What can you learn about a company from the outside?

Computing the Labor Market

SubscribeSign in

What can you learn about a company from the outside?<br>Why job postings are revealed preferences of org design

Mik Wornoo<br>Jul 27, 2026

Share

Nestlé sold its water business in 2021, and that business became BlueTriton. We have about 3.200 of their job postings from April 2021 to March 2023. No access to their HR internal systems. How much of the company build process can you recover from just postings?<br>A lot. We can see the entire company being built in real-time. See below. It was a question we were recently exploring with a private equity firm. Can you use job posting data to inform the transaction services agreement?<br>Thanks for reading Computing the Labor Market! Subscribe for free to receive new posts and support my work.

Subscribe

While one posting can be noisy and have a lot of boilerplate text, thousands tell us a lot more than you would think.

How companies understand their work

Every company has an underlying structure of work: jobs, tasks, skills, and the relationships between them. That structure can’t really be observed directly.<br>What is observable is the job architecture. For many companies, it’s the only enterprise-wide view of work that is available. Think of it as their own labor classification system. Their own ONET.<br>It’s mostly used by the compensation team in HR. A job architecture groups people into manageable pay categories, so it is coarse by design. A company with 50 people labeled Cyber Security Specialist has a cloud security engineer, a SOC analyst, and an IAM specialist sitting inside one code, all doing materially different work.

An example job architecture<br>Understanding your work is step 1 when modeling the impact of AI .<br>As companies do this, they are forced to pick between two compressed structures: (1) ONET, the government classification. Has 900 occupations. 1.7M people in the software developer category alone. Too coarse. (2) Their job architecture. Optimized for compensation. Backward-looking. Usually outdated by 5-7 years.<br>One of our customers has 15.000 employees and 2.000 different job titles. Maintaining a structure granular enough by hand is impossible (or really expensive). Consulting firms usually charge between 500K-1M for a full refresh.<br>Can we do better?

In statistics, this problem is called latent variable modelling. The thing you care about exists, but you never observe it directly. You only observe its consequences.<br>One of my favorite examples comes from biology.<br>Proteins belong to families: groups of proteins that perform the same biological function and evolved from a common ancestor. The family itself is never observed. What we observe are individual protein sequences.<br>Evolution constantly edits those sequences. Amino acids are inserted and deleted over millions of years. Two proteins performing roughly the same function can share only 20–30% (!) of their sequence.<br>Yet they still carry the fingerprints of the same underlying structure.<br>Computational biology learns those hidden relationships from millions of examples. By looking at patterns across many protein sequences, it reconstructs latent protein families and can even identify entirely new family members.<br>The same idea applies to work.<br>The true structure of a company’s work is also latent. We never observe “how work is really organized”. We observe imperfect models of it. The job architecture is one model. Designed around compensation.<br>Job postings are another, and sometimes an even more powerful one.<br>And we can use job postings to estimate the underlying structure that the job architecture is trying (and often failing) to represent. The Nestle/BlueTriton example is just one way that can happen.<br>We can aggregate across companies as well. Take the forward-deployed engineer . Palantir started hiring these in 2005. Everybody else followed about two decades later, and demand has absolutely exploded: about 400 FDE postings in 2024. 3,640 in 2025. From 2021-2025, we have about 5,180 postings in total. Analyze them, and 6 archetypes emerge that cover about 66% of postings:

Postings are revealed preferences of org design

Another way to think about job postings is as the revealed preferences of org design. Open headcount. Resource allocation. Geographic decisions. Seniority decisions, and so much more. One job posting doesn’t tell you that much. Not all roles get published either. But… combining historical job posting data gives you a ton of insight into corporate decision-making.<br>Two more examples below.<br>Setting up a standalone company after a divestiture. One of our customers is spinning off an entity. By comparing historical job posting data (2021-2025) of the corporate HQ, the SpinCo, and similar standalone companies, we found that the SpinCo was relying on HQ shared services for Tax and Internal Audit. Two functions the new business will have to build internally.

Reading seniority structure. If we...