"Our data is a mess" is the single most common thing we hear in an intake, and it is almost never useful information. Everyone's data is a mess. The question that matters is which kind of mess, because some kinds cost you a week and some kinds mean the system you want cannot be built yet at any budget.
So we stopped asking whether data is ready and started scoring it, one to five, per system in scope. The score goes in the CRFT Scan report next to each opportunity, and it does more work than any other number in there — it is usually what decides the order of the build.
The five levels
Level four is the line
Everything below four can be made useful with AI. Only four and above can compound.
That is the whole reason we score it. A team at level three asking for a system that "gets smarter over time" is not asking for something impossible — they are asking for something that requires one more piece of work first, and it is much cheaper to say so in week one than to discover it in month five.
The move that unlocks the most
When a Scan comes back with a high-value opportunity sitting at level two or three, the recommendation is frequently not the opportunity itself. It is a smaller, duller build that moves the readiness level — structuring one data source, or emitting one outcome event that nobody was recording.
That is a hard thing to sell and we say it anyway, because the alternative is worse. An agent built on level-two data does not fail loudly. It produces plausible output on bad inputs, quietly, for months, and by the time anyone notices, trust in the whole programme is gone.
Sometimes the honest first build is the one that makes the second build possible. We would rather say that in week one than defend a model that has been guessing since March.
Reading your own score
You can approximate this without us. Pick the workflow you most want to put a system inside and ask, in order:
- Is there a system of record, or is it in people's heads? (1 vs 2+)
- Could I count the cases, or only read them? (2 vs 3+)
- Can I tell, from data alone, how each case turned out? (3 vs 4+)
- Does anything currently read those outcomes and change behaviour? (4 vs 5)
The first question you answer "no" to is your level, and the gap between that and four is your real project — whether or not anyone has scoped it as one.
