Growhold
Two businesses ask for AI. One says add AI to our app. The other says make AI answer from our support history. The second request pays a third more, goes custom three times out of five, and its share of demand is climbing fast.
Two businesses ask for AI this week.
The first says: add AI to our app. The second says: make AI answer from our four years of support tickets.
Same technology. Completely different purchases. The second one pays roughly a third more, becomes a custom build three times out of five instead of two, and carries the highest complexity in the pool. The first one is comparing quotes on a commodity.
The market has noticed what many roadmaps have not: the models are identical for everyone with a card. Whatever an AI feature does for your product, it does for your competitor's product by Friday. What cannot be replicated by Friday is what the intelligence gets to read.
Buyers now name the data first
Right now, one in sixteen requests for AI names a specific body of records the intelligence must stand on: customer history, files and documents, a knowledge base. Not "add AI". Make AI work on what we have. And that share keeps climbing.
Requests that name records the model cannot reach on its own, as a share of all demand, indexed to the start of the window.
When the request names its data, it changes species.
The same technology, a different class of purchase, decided by what the intelligence gets to read.
Intelligence is rented by the token. The data underneath it is what actually gets bought.
Where the requests point
These buyers are not experimenting. They point the AI at the processes that produce new records every day.
The destinations are the record-producing processes of a business, not the experimental ones.
Lead flows first. Then analysis and extraction of what already exists. Then the support inbox, the content archive, the CRM. And the stores these requests name are deliberately ordinary: HubSpot, Google Sheets, Excel, Supabase. Nobody is asking for a data platform. They are asking for the one input AI cannot commoditize, exactly where it already lives.
What a builder does with this
The data does not have to belong to the buyer. It can be yours. Three concrete moves, in rising order of strength:
- Read what the customer already holds. "Answer from our tickets", "score leads the way our pipeline history says", "draft in our voice from our archive". The integration is unglamorous: HubSpot, a spreadsheet, a database. The moat is that the output is right for this customer and generic for everyone else.
- Make your product produce records nobody else has. Every answered question, every qualified lead, every correction a user makes to an AI draft is exhaust that only your product collects. Instrument it from day one. Six months in, your AI stands on something a competitor cloning your features on the same model cannot buy.
- Do both. Arrive with your own data asset and dig into theirs. Each month of use deepens a moat the customer is helping you build. This is the strongest position in the current market, and it is why a feature demo and a defensible product can look identical on launch day and be worlds apart a year later.
What a product cannot do anymore is arrive carrying only a model.
So flip the roadmap question. Not "where can we add AI". Ask: what does the intelligence in our product get to read that nobody else's can? If the answer is data you gathered, you have a moat. If it is data your customer holds, you have a moat they deepen for you. If the answer is nothing, you have a demo, and the market keeps moving while you decide.
