Microsoft Scout — Microsoft’s first “Autopilot”, built on OpenClaw
The biggest Microsoft 365 news from Build. Microsoft Scout is Microsoft’s first desktop-based, always-on personal agent — a fork of the open-source OpenClaw project, productised and secured for the enterprise. Microsoft is framing this new category as Autopilots: constantly running agent processes with their own identity that act on your behalf, sitting alongside Copilot as part of the family.
The key difference from Cowork: Scout runs locally on your machine as a persistent process (models still in the cloud), so your machine needs to be on — but in return, you get scheduled and recurring tasks, and a heartbeat that keeps work moving without prompting.
What can it do? It can read and write files, code, run scripts, and access your workplace data via Work IQ. It connects to Teams, Outlook, OneDrive, and SharePoint. You can interact with it in Teams, and it extends to your browser, local resources, and MCP servers. If you've followed OpenClaw or Hermes (which I'm currently using and am impressed with), it's the same idea.
To try it today, you need to be on Frontier, enrolled in early access via an Intune policy configuration and an opt-in attestation, and have a GitHub Copilot licence for the compute (setup instructions here).
Look out for two Teams Insider Podcasts coming where we go deep on this.
A Copilot “Super App” is coming
We got official word at Build that a Copilot Super App is on the way, bringing Chat, Cowork, Code, and Autopilot together into a single desktop app experience. If you’ve been following Anthropic and OpenAI, they’ve both started merging their separate scenarios into single desktop super apps — Microsoft is following the same pattern. Due this summer.
Work IQ APIs GA on 16th June — watch the consumption pricing
The Work IQ APIs are generally available from 16th June. This is a proper API surface onto your Work IQ data — via agent-to-agent (A2A), MCP, or directly via REST. There’s a deeper technical breakdown on the Microsoft 365 dev blog. I’ve been experimenting with these in different interfaces, and it’s genuinely handy to hit Work IQ data programmatically.
The important thing to understand is the commercial model. These are consumption APIs, even if you have an M365 Copilot licence.
If you’re consuming Work IQ through first-party experiences (Copilot, Cowork, and I hope Autopilots and Scout), it’s included in your Microsoft 365 Copilot licence. But hit the API from anywhere else — your own processes, third-party apps, or other AI tools like Claude or ChatGPT — and you’re paying per call in Copilot Credits.
Microsoft’s own scenario examples put even the “light” scenario at 20–40 cents per call, up to around $1.50 for the heavy scenario. Run heavy queries on a schedule, or point a busy third-party agent at it, and this gets expensive per user fast.
Microsoft also announced Web IQ, extending the same IQ framing to web data for agents. And for a look at how this plays out in practice, Microsoft’s Inside Track team has published a piece on how Microsoft itself uses Work IQ to power AI and agents internally — useful reading if you’re planning your own agent scenarios.
Claude Fable 5 lands in Copilot — with a 30-day data retention catch
Last issue, it was Opus 4.8 rolling out in Cowork; now Anthropic’s Claude Fable 5 is available in Microsoft 365 Copilot as a preview model. Fable 5 is the first of Anthropic’s “Mythos-class” models, built for long-horizon autonomous work.
Default off in Copilot Cowork (Frontier), with private preview in Copilot in Excel and PowerPoint. It’s also generally available in GitHub Copilot, again off by default.
The catch is data retention. Unlike every other Claude model, which can run with Zero Data Retention, Anthropic requires 30-day retention of all prompts and outputs on Mythos-class models — on every surface, first- and third-party — to run its safety classifiers. The data isn’t used for training and is deleted after 30 days in most cases, but it sits on Anthropic’s side. Microsoft has had to surface this explicitly: the Copilot announcement tells admins to review the retention consideration as part of their enablement decision, and in GitHub Copilot, enabling the Fable 5 policy constitutes acknowledgement of the requirement.
The telling detail: Microsoft has reportedly pulled Fable 5 from its own employees’ internal model picker while legal reviews the retention terms — earlier Claude models stay available internally because they’re Zero Data Retention (ZDR).
Mustafa Suleyman has told Bloomberg that Anthropic’s models are too expensive at enterprise scale, and Microsoft is cancelling the majority of its internal Claude Code licences by 30th June, moving thousands of engineers across Windows, Microsoft 365 and Surface to GitHub Copilot CLI.
Seven new MAI models — including Microsoft’s first reasoning model
Mustafa Suleyman’s Microsoft AI team launched seven new MAI models at Build, and they’re available in Microsoft Foundry across text, image, voice and speech. Some are new versions of models we’ve seen before, and they’re getting genuinely competitive — Microsoft claims its image model is now up there with NanoBanana. The standout is MAI-Thinking-1, Microsoft’s first reasoning model. Simon Willison has a good breakdown of the specifics: MAI-Thinking-1 is a 35B parameter model currently with select early partners, and MAI-Code-1-Flash is a 5B model purpose-built for GitHub Copilot and VS Code.
