What’s Changed in AI This Week? Microsoft Copilot Agents, Meta Muse Wonder, and the Governance Shift
What’s Changed in AI This Week? Microsoft Copilot Agents, Meta Muse Wonder, and the Governance Shift
Microsoft gives AI agents their own identities and memory, Meta launches Muse Wonder, OpenAI reports growing EU adoption and Anthropic opens its systems to independent scrutiny.
Microsoft Copilot Studio is introducing agent accounts, business process memory and automatic triggers, bringing AI closer to independently managed workflows. Meta Muse Wonder is targeting longer, more complex agentic tasks, while OpenAI reports more than eightfold growth in weekly active agentic AI use among EU SMBs since May. Meanwhile, Anthropic is giving independent evaluators greater access to assess model risks and practices.
AI agents are getting their own accounts, remembering what they've been working on and picking up tasks without someone having to prompt them, which is… quite the shift from opening ChatGPT and asking it to draft an email.
At the same time, Microsoft's giving agents more independence inside everyday business tools, Meta's making a play for longer, more complex tasks, and OpenAI's latest figures suggest businesses are already embracing the change.
Meanwhile, Anthropic's opening up its models to greater independent scrutiny, which raises another question: who's checking what these increasingly capable systems can actually do?
There's plenty to unpack this week, particularly for businesses looking to put AI to work without losing control of their data, processes or costs, so without further ado, let’s dive into this week’s roundup.
Headlines at a glance
Microsoft Copilot Studio
Agent accounts, business process memory and automatic triggers move agents closer to managed process workers, with clearer identity and auditability.
Meta Muse Wonder
Meta’s new models target long-running professional and coding tasks, but availability and independent validation still matter more than the launch claims.
OpenAI EU adoption data
OpenAI says weekly active use of Codex and ChatGPT Work among EU SMBs has risen more than eightfold since May, a strong signal that agentic use is moving beyond pilots.
Anthropic external evaluation
Six METR evaluators now have employee-like access and publication rights, raising the bar for what meaningful independent AI oversight can look like.
Microsoft Copilot Studio gives agents their own accounts, memory and triggers
Microsoft has announced new Copilot Studio capabilities for persistent business-process agents. The important pieces are agent accounts, business process memory and triggers in the GitHub Copilot harness.
Agent accounts give an agent their own approved credentials and scoped permissions inside Microsoft 365, meaning colleagues can then assign it work through Teams or Outlook, or @mention it in comments in Word, Excel, and PowerPoint. Actions can also be attributed to the agent, rather than the individual employee’s identity who summoned it - so to speak - which makes reviewing outputs considerably easier.
Business process memory also lets an agent carry forward relevant preferences and process context between cycles, while triggers allow an external event or a schedule to start work without a fresh prompt. Put those together, and you feasibly have an agent that could prepare a weekly budget review, chase missing explanations and retain unresolved questions, while people keep authority over approvals and spending.
The caveat: these capabilities are rolling out through Microsoft’s Frontier programme, and Microsoft says further access, quality, cost and lifecycle controls are planned over the coming months. So this is more of a ‘direction of travel’ and an early deployment opportunity, as opposed to you booting it up and handing every recurring process to an agent by Monday afternoon at the latest.
Start with one process that is repetitive, has clear inputs and outputs, and already has a named owner. Give the agent a distinct identity, the minimum permissions it needs, visible escalation points and a human approval step for anything consequential. If nobody can explain who owns the agent when its maker changes role, for example, the process isn’t ready.
Meta launches Muse Wonder for longer agentic work
Meta introduced Muse Wonder and Muse Wonder Full, its most capable models to date, for personal and professional agentic tasks. Meta highlights two ideas: Active Memory, which carries useful context across conversations, and Active Focus, which is designed to keep working through difficult tasks, recover from errors and ask for clarification when needed.
Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is available to Muse Code subscribers. Meta says the models cover expert domains including law, finance and health, alongside coding tasks. It also claims Muse Wonder will power more experiences in Muse, beginning with subscribers (as if we didn’t need MORE subscriptions to stay on top of, right?).
Now, we’re not saying every SMB should switch mofels to Meta’s offerings this week - or, ever. But it’s interesting that Meta is trying to compete directly for market share in that layer where agents remember, persist and act - rather than doing what every SaaS platform seems to be doing, and treating AI as a feature tucked inside their service, which in their case would be Facebook or WhatsApp. More credible competition should improve choice for developers and, eventually, pricing leverage for buyers.
The caveat: the performance claims are Meta’s own, and they’ve famously never lied to anyone, while they go on to say that a fuller safety evaluation report is still to come. Long memory and greater persistence are useful only when permission boundaries, deletion, audit trails and error recovery are equally mature.
For most firms, Muse Wonder belongs on a watchlist or in a bounded technical evaluation. Test it on a real but non-sensitive workflow if you’re curious, record the baseline time and quality, and check how it behaves when information is missing or an action needs approval. A memorable agent that remembers the wrong thing isn’t a productivity gain, it’s just yet another AI platform.
OpenAI says agentic AI use among EU SMBs has grown more than eightfold
OpenAI’s new analysis of agentic AI use across the EU offers a useful view of how quickly the market is changing. OpenAI defines SMBs in this analysis as organisations with 200 or fewer associated users and says weekly active users of Codex and ChatGPT Work among EU SMBs have grown more than eightfold since May 2026.
