The prediction economy
The prediction economy is the phase in which the winners in vertical markets are no longer the companies that automate the work, but the ones that see the market through the work. It is the third level of value creation in software: after selling the tool and selling the outcome, companies sell what their distribution lets them see.
The three levels
| Level | What is sold | Share of the value created that the seller keeps | Why it ends |
|---|---|---|---|
| 1. Sell the tool (SaaS) | Software the customer runs | 2–5% | AI rebuilds most software in weeks |
| 2. Sell the outcome (service as software, AI rollups) | The finished work | 30–60% | Every operator buys the same automation |
| 3. Sell what you see | Early signal on how a market is moving | Set by what outside buyers pay | Does not end: history cannot be copied |
Worked example
A sales tool saves a team 30% of its time and closes 20% more deals. On a subscription, the customer keeps about 95% of that value. Run the same work as a service and the provider keeps 30 to 60%. Now hold the platform across 1,500 tax advisors in one country: the tool and the service still run, but the asset is seeing which sectors stretch their cash cycles months before any statistic shows it.
The idea is old. In June 1815 Nathan Rothschild's private courier network told him the outcome of Waterloo a day before the government knew. The couriers were not a messaging business. They were sensors.
How it differs from
- Prediction markets. Kalshi and Polymarket let people trade contracts on events. The prediction economy is not about betting; it is about who owns the early view of a real market.
- Prediction Machines. Agrawal, Gans and Goldfarb argue AI makes prediction cheap. The prediction economy argues the opposite input stays scarce: ground truth about the present.
Questions
Why would anyone stop charging users? Because friction slows distribution, and distribution is the sensor grid. The user becomes the instrument, and a third party pays.
Does superintelligence make this obsolete? No. Self-play works on closed worlds like Go. An economy has to be observed; no model can deduce this quarter's margins in Austrian dental practices from first principles.
What would prove it wrong? Three dated predictions are on the record: at least five more acquisitions of vertical information businesses by serial software acquirers by end of 2027; a paid alternative-data panel in at least one US vertical by end of 2028; and a state institution buying a real-time SME panel by 2030.
Origin
Defined in this sense by Viktoria Izdebska in The Prediction Economy (Part 1, 29 July 2026) and the essay series The Prediction Economy (1/4) to (4/4).