On 8 October, Google introduced an agent that gets its own account inside a company: an email address, its own storage and a place on the team. You tag it in a chat, send it an email or share a document with it, and it works on the task, for days if needed. It knows who on the team works on what and who has to sign off.
Four weeks earlier, OpenAI released an API for developers to build exactly this kind of agent, one that stays on a task for hours or days. Meta now offers its agent Muse to small businesses, connected to Shopify, their accounting software and Slack.
That sounds like the product announcements that arrive every week. Underneath is a real shift. Until now, you asked AI questions. Now you give it tasks.
What changed
A chatbot answers a question and forgets it. You copy the text into the email yourself and type the number into the ERP yourself. The work stays with you; the AI supplies parts.
An agent gets a goal and works through it. It reads the order, looks up the customer in the inventory system, creates the sales order and gets back to you when something does not fit. Three things made this practical this year.
Models keep going for longer. OpenAI reported in June that more than 70 percent of sampled users of its coding agent Codex had given it at least one task that would have taken a person more than an hour. Two years ago, agents fell apart after a few steps. Today they summarise their own progress when their memory fills up and carry on.
There is a standard for connecting systems. With the Model Context Protocol (MCP), a system is connected once and every major model can work with it. OpenAI, Google, Microsoft and Salesforce support it. Connecting your ERP or ticket system is no longer a one-off build for one particular model. When a better model arrives next year, the connection stays and the agent simply gets better.
Agents become addressable. With an account of its own, an agent works in the tools your team already uses: the inbox, the chat, the document. Nobody has to learn new software to hand it work. You just write to it.
What this looks like in a company with 50 people
The examples in the announcements come from banks and large corporations. DBS Bank in Singapore has chains of 70 to 80 specialised agents prepare credit memos. ICON, a clinical research organisation, uses agents to generate analysis code for clinical trials and is about 30 percent faster at it.
I find it more interesting what this means for companies without an AI department of their own. A few tasks I find in almost every mid-sized business:
- Incoming orders. Orders arrive as PDFs, as email text, sometimes as a photo of a fax. An agent reads them, checks items and prices, creates the order and only asks the sales office when something is ambiguous.
- Quotes. A request comes in. The agent finds comparable quotes from past years, recalculates with current prices and puts a draft in front of sales. Sales negotiates instead of copying spreadsheets.
- Follow-ups. Overdue deliveries, quotes without an answer, unpaid invoices. An agent has the patience to check every day and send a friendly reminder.
- Questions about your own numbers. "Why was September revenue below last year?" Management asks the agent, which reads the data from the ERP and backs its answer with figures, instead of commissioning a report and waiting a week.
None of these tasks needs superintelligence. They need an agent that is connected to your systems and knows how you work. That is also the difference between the agent Google and Microsoft ship with their office software and one that gets work done in your business. The bundled agent knows your inbox. It does not know your ERP.
How to start
Like with a new colleague. Nobody gets power of attorney over the bank account on day one, but nobody makes a new hire watch for three months either.
- Pick a task that comes back every week and where someone currently carries data from A to B. The more boring, the better.
- Describe the goal the way you would to a person. What a good result looks like, where the data lives, whom to ask when something is unclear.
- Let it suggest first, then let it act. In the first weeks the agent prepares drafts and a person clicks send. Once the drafts are right, it does it on its own.
- Measure on real cases. With work from your actual day, not with a demo.
After a month you know whether it holds up. Usually the team has already thought of the next task it would like to hand over.
What comes next
Google and OpenAI already have agents that split a task, hand the parts to sub-agents and assemble the result. The first retailers are deciding which outside agents may shop with them. From there, it is not a big step until your agent negotiates delivery dates with your supplier's agent.
Companies that start with one task now will, a year from now, have a team that knows how to work with agents. That is the real head start, and it cannot be bought at short notice when you need it.
If you want to know which tasks in your business are a good fit: in the AI Potential Check I pick three of them with you, with effort, benefit and a fixed price for the first agent.
Sources
- Google Cloud: Gemini at Work 2026, 8 October 2026
- TechCrunch: Google brings agentic AI to Gemini, starting with businesses, 8 October 2026
- OpenAI: Introducing the Agents API, 10 September 2026
- TechTarget: AI agents are running into the limits of access, 9 October 2026
- R&D World: ICON puts clinical-trial agents into production, 9 September 2026