A bank implemented Data Vault for auditability. Analysts say it is unusable and are extracting to spreadsheets. What went wrong?

What the interviewer is testing

Whether you understand that Data Vault is an integration layer, not a serving layer.

What went wrong

The marts were never built. Data Vault's structure — hubs for business keys, links for relationships, satellites for time-stamped attributes per source — is designed for absorbing change and demonstrating lineage. It is deliberately unpleasant to query: many more tables than a dimensional model, and a simple business question requires many joins with effective-date filtering.

That is the intended trade. The vault integrates; dimensional marts on top serve. Exposing the vault directly to analysts skipped the layer that makes it usable.

The consequence is predictable and is what happened: analysts route around it, extract to spreadsheets, and the organisation now has ungoverned copies — the exact opposite of what the programme was funded to achieve. And every spreadsheet has its own join logic, so the numbers diverge.

The remediation

Build the marts. Dimensional models over the vault, conformed dimensions, at the grain analysts actually work at. This is the missing deliverable, not an enhancement.

Start with the questions analysts are answering in spreadsheets, which is a precise specification of what the marts must contain and gives immediate adoption.

Add a semantic layer over the marts so metric definitions are single-sourced.

Then decommission the extracts — but only after the replacement is genuinely easier, or they will persist.

What a strong answer adds

Defending the vault itself: for a bank with many sources and a regulatory requirement to demonstrate what arrived and when, it is a sound choice and the auditability is real. The failure is in the delivery plan, not the modelling approach, and saying so protects a large prior investment from being discarded for the wrong reason.

Common weak answers

Concluding Data Vault was the wrong choice. Training analysts to write the joins, which does not scale and does not fix the underlying gap.