AI News: EU Enforcement, Agent Boom and New Features (Oct 2026)

Two weeks, five developments: the EU moved into enforcement mode, the agent-based AI market accelerated further, and major platforms began shifting their business models towards agents. Here is what each of these means in practice.

The EU AI Office Issued Its First Formal Requests

The European AI Office sent formal requests for information to the developers of leading AI models in early October. The reason: AI agent control-loss incidents documented over the summer, which occurred at several major developers. The AI Act's enforcement powers have been in effect for months; the office has recorded several similar cases throughout the year. No fines have been issued yet, but the law allows for serious sanctions calculated as a percentage of global annual revenue.

What does this mean for you?

The “wait and see whether they'll actually do anything” phase is over. The AI Office has authority, and it is using it. This is not a message directed only at large platform developers: it indirectly affects every company integrating AI agents into its operations. If an agent is running in your systems, think about who supervises it, what it does, and how you can demonstrate accountability if questions arise. Responsibility does not end when the developer hands over the keys.

Agent-Based AI Among European SMEs: No Longer a Promise, Now a Number

According to OpenAI's 'Europe's Agentic Opportunity' report, the number of weekly active users of the Codex and ChatGPT Work agent tools among EU small and medium-sized businesses has multiplied over the past six months. Every member state recorded growth. The acceleration picked up sharply after ChatGPT Work launched last summer, and agent-based applications now account for the majority of outputs generated by EU corporate subscribers.

What does this mean for you?

An OpenAI report is not exactly a disinterested source on its own products, but the direction of the metrics is instructive regardless. Agent-based use is not a niche for tech enthusiasts: it is spreading into everyday business work. Most business owners we encounter still think in terms of chatbot-style, question-and-answer interactions. Agents go further: they execute steps independently, call tools, connect systems. If your competitors are already there, it is worth knowing where you stand.

ChatGPT Business Can Lock and Organise

ChatGPT Business subscribers can now lock individual conversations with a PIN or device authentication. Locked conversations are excluded from search and history but are not deleted. With an expanded Admin Console, workspace owners and administrators can manage membership and workspace settings directly. Administrators can also activate or disable dozens of plugins at once from the console.

What does this mean for you?

If your company already uses a ChatGPT Business subscription, it is worth spending some time in the Admin Console. The lockable conversation feature is particularly relevant where commercially sensitive topics enter your prompts: legal matters, client data, strategic negotiations can be kept protected. The plugin controls address exactly the concern many IT managers have: that a colleague switches something on that should not have been. Not a dramatic change, but a useful step for operational control.

Microsoft Splits the Copilot Bill in Two

Within Microsoft 365 Copilot, Microsoft is separating core daily AI features, which remain part of the fixed-fee subscription, from advanced agent-based capabilities, which will be billed under a consumption-based Copilot Credits system. SharePoint and OneDrive Copilot features are generally available; agent-based, scalable operations will be billed from the Credits system. Administrators have until the end of October to set spending limits if they want to restrict automatic agent and extension charging within their organisation.

What does this mean for you?

If your company has a Microsoft 365 Copilot subscription, the end-of-October deadline is a real task, not an administrative formality. If your IT administrator does not set a spending limit, agent-based operations can charge automatically. The “we'll see what it costs” approach can produce a concrete, unplanned bill. The core features are unchanged; agent-based capabilities extend the system beyond them. But cost predictability drops if you do not get ahead of it in time.

Anthropic Shows Where the AI Assistant Is Heading

Anthropic has unveiled the Claude Duet application as a research preview for iOS, Android, and the claude.ai web interface. The feature is currently only available in the United States, for Pro and Max subscribers, via a waitlist. Unlike a conventional chatbot, Claude Duet is proactive: it can initiate and independently execute tasks on the user's behalf.

What does this mean for you?

There is nothing to act on for now: it is not available in Europe. The direction, however, is worth noting. AI assistants have so far reacted; now they are beginning to act. The boundary between agent and assistant is dissolving, and a tool that does not only answer but also takes steps requires a different kind of trust from the user. If this reaches Europe, the interesting question will not be the feature itself, but this: what decisions are you prepared to let an automated system make in your name?

Sources

• EU Announces First AI Act Enforcement Actions After AI Agent Incidents – Unite.AI, 2026-10-09
• OpenAI EU Agentic Opportunity Report – CellCog / OpenAI, 2026-10-14
• ChatGPT Business: lockable conversations and expanded Admin Console – Releasebot / OpenAI, 2026-10-21
• Microsoft Copilot: separating core and advanced AI feature billing – ChangePilot, 2026-10-15
• Anthropic launches Claude Duet proactive AI agent app – Releasebot / Anthropic, 2026-10-21