🌟 Vasilij’s Note
This week crystallised something I keep coming back to with clients: the tools are ready, but the decision to act isn't. Claude Cowork is now inside Microsoft 365 via Copilot Cowork. Anthropic's Mythos model has leaked - and it's being described as the most capable system they've ever built. Claude paid subscriptions have more than doubled this year. And yet, most consultancies I speak to are still waiting for the "right moment" to start. There is no right moment. There is only a first workflow. The firms building discipline around one repeatable task now will be the ones with a genuine operational advantage in twelve months. The ones waiting will be reading about why they fell behind.
In Today's Edition:
This Week in Agents | What Changed
Copilot Cowork rolls out to Microsoft Frontier - Claude is now inside M365 for enterprise. Microsoft shipped Copilot Cowork to its Frontier programme on 31 March, powered by Anthropic's Claude. The agent runs in the cloud within a tenant's Microsoft 365 environment - accessing Outlook, Teams, SharePoint, Excel, and calendar data simultaneously. It is designed to manage multi-step workflows, not just answer single queries. The M365 E7 bundle is priced at $99/user/month; the standalone Copilot Cowork add-on at $30/user/month. Windows Central
Anthropic leaks Claude Mythos - a model tier above Opus. A configuration error exposed a draft blog post describing Mythos, a new tier of Anthropic model larger and more capable than Opus 4.6. It reportedly scores dramatically higher on software coding, academic reasoning, and cybersecurity benchmarks. Anthropic is working on cost and compute efficiency before any general release.
Claude paid subscriptions more than doubled in 2026. TechCrunch reported that Anthropic's paid subscriber base has grown at record pace this year, with the majority of new subscribers on the lowest-tier plan. OpenAI remains the largest consumer AI platform, but the gap is narrowing.
Top Moves - Signal → Impact
Microsoft bets on Anthropic over OpenAI for its flagship enterprise agent.
Despite a $13 billion investment in OpenAI, Microsoft built Copilot Cowork on Claude. The product uses Anthropic's Cowork engine, runs inside M365's security and compliance infrastructure, and is tied to Work IQ - Microsoft's intelligence layer across emails, files, meetings, and calendars. Microsoft's own VP of Copilot described Claude Cowork as demonstrating "the value of agentic capabilities" before Microsoft commercialised it at scale.
→ The multi-model enterprise is now real. If you are advising clients on AI infrastructure decisions, the question is no longer OpenAI vs Anthropic - it is which model fits which task within a governed stack.
MCP crosses 97 million monthly installs - it is now infrastructure, not an experiment.
Anthropic's Model Context Protocol reached 97 million monthly SDK downloads in March 2026, up from roughly 2 million at launch sixteen months ago. Over 5,800 community and enterprise servers now exist. Every major AI provider ships MCP-compatible tooling.
→ If you are still evaluating whether to build agentic workflows on MCP, that decision has been made by the ecosystem. The question now is which connectors you prioritise - and whether your governance is in place before your team starts building.
UK CMA publishes binding guidance on AI agents - businesses are legally responsible for what their agents do.
On 9 March, the Competition and Markets Authority published "Complying with Consumer Law When Using AI Agents," making clear that existing UK consumer protection law applies in full to agentic AI. Businesses deploying agents in consumer-facing roles face fines of up to 10% of global annual turnover for breaches. The CMA's central warning: agents optimised for commercial outcomes - not user outcomes - risk constituting unfair trading practices, regardless of whether the manipulation was intentional. The watchdog stopped short of proposing new legislation, instead stating that current law already covers the risk.
→ For consultancies advising clients on agentic deployments - particularly in financial services, retail, or any consumer-facing context - this guidance is now live and enforceable. The four requirements to know: transparency about AI use, compliance by design, human oversight of agent outputs, and fast remediation when errors are identified. If you are building or recommending agent workflows that touch consumers, a compliance review against this guidance is not optional.
Upskilling Spotlight | Learn This Week
Claude Cowork Changelog - what shipped in March.
Thirteen new enterprise connectors launched this month alongside scheduled tasks, computer use in research preview, persistent agent threads via Dispatch, and the Sonnet 4.6 default model update. If you configured Cowork more than four weeks ago, you are running a materially different product to what is available today. The changelog is the most time-efficient way to close that gap.
Anthropic's Computer Use documentation.
The 23 March research preview for Pro and Max subscribers allows Claude to open apps, click, type, and navigate a Mac screen on your behalf. The documentation covers what it can and cannot do reliably, and where human-in-the-loop checkpoints are required before client-facing deployment.
Maker Note | What I built this week
This week I recorded a complete walkthrough of how I set up Claude Cowork to run sales operations for a consultancy - replacing workflows that would otherwise require a dedicated £65–80K sales ops hire. Six connectors, five custom skills. Prospecting, call prep, follow-ups, pipeline reviews, win/loss analysis. The full video is below.
Decision: the gap at most consultancies isn't capability, it's configuration. Partners are spending 70%+ of their week on non-selling tasks - not because the tools aren't available, but because nobody has mapped the workflows and built the skills around them. That's what this video covers.
In this video, I show you how to set up Claude Cowork as your AI-powered sales operations layer - six connectors, five custom skills
Operator’s Picks | Tools To Try
Apollo.io - Use for lead enrichment and qualification inside a Cowork prospecting workflow. Feed in a LinkedIn post URL or company list; Apollo returns company size, sector, job titles, and contact data. Standout: 275 million+ contact database with API access that pairs cleanly with Apify for LinkedIn scraping.
