For a long time, I have argued that AI is not the new 3D TV. It is not blockchain-for-everything. It is much closer to the internet, spreadsheets, or ESG: once it becomes part of the operating system of work, everything around it changes. This month, that argument has become much easier to make. The reason is not only that the models are better (although they are). It is that models, agents, and deployment tools are all improving at the same time. We now have frontier models that are excellent at office work, agents that can actually complete tasks, and senior leaders who increasingly understand why this matters. That combination feels significant.
The new models are aimed directly at work
The new GPT-5.5 seems really good. What impresses me is not only that it is smarter in some general sense, but that it is so good at the things professionals actually spend their lives doing: Excel modelling, data analysis, document drafting, coding, research, and turning messy inputs into usable outputs.
It is also remarkable value for money. Compared with many enterprise AI tools, ChatGPT still feels absurdly capable for the price. When I need rigour, research depth and a convincing argument, I usually find it the strongest option. It is sharp, direct, and often quite unsentimental. Sometimes that hurts my feelings. Usually in a useful way.
Claude Opus 4.7 is also important, particularly because it seems to be aimed at office-type work and coding agentic workflows. My sense is not that one model now wins everything. For research-heavy work and reasoning, I still often prefer ChatGPT. For structured business task execution, Claude can be excellent. The important point is that the frontier has moved straight into the work done by analysts, researchers, asset managers, consultants, executives and academics. It’s no longer about benchmarks on math and coding.
Copilot Cowork is also becoming more stable and capable. When it works it is a true powerhouse giving AI (Claude Opus 4.7) access to a virtual machine that can work within your Microsoft environment using your files. It will work for a long time until the task is completed even with large numbers of files. It still gives me an error message occasionally, but when it works, it really hits the spot! Copilot in PowerPoint and Excel has also gotten significantly better.
Agents turn AI from answering into doing
The bigger shift is agents. For the last two years, most people used AI as a chatbot. You asked a question and it gave an answer. That was useful, but limited. Agents change the interaction model. Instead of asking, 'Can you answer this?', we can increasingly say, 'Here is a task. Go and do the work.'
ChatGPT workplace agents and Anthropic's managed agents are both moving in this direction. They connect to applications, work with files, follow instructions, use skills, and can be taught to complete tasks in a particular way. ChatGPT currently feels easier for non-technical users, while Claude's agent infrastructure can be very powerful once configured properly.
ChatGPT workplace agents and Anthropic's managed agents are both moving in this direction. They connect to applications, work with files, follow instructions, use skills, and can be taught to complete tasks in a particular way. ChatGPT currently feels easier for non-technical users, while Claude's agent infrastructure can be very powerful once configured properly.
This is a tremendous capability. I have deployed agents to amazing effect. They can clean files, prepare summaries, check documents, draft reports, update analysis, organise information and run workflows that used to require repeated manual effort. In real estate, the use cases are obvious: investment memo drafts, lease extraction, DCF checks, market updates, tenant summaries, ESG reviews, file clean-up and recurring internal briefings.
The hard part is translating work into instructions
This is also where many people will misunderstand agents. They will see a demo, assume the agent just does the work, try it on a real task, get a mediocre result, and conclude that agents are overhyped. That would be the wrong conclusion.
The hard part is not usually connecting the tool. It is translating what is in your head into something AI can understand and follow. Most professionals underestimate how much tacit knowledge they use. When you ask a good analyst to prepare an investment memo, you are relying on hundreds of assumptions: which sources matter, what risks are material, what tone is appropriate, what the investment committee cares about, and what 'too optimistic' looks like.
Most of that knowledge is not written down. Agents force us to make it explicit. That is uncomfortable, but incredibly valuable. To make an agent work well, you need to describe the task, the data, the process, the judgement calls, the output, the review criteria and the failure modes. In other words, good agents require operational clarity.
Senior leaders now understand the significance
The most encouraging change I see is not technical. It is cultural. When I speak to senior leaders now - CEOs, partners, board members and investment leaders - the conversation is very different from last year. I no longer spend most of the time explaining that AI is not a toy. I no longer need to convince people that this is not another overhyped technology trend. Many of them get it. They may not know the latest model names or which tool is best for which workflow, but they understand the significance.
That is new. It means we are moving from 'Is AI real?' to 'How do we use it properly?' That is much more productive, but also more difficult. Once people accept that AI matters, they start forming strong opinions about what it can and cannot do. Some of those opinions are based on serious use. Some are based on one bad Copilot experience. Some are based on a demo. The right posture is humility. AI is changing too quickly for anyone to say they are finished learning.
What this means for real estate firms
We are seeing the shift directly in our courses. The recent public course (in partnership with ULI and HSG) has just finished and was very well received (it was full and had a waiting list). Our private courses are also seeing renewed interest. Companies are no longer asking only for inspiration. They are asking how to deploy AI safely and effectively.
The use cases are no longer theoretical. Two or three years ago, it was obvious to some of us that AI would eventually be useful for financial modelling, data extraction, document processing, file clean-up, research, strategy, automation and internal knowledge work. Now companies are actually doing it. Not perfectly. Not autonomously. Not without supervision. But they are doing it.
This is why I think AI is now universally consequential. I do not mean that AI is equally useful for every task. I do not mean that it replaces human judgement. I do not mean that every company should build agents for everything and let them loose across confidential systems. I mean something simpler and more important: every professional activity that involves text, data, documents, spreadsheets, images, code, communication, analysis, research or decision support is now affected by AI.
The companies that understand this will not simply buy AI. They will redesign workflows around it, train people properly, build agents for repetitive but valuable tasks, and define where automation is safe, where delegation is useful, and where human judgement must remain central. Buying access is not the same as building capability. That is the real lesson of this month. Very exciting times. Also slightly terrifying. Which, in my experience, is usually where the interesting work begins.
Created with