Among the ChatGPT Enterprise customers in OpenAI’s EU sample, Codex and ChatGPT Work now account for more than half of output tokens. The company also found a wide spread in adoption intensity between countries and says the most intensive EU users still trail the global frontier. In other words, access is widening faster than deep operational use.
Now, obviously, this is vendor telemetry and not an independent measure of productivity, and token volume doesn’t tell us whether a process became faster, cheaper or better. Still, the direction appears to be that businesses are moving from asking AI for an answer to asking it to complete a body of work.
Don’t use adoption as the success metric. Pick a small number of workflows and measure cycle time, rework, error rate, human review and total cost before and after. If usage rises but the checking burden rises with it, you have more AI activity, not necessarily more productivity.
Anthropic expands independent AI model evaluation
Anthropic has begun giving a six-person team from the nonprofit METR employee-like access to its internal tools and systems for model-risk assessment. The access excludes defined areas such as customer data, privileged legal information and model architecture details.
The most important part is publication independence, as METR can publish its assessment of Anthropic’s risks, incidents and practices. Anthropic cannot veto publication, although it can require narrowly defined confidential information to be redacted; METR can publicly characterise those redactions and whether it considers them justified.
Anthropic is also working with SecureBio and Accenture under more limited arrangements. This is not a regulator certifying that Claude is safe, and it doesn’t remove the need for customer due diligence, but it’s more substantive than asking an external party to test a polished model shortly before launch and calling that independent oversight.
Procurement teams should ask AI suppliers what independent evaluators can actually see, whether findings can be published without approval, how incidents are disclosed and what evidence is available for the specific model or service being purchased. A safety claim is more useful when someone outside the vendor has meaningful access to challenge it.
Quick answer: What should SMEs do this week?
- Choose one recurring process that could benefit from an agent, then map its owner, data, permissions, approvals and failure path before testing a tool.
- Create separate identities for operational agents wherever the platform supports it, with minimum permissions and audit logs that distinguish human actions from agent actions.
- Measure business outcomes rather than prompts, tokens or time spent inside the product. Include review time and exception handling in the calculation.
- Add independent evaluation and incident transparency to your AI supplier checklist, especially for tools that can take actions across business systems.
Fifosys View
Giving an AI agent its own credentials, memory, and the ability to work independently sounds brilliant on paper if it’s taking care of the tedious, manual admin. That is, until it makes a mistake, accesses something it shouldn't, or nobody knows who's responsible for it - and that's the challenge businesses need to get ahead of.
You don't need to deploy agents across every department to see value from AI. Start with something useful, understand what information it needs, decide what it's allowed to do and make sure someone's accountable for the outcome.
And please, whatever you do, don't mistake ‘more AI activity’ for better productivity. If your team spends as much time checking an agent's work as they would've spent doing it themselves, what's really been gained?
We're exploring exactly these challenges in our upcoming AI in Business webinar on the 23rd of October. We'll be looking at governance, security, integrations and costs, along with what businesses should consider as they move beyond individual AI experimentation.
Join us for AI in Business: Part 2
AI's becoming a bigger part of how businesses operate. Making sure it's working for your business, rather than creating more work behind the scenes, is where the real value lies.
This week's AI news FAQs
What are Microsoft Copilot Studio agent accounts?
Agent accounts give AI agents their own approved identities and scoped permissions within Microsoft 365. This helps organisations distinguish actions performed by an agent from those performed by an employee, improving accountability and auditing.
What is business process memory in Copilot Studio?
Business process memory allows an AI agent to retain relevant preferences and process context between recurring tasks. Combined with event-based or scheduled triggers, it can help agents continue work without needing a new instruction every time.
What is Meta Muse Wonder?
Muse Wonder is a family of AI models introduced by Meta for longer-running professional, coding and agentic tasks. Its Active Memory and Active Focus capabilities are intended to help agents retain context, work through complex tasks and recover from errors.
How quickly is agentic AI adoption growing among EU businesses?
OpenAI reports that weekly active use of Codex and ChatGPT Work among EU organisations with 200 or fewer associated users has increased more than eightfold since May 2026. These figures describe product adoption, not independently verified productivity gains.
What has Anthropic changed about independent AI evaluation?
Anthropic has given six evaluators from the nonprofit METR employee-like access to selected internal tools and systems. METR can publish assessments without Anthropic having a general right to veto its findings, although narrowly defined confidential information may be redacted.
How should businesses govern autonomous AI agents?
Every operational AI agent should have a named owner, a separate identity where supported, minimum necessary permissions, audit logging and clear escalation routes. Consequential actions should require human approval, with regular reviews of performance, cost and access.
How can SMEs measure whether AI agents are delivering value?
Measure outcomes such as process completion time, rework, error rates, human review time and total operating cost. Increased prompts, token consumption or agent activity don't necessarily mean improved productivity.
Make AI work for your business, without losing control
From identifying useful processes and connecting the right systems to managing access, costs and governance, Fifosys helps businesses take a practical, consultancy-led approach to AI adoption.
Talk to our team
Explore where AI could create measurable value in your organisation, and the safeguards needed to support it.
AI in Business: Part 2
Join our next webinar for a practical look at AI governance, security, integrations and costs.