01
The data nobody planned for
Projects plan for the data that goes in. The more valuable data is what comes out of use.
Questions:what people need, in their own words, including what no system answers yet.
Corrections:where an expert changed an answer, and why.
Decisions and outcomes:what was decided with the system’s help, and what happened next.
Usage:who uses what, when, and where they give up.
Definitions:terms the business had to agree on before the system could answer.
02
What doesn’t compound
More of the same records. A tenth year of transactions adds less than the first.
Raw logs nobody owns or reads.
Data you aren’t allowed to use for a new purpose.
03
Level by level
Check the contract either way: what the vendor may do with your prompts and outputs, and whether you can export the usage data.
04
What Level 4 lets you do with it
Statistical tests:did a change make the answers better, for whom, and is the difference real?
Behavior and usage analysis:where people get stuck, and which features carry the value.
Network and graph analysis:how documents, decisions and teams connect.
Documents and conversations into tables,then new prediction models on top.
Machine learning and computer visionon operational data such as machine signals or product images.
Continuous improvement:each release measured against the last on the same test set.
05
Choosing the level
The level is chosen first by what the answers have to do, as in our pillar on the four levels. Who keeps the data is a second question, asked at the same time: if this system’s data could feed the next one, count it.
Sometimes that moves the decision up a level. Often it doesn’t: a subscription that pays back this year beats a build that might compound in three.
06
Before you sign or build
Name the data the system will produce as people use it.
Decide who owns it, and who acts on it.
Check what the contract or the platform lets you keep and export.
Choose the level that keeps what you need, and price it.
Related use cases
ChatGPT Enterprise, Claude and CopilotThe seats you already pay for, set up to pay back.Workflow automationRepetitive document and data work, on data you already have.