Headcount is the frame. Human capital is the story.

Every company dataset starts with headcount. How many people worked there, and how that number moved. It’s useful, it’s everywhere, and AI models already know it well. When we asked two current frontier models for the 2019 headcount of 150 listed US companies, they landed within about 5% of the filed figure at the median and called the direction of change correctly 92% of the time.

Headcount isn’t where the edge is. The edge is what sits inside it: which functions a company was built on, which capabilities its people brought, how long they stayed, and how all of that shifted year by year. That’s what the Vivameda panel records, for 4.2 million companies from 1950 to 2020. This week we put that record to the test.

What Frontier Models Don’t Know

We asked the same models about private companies, this time about their inside rather than their size.

First, how each company was staffed by function in 2019. Both models did worse than a simple sector average. On the share of staff in engineering they were 7 to 8 points off, where the sector average was about 5. They named the largest function correctly 57 to 60% of the time, against 69% for the sector average.

Second, how each company’s capabilities changed. We gave the models a shuffled list of skills from a company’s workforce and asked which became more prominent between 2016 and 2019 and which faded. They got 52 to 55% right. A coin flip gets 50. The plain industry trend gets about 70.

Public companies are well covered in the material models learn from. The inside of private companies isn’t.

Does It Matter for What Happens Next?

A record is only worth something if it says something about the future. So we checked what became of the companies. For 253,739 companies in the panel we established their status today: still operating, absorbed into another company, or gone. Then we asked a hard question of 85,311 US private companies with known outcomes.

Companies built on four or fewer functions in 2019 disappeared at 1.5 times the expected rate. Broadly built companies of the same size, industry and growth came in at 0.85 times. Companies where one function held 60% or more of the staff closed at 1.27 times the expected rate. On headcount these companies look identical. On human capital they don’t.

Then we trained a model. One version only saw the headcount frame: size, growth and industry. The other also saw the human-capital layers. Tested on companies it had never seen, the human-capital model caught 31% more of the companies that later disappeared, in its highest-risk tenth. We repeated it on six independent splits of the data, and the gain stayed between 26% and 34% every time.

Reading Isn’t Learning

One more result surprised us. When we handed the frontier models the full workforce history of each company inside the question, their predictions didn’t improve at all. The signal is there, but a model can’t pick it up from a few lines of text. It has to be trained on the record.

Historical human capital is training material, not background reading. A history book of how companies were built from the inside, with the endings attached.

Why the Past Matters

None of this can be collected again today. Of the companies active in 2019, about 10% no longer exist and another 11% have been absorbed or renamed. Their workforce history isn’t sitting on a live profile waiting to be gathered. The record was assembled over a decade, from multiple sources, and it reflects companies as they were, not as they look now.

Headcount tells you how big a company is. Human capital tells you what it’s made of, and that’s what decides whether it’s still there in seven years.