Teams Insider

Copilot Cowork vs. Scout with Microsoft's Alev Tamer and Chris Wheeler

Prefer audio?

Who should listen: Microsoft 365 service owners, Copilot programme leads and partner architects deciding between cloud-based Cowork and local desktop agents, and anyone fielding questions about Anthropic model processing, EU data boundary scope or agent governance.

Guest: Alev Tamer and Chris Wheeler — Sr. Partner Solution Architect; Senior Copilot Solution Engineer, Microsoft

Alev Tamer, Sr. Partner Solution Architect, and Chris Wheeler, Senior Copilot Solution Engineer aligned to UK healthcare, unpack the Copilot announcements made the day before at Microsoft Build. They compare cloud-based Copilot Cowork with the local Scout/"ClawPilot" desktop agent, explain the multi-model and critique approach, and work through the data residency questions European and NHS customers are raising about Anthropic processing.

Many thanks to Shure, sponsor of this episode.

Key insights

Multi-model isn't just a picker: pairing models so one drafts and another critiques catches hallucinations that a single training set misses. On the DRACO benchmark (Deep Research Accuracy, Completeness and Objectivity), Research with Critic scores 57.4% against 31.7% for OpenAI's deep research on o3-mini. ▶ 5:34
The practical split: Copilot Cowork currently runs Anthropic's Claude models only, while Scout (formerly nicknamed ClawPilot) lets you select from OpenAI, Anthropic and Microsoft AI models — but you're locked to one model per session, so a different model means starting a new chat. ▶ 7:05
Scout does computer use — it opens a real browser under your credentials — which is why Tamer used it to trawl an internal marketplace for partner offers and build a performance-review deck in 20 minutes. Cowork cannot do that today. ▶ 10:10
Cowork's advantage is the cloud container: Wheeler fired off 18 persona-based demo packs (Excel, Word, PowerPoint dummy data plus prompts), shut the laptop and came back hours later. Local desktop agents need the machine open and awake — hence the joke about selling plastic props to stop laptops closing. ▶ 11:10
Skills are just markdown instruction files, and the industry has converged on them — SharePoint, Cowork and Copilot all take them, so a demo-build or dashboard skill written once is reusable across tools even though each tool's execution capabilities differ. ▶ 16:15
Scheduled autonomy needs care: Tamer runs a filterable announcements dashboard that refreshes at 5am daily, but warns against granting blanket "do it all the time" permissions on mail and OneDrive — colleagues have ended up with missing files and emails. ▶ 18:48
The enterprise blocker is accountability, not capability: if an agent impersonating a user schedules a PowerShell script, admins need to log and attribute it. Expect execution containers, policy layers and Agent 365 conversations before local runtime agents get past desktop admins. ▶ 19:20
On Anthropic data handling: processing is still US-based, but follows the same "other M365 services" clause in the data protection addendum as PowerPoint Designer, spelling and grammar, and Excel advanced analytics. Data at rest is unaffected, processing is stateless, the EU Data Boundary doesn't cover the UK, and flex routing lets EU customers burst to US compute when capacity is short. Blocking Anthropic means losing Researcher with Critique and Council. ▶ 25:27
Availability at time of recording: Cowork requires the Frontier programme with Anthropic models enabled, and Scout — the first inflection of Autopilots — is also Frontier-only with additional security configuration to complete first. ▶ 35:42