Caveat: data accuracy varies by geography - EU contact data is patchier than US.
Unipile - Use for LinkedIn outreach automation once Cowork has qualified your prospect list. Handles connection requests, follow-up message sequences, and supports WhatsApp and X channels. Pair with: Apollo for enrichment, Cowork for message personalisation.
Caveat: outreach sequences still need human review before scaling - platform detection policies change without notice.
n8n - Use as the automation backbone connecting Cowork skill outputs to your team's existing tools. Particularly useful for routing pipeline review outputs to Slack or triggering CRM updates from Cowork's follow-up drafting skill without touching the CRM manually. Open-source; self-hostable for firms with data residency requirements.
Deep Dive | Thesis & Playbook
The Consultancy Sales Ops Problem - and the £100/Month Fix
Salesforce's State of Sales survey found that sales professionals - across all industries - still spend 72% of their week on non-selling tasks. At consultancies with 15–50 staff, those tasks do not fall on a sales ops team. They fall on partners who are simultaneously billing 40 hours a week. The cost of a dedicated sales ops hire at UK rates is £65–80K in year one when salary, NI, pension, and recruitment are factored in. At five to ten BD leads, the business case for that hire does not exist.
The result is a predictable pattern: pipeline management becomes a side task, follow-ups do not happen, proposals go out late, and margins thin. Deltek's 2025 benchmark put professional services EBITDA at 9.8% - the lowest in five years. The status quo is not a holding pattern. It is a slow deterioration.
On Paper
Cowork launched in January 2026. By March, it had thirteen enterprise connectors, a plugin marketplace, scheduled tasks, computer use in research preview, and persistent Dispatch threads via mobile.
The core proposition for a sales workflow is six connectors - HubSpot, Gmail, Notion, Apollo, Apify, and Unipile - and five custom skills that handle prospecting, call prep, follow-up drafting, pipeline review, and win/loss analysis.
At Team plan pricing, that stack costs approximately £100/month for five users. The full setup including Apollo and HubSpot Starter runs to £60–100/user/month.
In Practice
The prospecting skill reduces 2–3 hours of manual research and outreach to 10–15 minutes per run.
The call prep skill - scheduled to fire at 7am each morning - delivers an HTML briefing for every meeting before the partner opens their laptop.
The follow-up skill reads Notion call notes, checks the deal stage in HubSpot, reviews the email thread, and drafts a personalised follow-up for review in under two minutes.
The pipeline review skill flags stale deals, prioritises active ones, and surfaces lost leads worth re-engaging.
The win/loss analysis skill generates a monthly report that most consultancies either pay a consultant to produce or simply never see.
None of these skills are perfect out of the box. Each requires 2–3 iterations of criteria refinement before it runs reliably.
The discipline of writing explicit acceptance criteria is itself the value - it forces the documentation that makes any automation sustainable.
Issues / Backlash
Computer use remains a research preview and should not run unsupervised in client-facing workflows.
Scheduled tasks require the desktop app to stay open - cloud execution has not shipped.
LinkedIn outreach via Unipile needs human oversight; platform detection policies evolve and message quality degrades without periodic review.
Skills built on poorly structured CRM data will reflect that structure.
Data quality in equals data quality out.
My Take (What to do)
Startup (15–40 staff):
Install the six connectors this week. Run the call prep skill manually for five working days before configuring the schedule. Time the difference. That delta is your ROI case for every other skill in the stack.
SMB (50–120 staff):
Assign one person as your Cowork owner before building. Not an additional role - make it an explicit 2-hour weekly responsibility for someone already in ops. Document one workflow per month before automating it. The documentation is the work; Cowork runs the output.
Enterprise (150–250 staff):
Run a structured comparison of Cowork versus Copilot Cowork for your sales workflows this month. Your M365 estate makes both relevant. The key difference is data location: Cowork runs locally, Copilot Cowork runs in your M365 tenant. For firms with data residency requirements, that distinction matters before you deploy at scale.
How to Try (15-minute path)
Connect HubSpot and Gmail to Cowork. Verify both connectors are reading your live data. (5 min)
Ask Cowork to generate a call prep briefing for your next meeting. Review the output against what you would prepare manually. (5 min)
Write down the three things the briefing got wrong or missed. That list is your skill refinement brief. (5 min)
Success metric: a call prep output you would actually use before a real client meeting, produced in under two minutes without prompting.
Spotlight Tool | Claude Cowork
Purpose: desktop AI agent for professional services teams that cannot justify a sales ops hire, but cannot afford the pipeline discipline failures that result from not having one.
Edge: composable skills architecture means you build once and reuse across every partner's workflow - call prep, follow-up drafting, pipeline review, and win/loss analysis all run on the same connector stack.
→ HubSpot, Gmail, Notion, Apollo, Apify, and Unipile connectors in one setup
→ Scheduled tasks for recurring workflow automation
→ Dispatch for mobile-to-desktop task delegation
→ Computer use (research preview) for hands-off execution
→ Plugin marketplace for pre-built and custom skills
What did you think of today's email?
Sponsored - Partner
n8n – An open‑source automation platform that lets you chain tools like DeepSeek, OpenAI, Gemini and your existing SaaS into real business workflows without paying per step. Ideal as the backbone for your first serious AI automations. Try: n8n
Did you find it useful? Or have questions? Please drop me a note. I respond to all emails. Simply reply to the newsletter or write to [email protected]