Eight ways AI comes for a data business. Then the buyer leaves the room.

Two weeks ago I listed four ways AI adds supply to a data business. The demand side is quieter and, I think, worse. Supply attacks lower your price. Demand attacks remove the reason anyone paid it.

5. The buyer shrinks.

This is the weakest of the eight, and I will say so. Equity analysts at the 15 largest banks fell to about 3,000 from almost 4,600 in a decade, on figures Bloomberg reported in January 2025. Almost all of that decade came before AI agents, and Bloomberg pointed at older pressures: regulation that unbundled research, money moving to index funds. Those who stayed cover more companies each. And FactSet says its user count is up 12 per cent. Data pointed at a decision taken by a shrinking job title shrinks with it, but so far the job is stretching, not vanishing. Watch the headcount of whoever takes your buyer's decision. It is a better forecast than your pipeline.

6. Seats turn into agents.

In July FactSet told investors it is "accelerating the shift in our business model from seat linked contracts to flexible enterprise agreements". On the same call it said users were up 12 per cent, mostly in the long tail, without much effect on revenue, and that "more work shifts from humans doing the work to agents doing the work". That is under-capture in two sentences: more use, little more money. PitchBook charges for agent access three ways and is testing a fourth, priced on consumption, which tells you the price of an agent is still being found. A seat was a stand-in for one analyst. The analyst now runs agents on it.

7. The decision leaves the person.

On 8 September Catalist Sports gave Kalshi exclusive US prediction-market access to its live streams, plus official low-latency data to create, run and settle the markets, across more than 65,000 tennis matches a year. That data's job is not to inform anyone. A market is settled on it. Data that informs a person's judgement competes with everything else the person weighs. Data a system acts on is either in the loop or out of it, and when it is in, its price can follow the actions. The decision used to be a person's. More and more it is a system acting, and data pointed at the person's view loses to data wired to the action.

8. Nobody opens the product.

On 16 September Memorial Sloan Kettering agreed to put OncoKB, its precision-oncology knowledge base, inside OpenEvidence, and to put OpenEvidence inside MSK's Epic records system. Once it is in, the oncologist gets the answer without visiting OncoKB. Bloomberg's Enterprise MCP lets client agents pull Data License Plus content directly. On Redpine, six publishers will let AI agents read more than 20 million research papers, each publisher setting its own price per transaction. For decades the screen was the boundary: a terminal showed the data and kept it. An agent does not look; it takes. A boundary that lived in the interface leaves with the interface. It has to move into the contract: what the agent may take, keep and pass on. Bloomberg has started there: clients license the specific content domains they need, under its standard data licence.

The ranking, all eight.

My ranking, by how soon each hits a data business that sells seats to people, worst first:

1. The platform faces your buyer. In market in finance now.
2. Seats turn into agents. Under-capture is already on the earnings calls.
3. Nobody opens the product. The boundary leaves with the screen.
4. The buyer builds it. No press release, only smaller renewals.
5. The model swallows the product. Paid well, while the model is still learning.
6. A competitor catches up. Slower than it looks where the data is maintained.
7. The decision leaves the person. Good news for data wired to an action.
8. The buyer shrinks. Not yet.

Read down the list and one pattern holds. Compiled data, sold to a person through a screen, is exposed to almost all eight. Maintained data, wired into an action and bound by its contract, is exposed to a few, and later.

Before you ask which attack is coming, ask what you were really being paid for.

Compiled data goes first. Maintained, acted-on data goes last.