World models, and why HR needs one
In April I went to Imagination in Action, an AI conference at MIT. Someone used the term world model. It had nothing to do with HR, and the people using it did not think it was a new idea. I have been thinking about it ever since, because it named something I had been trying to build for years without having a word for it.
Physicists have had these for a century
A world model is a model of how a system behaves. You give it a state and you give it a force, and it tells you what happens next.
Gravitation is the easy example. A model of gravitation does not describe the solar system. It simulates it. Give it the masses and the velocities and it will tell you where the planets are in fifty years. Nobody had to observe those fifty years first. The same idea runs fluid dynamics, where you give it pressure and geometry and it tells you where turbulence forms, and climate models, where you give it the forcings and it tells you what the system does over decades.
This is old and well understood. The interesting part is how small the leap is to an organisation.
An organisation is a system too. It has a state: who is where, doing what, under whom, for how long, paid how much. It has forces acting on it: pay, managers, promotion, workload, the labour market outside. If you can describe the state and you know how the forces behave, you can simulate what happens next.
How this differs from predictive analytics
This is the question I get asked most, and it is a fair one, because on the surface both of them produce a number about the future.
Predictive analytics looks at what happened before and finds a pattern in it. It learns that people who have been in the same role for four years, whose manager scores badly, who have not had a raise recently, tend to leave. Then it scores your current people against that pattern and hands you a list. This is useful. I have built plenty of it and I am not going to talk anyone out of it.
The limit is what it can answer. A prediction tells you where you are heading if everything carries on as it is. Ask it what happens if you change something and it has nothing to say, because the change you are considering has not happened yet, so there is no pattern of it to learn from.
A world model is built the other way round. Instead of learning who leaves, it learns what moves leaving. How much a manager change moves it. How much a pay increase moves it, at which levels, and how long the effect lasts. Once you have that, you can set a state, apply a force that has never been applied before, and watch the organisation respond.
| Predictive analytics | World model | |
|---|---|---|
| Question | Who is likely to leave? | What happens if we do this? |
| Learns | Patterns in what happened | Cause and effect, and how strong |
| Handles a new decision | No, there is no history of it | Yes, that is the point |
| Output | A ranked list of people | A simulated outcome, with a cost |
| Good for | Spotting risk early | Choosing between options |
Both are worth having. They answer different questions and the second question is the one an executive actually asks.
What this looks like in HR
Take a decision you make every year. Someone proposes a three percent increase to base pay. Finance wants to know what it costs. Everyone in the room already knows the answer to that, because it is arithmetic.
Nobody in the room can answer the second question, which is what it buys. Some of that money comes back. Fewer people leave, so you spend less on recruitment and less on the months a seat sits empty. Absence moves. Productivity moves a little. The effect does not arrive on the same clock as the cost either, because the cost lands in month one and in full, while the return builds over a year.
A world model can walk that month by month and give you the net. The useful number is what the increase costs once the organisation has responded to it. The gross figure is only where the conversation starts.
The same shape covers most of the decisions HR is asked about. What a reorganisation costs in attrition eighteen months out. Whether filling senior seats from outside instead of promoting internally quietly jams the organisation and how long that takes to show. What a richer benefits package returns, priced in retention rather than in premium.
What it takes
A world model needs three things, and the first two are the reason so few organisations have one.
It needs connected data, joined at the person, so that the pay system and the performance system and the absence system are describing the same human being. It needs managed definitions, so that headcount means one thing across the whole model. Then it needs the relationships themselves, learned from your organisation rather than from a benchmark, and kept current, because a relationship measured three years ago in a different labour market is not reliable now.
None of that is exciting and all of it is the work. The model on top is comparatively straightforward once the layers underneath hold.
Why now
Two things changed. Enough HR data is now in one place often enough that learning real relationships from it is possible. And the language models everyone is talking about turn out to be the wrong tool for this, which has made the gap easier to see. A language model predicts the next word from everything ever written. It has never read your reorganisation, your pay round, or this morning’s vacancy list, and it never will, because that information arrives after the model is sealed.
What HR needs is a model of the next action inside one specific organisation, learned from that organisation, still true this quarter. That is a different thing, and physicists have known how to build it for a hundred years.