OpenAI): excluded from the Pro plan in the European Economic Area, Switzerland and the United Kingdom. They are open to ChatGPT Business Premium customers in every region where ChatGPT is available, France included. On Enterprise, a beta exists and is switched off by default.
- Muse for Small Business (Meta): Meta describes Muse as available in the United States and Canada. No date has been announced for Europe.
- Gemini agent (Google): "in early access with SMB customers", then open to everyone "soon", according to Google. The countries concerned are not specified.
So a company in France or elsewhere in Europe can test this kind of agent now, through a business plan. I see an advantage in that: an administrator keeps control of access.
Why I call it a colleague
An always-on agent takes initiatives within the limits you set, without waiting to be asked. OpenAI's help page says a Dot can proactively review data from connected apps and turn it into memories. It adds an important detail: disconnecting an app does not delete what the agent has already retained. To erase it, you have to delete the agent.
That sentence stopped me. It means you are granting access as you would to a new hire, with a memory you cannot control line by line.
I was an advisor to Purples.ai, a platform for beauty brands. Its agent, Purla, reads Amazon reviews overnight, tracks competing products and listens to social media. Every morning, the brand receives a summary and drafted fixes waiting for its sign-off. The Purples.ai team built the product that way, and I use this way of working as my reference.
The lesson of a Canadian case
In November 2022, after his grandmother died, Jake Moffatt asked Air Canada's chatbot about fares. It told him he could claim the bereavement fare after his trip. The airline's policy said the opposite. Before the British Columbia Civil Resolution Tribunal, Air Canada argued that its chatbot was "a separate legal entity that is responsible for its own actions". The tribunal member called this a "remarkable submission". In February 2024, the airline was ordered to pay 812 Canadian dollars.
The amount is small. The principle is clear: the company answers for what its agent says. When the agent carries out actions, the exposure grows.
The incidents OpenAI published on 15 October point the same way. During an internal test, a model asked to deploy an application used funds from a test wallet it had no right to. According to Ars Technica, it wondered whether it had permission, then judged the use acceptable. These models were in training, with some guardrails switched off. Still, an agent can replace an authorisation with its own judgement.
What I recommend before switching on an agent
- Name one human owner per agent. That person approves its scope and answers for its actions, as with an intern.
- Write its scope on one page. List the connected apps, the actions it may take without asking and those that need approval. Dots offer rules such as "ask before acting" or "hand over to you". I would set them up on day one.
- Start read-only. First connect sources the agent can consult without changing anything, such as calendars or dashboards. I would allow it to send messages only after a few weeks of observation.
- Cap the spend. Google highlights per-project caps. Set a monthly budget and an alert. In the report G2 published on 14 October, 46% of agent buyers report higher operating costs than they expected.
- Tell the people involved. Article 50 of the EU AI Act, in force since 2 August 2026, requires that people be informed when they interact directly with an AI system. If your agent writes to customers, say so clearly.
- Plan the exit. Check how to retrieve the work produced and how to erase the agent's memory before you hand it any data.
Framing this first use is part of my work on deploying AI solutions at Hiclix.
Decision checklist: can I switch this agent on?
| # | Question | Yes / No |
|---|---|---|
| 1 | Is a human owner named for this agent? | |
| 2 | Is its scope (apps, free actions, actions that need approval) written down? | |
| 3 | Is sensitive data excluded, or authorised in writing? | |
| 4 | Are a monthly budget and a spending alert in place? | |
| 5 | Will outside people in contact with the agent be informed? | |
| 6 | Do you know how to retrieve its work and erase its memory? |
How to read it: 6 yes, run a 30-day pilot on an internal task. 1 or 2 no, fix them before you start. 3 no or more, wait.
My takeaway
I would test an always-on agent this autumn on an internal, repetitive task, for example preparing a project's weekly update. I would connect it to a customer or a bank account only after completing the checklist. I will come back to the subject when Muse and Gemini agent reach Europe.
If you would like to discuss this for your organisation, I offer a first conversation of 30 to 45 minutes, free and with no commitment.