TechWolf

Remember our Summer launch? That one was about going faster. This season is about trust. Your agents are fast and smart. They just don't know your workforce. This quarter, we fixed that. We put trusted, governed Work, Skills and Market Intelligence inside the agents and tools your teams already use, Excel, Copilot and Claude included. So they can redesign work and plan the workforce on your reality, not a guess. Three are live now: → Agentic Work Redesign: your roles and tasks, the skills behind them, every task scored from AI-driven to human-driven. → Agentic Workforce Planning: skill demand and supply, market cost, and task-based modelling of future demand. → Market Intelligence: the global labor market, in your own skills and tasks. Our CPO Yasamin and our Director of AI Jens-Joris break it down in the teaser below, and there's more coming. The pumpkin spice latte didn't make it to set, so grab your own and hit play. Explore the Fall Launch

📣 Issue two of our co-founder Mikaël Wornoo🐺's summer series digs into Unified Work Embeddings (UWE). How do you measure AI's impact across 1,500 companies when no two describe work the same way? The same task gets written ten different ways, and one job title can mean very different things depending on who's hiring. Before you can compare anything, you have to know what's actually the same underneath. That's what UWE does. It reads past the labels to the work itself, mapping 2,500 skills to 50,000 tasks across 20,000 jobs to power our Work Intelligence Index. Where JobBERT read job titles, UWE learns the structure of work itself: a focused model that outperforms general ones many times its size. Mik breaks down how, link's in the comments 👇

Issue four of Mikaël's summer series lands today. When Nestlé sold off its water business, a whole new company had to be stood up from scratch. Mik reconstructed most of that build from the outside, using 3,200 public job postings and nothing else. No access to their systems. Turns out job postings, read in bulk, are the revealed preferences of how a company is designed: where it's hiring, where it isn't, what it's missing. The analysis that used to take months of consulting (and a 500k-1m bill) now takes about an hour. Have a look, link in the comments 👇

📣 New Substack series. How do the TechWolf AI models actually work? Where do they differ from other models in the market? And why is it crucial to invest so heavily in AI research? After almost a decade building an AI company in the work domain, our co-founder Mikaël Wornoo🐺 took some time to unravel some of TechWolf’s AI research and models. Straight from the lab, onto his Substack. First up, JobBERT. How do you teach a computer that "Java Rockstar", "Developer II" and "Backend Engineer, Ohio" are all the same job? That's what JobBERT does. Our open-source model for making sense of jobs, now the most downloaded in HR tech at 2M+ downloads. Mik breaks down how it turns messy titles into a map, plus the career tool Morad Cherradi at UCB built on it. Big credit to Jens-Joris Decorte, our Head of AI, who put his PhD into this problem and is a huge part of why JobBERT works as well as it does. More issues to come. The JobBERT one's in the comments 👇

Search 500 million job ads for "API testing" and you'll only find 5. The demand is everywhere, it's just written in a hundred different ways. That's why counting keywords fails: the same word means different jobs, and most skills are never named outright. So we built and open-sourced ConTeXT, a model that reads instead of counts. It maps every sentence to the right skill, at roughly 7,000x less than running it through GPT-4. Now we can see how work is changing, job by job, down to the skill. The latest from our co-founder Mikaël Wornoo🐺 , link's in the comments 👇

I started writing an article about WorkRB: the first open-source AI benchmark for the work domain. At first glance, it’s a shared yardstick for the industry, academia & governments. But I wanted to know “what exactly are we benchmarking?” I ended up going down a rabbit hole that started with model evaluation, went through 100 years of intelligence research, and ended with the realization that IQ tests and AI benchmarks are converging on the same problem. After writing this, I’m more excited than ever about the progress of AI in the work domain. Kudos to Matthias De Lange, PhD for bringing the industry together in this way 🙏🏽🫡