Eight ways AI comes for a data business. First, your data gets cheap.

A business loses its price in two ways. Someone makes more of what it sells, or the people who bought it stop needing it. AI is doing both to data businesses at once. This issue is the first kind: four ways what you sell gets easier to find somewhere else. The next issue, on 11 November, is the second kind.

None of the four is new. What is new is that three of them now have a deal behind them, and the fourth never will. And none of them hits a data business in general. Each one finds a specific choice the business made, usually years ago, usually without noticing it was a choice.

1. The buyer builds it.

This is the one with no press release. Take a buyer that licenses 400 fields and uses three of them. An analyst with an agent can now rebuild those three from filings, websites and the buyer's own records in an afternoon, and nobody announces it. It shows up as a renewal that comes back smaller, with a polite email about budget. The choice underneath is what you let cross. If what crossed was compiled from sources the buyer can reach, the buyer was paying for your labour, and labour just got cheap.

2. A competitor catches up.

On 8 September Circana added Meridian, the marketing-mix model Google gives away as open source, to its Liquid Mix product, combining it with Circana's own first-party data and measurement. Anyone can download the model. So the model cannot be what the client pays for. That is what parity looks like. When the method is free, everyone has it, and the price of anything that was only method falls toward the cost of copying it. Circana has to earn its fee on what cannot be downloaded. A data business whose price rested on its processing is pricing something a competitor can now copy.

3. The platform faces your buyer.

On 10 September PitchBook and Quartr both became data partners for OpenAI's ChatGPT for Financial Services. The analyst who used to open PitchBook can now ask a chat window, and PitchBook is one of the sources behind the answer, next to Quartr and whoever signs next. That is the supply problem in this attack: in someone else's window you are one supplier among several. The partner now fronts the buyer, and the party facing the buyer holds the stronger hand at renewal. PitchBook drew a line. The connector carries the firmographic facts; the judgement products, the scores and the valuation estimates, stay in PitchBook's own platform. The facts went to the agent. The judgement stayed home. On 30 September LSEG and Snowflake extended their partnership for five years. LSEG's licensed data now sits inside customers' own Snowflake accounts, where their AI tools use it without opening LSEG's product. Same risk as PitchBook, one layer lower: not the chat window the analyst reads, but the platform where the data is stored and worked on.

4. The model swallows the product.

CuriosityStream reported $14.1 million of licensing revenue for the second quarter, up 48 per cent on new partnerships. The figure mixes traditional media licensing with AI deals and does not split them; the AI deals include code licensed for training and footage licensed to a video research lab. That is real money today. What it sells is the right to learn from you, and a model that has learned from you needs you less. Thomson Reuters held the line: on 29 September a US appeals court ruled that ROSS infringed its copyright, and was not making fair use, when it trained a legal-research tool on Westlaw's headnotes. One licensed its material to the models. The other went to court to stop a competitor training on it.

The ranking.

The four do not hit every data business equally. Rank them by what each needs in order to work. The buyer building it and the competitor catching up need data compiled once, from sources others can reach. The model swallowing the product needs data that stays true once learnt. Those three take compiled data first. The platform attack is different: it needs only that your buyer prefers someone else's window, so it reaches anyone who let a partner face the buyer, whatever the data.

Data that has to be collected again every day, from sources nobody else holds, is the last to fall to those three. A model cannot learn tomorrow's reading today. It still needs its own window, or the platform takes it anyway.

Which of your defences only worked while compiling was expensive?

Next issue: the demand side, four ways the buyer stops needing you.

Compiled data goes first.