Human conversations
When I was in my late teens, there was a time I wanted to be a journalist. The investigative type. The one who can go on a little adventure. Talk to people. Ask questions. Learn something new and connect the dots.
Journalism was not the path I ended up walking, but what stayed with me is using conversations to learn and understand things better.
In one of my previous roles, I worked with an executive coach. We explored my intrinsic motivators and “unique” capabilities. We assessed feedback from colleagues, and she drilled down to understand my way of working. After a few sessions, she summarised it as “elevating conversations”.
I remember thinking that it sounded uninspiring. Not “leadership under complex circumstances”, or “designing creative solutions for hard problems”.
No.
Elevating conversations.
Reflecting on it now, I’ve come to appreciate it more and see how it’s a helpful capability in this new era of artificial intelligence (AI).
What’s more, it sits at the core of my vision for human and AI collaboration, and, by extension, for Third Vector.
Machine conversations
I started to develop a deep interest in AI in 2014, when I first learned about it during my time at Deloitte Innovation. Coming from a finance background, with few AI people in my network to talk to, I took Andrew Ng’s machine learning course on Coursera in 2016 to better understand the technology. A few years later, when I was working at AI software company Tamr, we deployed supervised machine learning (ML) models to transform and unify massive datasets for large enterprises. With subject-matter experts in the loop training the models, customers could do at scale what had been impossible before. Using this new technology in combination with human expertise, we were unlocking millions in business value.
With machine learning, humans interacted with the AI model by providing training labels, examples of good and bad answers. The model applied what it had learned to unseen data and, after a couple of training rounds, often performed on par with, or better than, its teachers. The ML model learned from a very basic “conversation”.
Today, with the large language models (LLMs) powering Claude, ChatGPT, Gemini and others, you interact with the AI through language. Besides the training data and model weights, it is the quality of the conversations (or “prompts”) that drives the quality of the output.
We have all had the experience of getting a confident-sounding answer from an AI, only to second-guess it and realise it was flawed after a more thorough reading. Automation bias, the human tendency to over-trust AI-generated answers, is a real thing.
A high-quality conversation is one where you challenge the AI, stress-test its assumptions and check for logical flaws. If you treat these tools like a question-and-answer box, value is limited. But if you treat them as a collaboration tool, to sharpen your own thinking and capture what you learn, value compounds.
Judgement architecture
The kind of questioning described in the previous section has been around for more than 2,000 years and is known as the Socratic method.
From Wikipedia: The Socratic method is a form of argumentative dialogue in which an individual probes a conversation partner on a topic, using questions and clarifications, until the partner is pressed to come to a conclusion on their own, or else their reasoning breaks down and they are forced to admit ignorance.
This type of conversation helps you make better decisions. It improves judgement and makes the reasoning explicit.
As we hand over more complex and high-stakes tasks to AI, the reasoning behind them needs to be explicit. That explicit version is judgement architecture: how decisions are made, including thresholds, trade-offs and escalation triggers.
Important decisions are often made by experts who rely on tacit knowledge to apply judgement. Tacit knowledge is what experts know but have never written down, and often can’t easily put into words. As organisations adopt AI and delegate more to autonomous systems, turning tacit knowledge into explicit judgement architecture becomes a precondition for delegating well.
Take the example of a Dutch installer of solar panels and heat pumps. Five seasoned experts review every sales order, text and photos, before anything is installed. Does the roof have the right slope for the panels sold? Where should the heat pump go? They either send the order back to the sales rep or turn it into instructions for the installation team. On paper, both are checklist questions. In practice, every roof and every garden is different. Where the heat pump goes usually comes down to distance, because of noise rules, but it also needs enough airflow, and every choice affects how well the system performs. The experts weigh those factors together. A checklist tests them one at a time. Until that judgement is made explicit, the review can’t be handed to AI without losing what makes it work, and it stays with five people.
Within ten years, organisations will maintain their judgement architecture the way they maintain their financial records (updated and audited). Only part of that architecture can be found in their systems. System logs, email and chat history, and permission settings show what gets decided and by whom. Conversation is how you find out why.
In the age of AI, the Socratic method works in new ways. In addition to using it in conversations with humans, you can use it to question the AI, and the AI can use it to question you.
Third Vector’s thesis 1.0
Third Vector’s thesis is this: the data that matters most for AI in an organisation sits in no system. AI can automate large chunks of intelligence, but for complex tasks human experts rely on tacit knowledge: exactly the data no system holds. Drawn out through in-depth conversation, this knowledge becomes explicit judgement architecture, built into skills a whole organisation can run. Experts know which questions to ask. An AI equipped with a skill can ask those questions of everyone who uses it.
A year ago, the skills I am talking about in the thesis barely existed. Anthropic introduced them as Agent Skills on 16 October 2025. The Agent Skills standard was open-sourced by Anthropic in December 2025 and is now supported by OpenAI’s ChatGPT and Codex, Google’s Gemini CLI, GitHub Copilot, Mistral and other AI tools. That makes skills portable: the judgement an organisation captures isn’t locked to one model or vendor.
A skill in its most basic form is a text file in .md (Markdown) format that contains written instructions in plain language. An AI agent reads the file and follows the instructions. A skill is close to a process document. But process documents hardly ever get read, while skills get read every single time an AI runs them.
Compare that with the supervised models I described earlier. What they learned from training labels ended up in their weights, where nobody could read it. Anyone can read a skill, edit it, argue with it and improve it. That makes authorship possible: expertise becomes a tangible artefact you can own and improve over time.
Organisations that adopt skills in their way of working create capacity to pursue new, meaningful work. Skills capture specialised knowledge and provide a scalable way to share the expert’s judgement architecture with the broader team. This enables organisations to move from individual AI to institutional AI: the difference between each person’s private chats and shared skills the organisation owns and improves.
Intense curiosity
AI-native organisations need explicit judgement architectures. Where judgement is tacit, conversation is how you capture it. Skills are how the organisation runs it, with the expert keeping authorship and gaining reach.
Third Vector’s name holds the idea. A vector carries two things: magnitude and direction. And we are in the third wave of AI. First it was predictive (machine learning), then generative (LLMs), and now agentic, where AI systems make decisions on their own. That gives enormous magnitude: speed, scale, autonomy. But magnitude without direction just gets you to the wrong place faster. Inside the models, ideas are stored as vectors, and similar ideas point in the same direction. In organisations, direction comes from human judgement.
I’m building Third Vector with intense curiosity about applying AI, understanding and scaling expert know-how, and creating methods to capture tacit judgement. There is so much to discover. The investigative journalist in me gets his little adventure after all.
The approach is elevating conversations: in interviews with experts, and in skills designed to apply expert judgement and stimulate critical thinking when people engage with the AI.
I’m looking forward to our first conversation.
Photo by David Underland on Pexels. Thank you, David.


