Building the data foundation an AI roadmap needs
Before the model: the warehouse, the contracts, the lineage and the access rules that decide whether AI is possible at all.
Most stalled AI programmes are not stalled on modelling. They are stalled because the data needed to train, ground or evaluate anything is scattered, undocumented or legally unusable.
Inventory before architecture
Unglamorous, and it routinely reshapes a roadmap in the first fortnight.
- 1
What exists
Every dataset, where it lives, how fresh it is.
- 2
Who owns it
A named person who can grant access and answer questions.
- 3
What it may be used for
The legal answer, not the technical one.
- 4
Where the gaps are
Write them down. This is the cheapest time to find them.
Inventory before architecture
Before choosing tools, write down what data exists, who owns it, where it lives, how fresh it is and what it may legally be used for. This is unglamorous and it routinely kills or reshapes roadmaps in the first fortnight — which is the cheapest time for that to happen.
One place things are true
AI multiplies the cost of inconsistency. If three systems hold a slightly different customer record, a model will surface those contradictions confidently and at scale. Agreeing a source of truth per entity is prerequisite work, not cleanup.
Contracts at the boundaries
An upstream team renaming a column should not silently degrade a model in production. Explicit schema contracts, validated in the pipeline and alerting on breach, turn a class of silent failure into a loud one.
Lineage, because you will be asked
When a downstream number looks wrong — and it will — you need to trace it back through every transformation. Retrofitting lineage after the fact is painful; capturing it as you build is close to free.
- Where each field originated and what transformed it
- When it last updated, and whether that is within expectation
- Which models and dashboards consume it
- What access class it carries, and whether that survives the join
What has to exist under a model
Build the slice your first use case needs — foundations should be laid under something.
The model
The part everyone budgets for
Access control
Permissions that survive the join
Lineage
Where each field came from and what transformed it
Contracts
Schemas validated in the pipeline, alerting on breach
Source of truth
One agreed record per entity
Inventory
What exists, who owns it, what it may legally be used for
Don't build the whole warehouse first
The failure mode at the other extreme is a two-year platform programme that delivers no product. Build the slice your first use case needs, ship it, and let the next case pull the next slice. Foundations should be laid under something.