The strategic context matters here. Under the old OpenAI agreement, Microsoft used OpenAI models and didn’t build competitors. The renegotiated deal freed Microsoft to develop frontier models in-house.
Six Microsoft 365 Copilot updates for June
A quick roundup of what’s new across the Copilot release notes and the What’s New in Microsoft 365 Copilot blog:
- Targeted organisational adoption messages — target by usage data (e.g. no Copilot use in 28 days), now with email as a transport and a template library.
- Federated connectors are now GA — Microsoft’s blessed third-party MCP connections (Google Calendar, Notion, Moody’s), accessed via Copilot Studio.
- Copilot in Word edits documents by default — what was “agent mode” is now the default. Exactly what we want.
- Copilot in model-driven apps — a Copilot pane that can interact with and drive your app’s data (we use model-driven apps for ChangePilot).
- Agent evaluation in Copilot Studio — audit agent responses and see which models perform best for your scenarios.
- Policy-based agent automation — auto-rollout of agents, plus managing and reassigning abandoned or ownerless agents.
A rebuilt Copilot Studio
Copilot Studio itself has been rebuilt for more complex, multi-step work — a new streamlined authoring experience and a modern AI core aimed at making agents and workflows more consistent and reliable. Given how much of the agent governance story (evaluation, policy automation, federated connectors) routes through Copilot Studio.
Windows and Hardware for the Agent Era
Alongside Scout, we got an extensive demo of OpenClaw running on Windows. The Windows team has been working directly with the open-source project and has done significant security and containerisation work so that on-device agents can be controlled and observed, with future ties into Agent 365 and the option to run local models alongside cloud ones.
NVIDIA is producing its own laptop chips — the new RTX Spark — and the Surface Laptop Ultra is Microsoft’s take, with major OEMs involved too. These are very high-end machines — lots of RAM, bandwidth and compute for running local models. Satya framed it at Build as “unmetered intelligence”: pay for the beefy hardware once rather than paying cloud consumption per token.
Project Solara — a platform for agent-first devices, built on MDEP
One of the most interesting announcements at Build. Project Solara is a new chip-to-cloud platform for agent-first device experiences — shown with two proof-of-concept devices: a desk unit with Hello for Business (think Amazon Echo territory) and a compute badge with display and camera. Importantly, these aren’t products, and Microsoft isn’t planning to ship devices — the idea is OEMs pick up the platform and build whatever form factors make sense: watches, glasses, anything.
The MDEP team has its own post on powering new devices and agent-first experiences, with more on the platform thinking.
Chris L Johnson on the MDEP team helped explain it to me.
Teams Rooms and Devices
Announced at Cisco Live and now certified: the Cisco Board Pro G3 in 55-inch and 75-inch variants for Microsoft Teams Rooms (certified hardware list). They’re looking really nice — and with the Microsoft UC User Group London being hosted at Cisco’s office in September, I’m looking forward to seeing them in person.
Last issue I flagged QSC joining MDEP and said to expect device news soon — here it is. Q-SYS announced their first video bar, a Microsoft Teams Rooms on Windows bar, alongside an 11-inch QSP-11 scheduling panel that’s just been certified for MDEP. Interesting to see them coming into the bar market — David Danto has a closer look at the full Q-SYS June announcements on TalkingPointz.
Barco also announced another bundle, this time pairing with the Logitech MeetUp 2.
I caught up with Barco, Pexip and Shure at the Microsoft Teams Bootcamp this week, and Irena Andonova gave me a rundown of what's coming in Devices and the Pro Portal.
As always, if you have any questions, corrections, thoughts, or feedback, please reply. I read all the replies. I really do want to hear your thoughts.
Thanks
Tom
New on Empowering.Cloud
Teams Insider Podcast
- Microsoft Teams Facilitator: The First Group AI Agent for Meetings With Madhu Sudan, Microsoft
- Microsoft Teams Call Quality Dashboard (CQD): Intelligent Classifiers, Silent Test Call and Power BI - James Parkes, Senior Product Manager for Microsoft Teams Call Quality Management, Siunie Sutjahjo, Principal Product Manager for Microsoft Teams Meeting, and Victor Guzman, Senior Technical Program Manager.
- Microsoft 365 Message Center and M365 Change Explained with Brian McGough, Principal Program Manager
- Microsoft 365 Agents Explained - Declarative, Copilot Studio, Pro-Code or Skills in Copilot Cowork? with Graham Hosking, Senior AI Solutions Engineer at Microsoft
Apple Podcasts, Spotify, Amazon Music and YouTube