I started this month expecting to write almost entirely about Claude. That would have been reasonable because Claude has become genuinely extraordinary over recent months, enough to reshape how I think about daily knowledge work. But as I reflected on what actually mattered in practice, I realised that focusing on one model would miss the bigger story. What stood out instead was something more fundamental: Automation is no longer theoretical. It is operational, accessible, and increasingly unavoidable.
From Assistant to Agent: Claude CoWork
Claude remains the right place to begin. The most interesting development is Claude CoWork - an experimental Mac interface built on Claude Code. Beneath the somewhat clunky name sits something far more ambitious: an agentic framework with direct access to your files, browser, and local machine. It builds applications. Creates presentations. Downloads and analyses documents. Browses the web. Operates your computer on your behalf. In short, automates everything you can do on a computer. This goes beyond a chatbot - it's software that acts.
For years we've discussed AI as an "assistant" and augmenting work using AI. Claude CoWork feels closer to an actual colleague, it doesn't just answer questions, it completes tasks and makes (pretty reasonable, but not always correct) assumptions about what you'd like it to do along the way. If you haven't experimented with this type of system yet, I'd strongly recommend doing so. This is what the next phase of knowledge work looks like.
The Unexpected Win: Copilot Workflow Builder
What surprised me was that Claude wasn't the only place meaningful progress appeared. Another genuinely useful capability came from an unexpected source: Microsoft Copilot. I'm not generally a Copilot enthusiast, and my scepticism has been well earned. But Copilot's Workflow Builder turned out better than expected.
After some experimentation, I created an automated workflow triggered by incoming Outlook emails (LinkedIn post here). The system scans each message, detects the language, and when it identifies German, automatically forwards it to Thomas (whose German is considerably better than mine). This may sound trivial, but it worked reliably. Once I understood how the tool wanted to be used (which took a while), setup took less than a minute.
More importantly, it illustrates the broader point: automation is quietly seeping into everyday processes that used to require manual intervention and doing so inside tools people already use. All you need to do is work out what automation you need.
Raising the Ceiling: Manus and End-to-End Workflows
That direction becomes even clearer when you look at platforms like Manus. Manus currently operates at an entirely different scale. It can scan the internet for opportunities in specific markets, identify listings on portals, download brochures, analyse unstructured text, identify missing information, compare locations, and report back with structured insights. It builds discounted cash flow spreadsheets, estimates potential returns, and automatically follows up when key data points are missing.
Once configured, these workflows run end-to-end with no (or minimal) human involvement. I got it to complete a proof of concept (chat history here) from a couple of prompts. It wasn't perfect, but definitely impressive. This is the crucial point: Perfect isn't a glimpse of a distant future. It's already possible today.
The bottleneck is no longer technical capability. It's imagination, governance, and judgment.
Governance Isn't Optional
Open Source Catches Up: Kimi and Local Deployment
Alongside automation, another shift is happening in parallel: open-source models are catching up at an astonishing pace. A new model called Kimi was released recently, and it's genuinely impressive. Not quite at the level of the very latest frontier models, but roughly comparable to where ChatGPT was a few months ago. That gap is shrinking fast.
The implications are profound. We now have open-source models that can be downloaded, run locally, and integrated into private workflows - while being only marginally behind the best proprietary systems. For anyone concerned about data privacy, regulatory exposure, or reliance on external providers, this changes the conversation entirely. There's very little excuse not to experiment with local deployment.
Even more interestingly, Kimi supports skills. When deployed locally, it can build spreadsheets, structure data, and execute complex workflows directly on your machine.
In other words: you can generate financial models tailored to your specifications without sending your data anywhere.
For real estate (where bespoke analysis and data sensitivity matter enormously) this should be setting off alarm bells in the best possible way.
The Real Advantage: Redesigning Work, Not Just Accelerating It
Stepping back, the pattern is clear. Automation is arriving across our devices and throughout our processes. Costs are falling rapidly. Capabilities are improving month by month. The competitive advantage won't come from "using AI" because everyone will do that.
It will come from:
• Choosing the right processes to automate
• Designing new workflows rather than digitising old ones
• Deploying systems in a way that is safe, thoughtful, and governed
This isn't simply a story about adaptation. It's a story about experimentation. The organisations and individuals willing to rethink how work is structured, rather than just accelerating existing routines, will benefit most.
These are, without exaggeration, very exciting times.
Created with
