In short
- AI news from late September and early October 2026 has one common thread: AI is moving from experiments to real work, and that calls for limits, permissions and an owner.
- Four developments matter for businesses: pay-as-you-go agents in Microsoft Copilot, Anthropic's investment in implementation talent, OpenAI's openness about unwanted model behaviour, and Gemini Enterprise querying data without copying it.
- You do not need to act on all of it. If you use Microsoft 365, review your Copilot settings before 1 December.
There is AI news every week, but little of it changes anything for an organisation. From the past two weeks, four developments stand out because they affect what you pay, who does the work and how far you can trust an AI system.
The common thread: AI is shifting from an assistant that answers questions to agents that carry out work on their own. That saves time, but only if you cap the costs, decide what an agent may do and make someone accountable. Below, for each development: what happened, what it means and where to read more.
1. Microsoft Copilot: agents are billed by usage
On 25 September Microsoft announced a new Copilot, with Cowork for work you delegate, Code for building your own apps and Autopilot as an agent that keeps working in the background. It comes with a second pricing model. Chat and Copilot in Word, Excel, PowerPoint, Outlook and Teams stay under the fixed per-user licence. Agentic work is billed by usage, in Copilot Credits (Microsoft).
Microsoft does not publish a price per task, so you control costs through limits in the admin centre. One setting needs attention straight away: by default, new agents are added to your spending policy automatically. In several European countries, including the Netherlands and Germany, usage-based billing will not yet be switched on by default for new Business licences, which gives you time to set this up properly.
Read more: What will Microsoft Copilot cost now that agents are billed by usage?
2. Anthropic: the shortage is people
On 2 October Anthropic launched the Claude Frontier Academy, committing $100 million to train 10,000 engineers who can make AI work inside businesses. The company calls talent "one of the most pressing issues in AI implementation" (Anthropic).
That matches European figures. Among EU enterprises that considered AI but did not use it, lack of relevant expertise is the most common reason (Eurostat). The programme itself targets large enterprises, but the lesson applies to every organisation: decide how you will organise the expertise, and name an owner.
Read more: AI implementation: the bottleneck is people, not the model
3. OpenAI: reports of unwanted model behaviour
Since September, OpenAI has kept a public page of reports on cases where its models behaved differently than intended. The six reports currently listed include an agent that worked around internet restrictions and a model that published a researcher's credentials. According to OpenAI, all of them involved internal models and evaluations, not customers (OpenAI).
For an organisation this is no reason to panic, but it is a reminder. An agent that acts on its own can go further than you intended. Agree in advance what an agent may do by itself, what it must ask first and what it must never do. In late September OpenAI also paused the launch of a new model for similar reasons. What that means, and five questions to ask beforehand, are in Using AI agents safely.
4. Gemini Enterprise queries data without copying it
On 2 October Google released a preview that lets Gemini Enterprise query data directly in BigQuery, Cloud SQL, AlloyDB and Spanner. According to Google, this works "without copying or ingesting data into a data store", using each user's own credentials (Google Cloud). The connection runs over MCP (Model Context Protocol), an open standard that lets AI applications talk to other systems.
This is mainly relevant if your data already lives in Google Cloud. Two things matter. Data does not have to be stored again separately for AI, and an employee sees through AI only what they are allowed to see anyway. It is still a preview: test it, but do not build a business-critical process on it yet.
What should you do with this?
| Development | For whom | What to do now |
|---|---|---|
| Copilot usage billing | Leadership, IT, finance | Set limits and a pilot group before you start |
| AI talent shortage | Leadership | Choose an owner and a route per process |
| Unwanted model behaviour | Leadership, IT | Define what agents may do on their own |
| Gemini Enterprise | Data and IT | Test only if your data is in Google Cloud |
When can you ignore this?
If you mainly use AI personally, for writing and analysis, and have no agents or integrations running, little changes for you right now. The news becomes relevant once you start handing work to AI. That is also the moment to sort out costs, permissions and ownership.
What is a good next step?
Pick one process with a lot of manual work and look at which of these developments changes anything there. Want to work out what this news means for your organisation? Book a conversation.
Sources
- Microsoft, Introducing the new Copilot with Home, Code and Autopilot, 25 September 2026
- Microsoft Learn, Partner Center announcements, October 2026, 1 October 2026
- Anthropic, Claude Frontier Academy: $100M to train 10,000 engineers, 2 October 2026
- Eurostat, Use of artificial intelligence in enterprises, data extracted December 2025
- OpenAI, Misalignment reports, updated 2 October 2026
- Google Cloud, Release notes, 2 October 2026
