When two departments report different figures for the same measure, the instinct is to look at the reporting tool. It is almost never the reporting tool.
What has usually happened is that the two figures are answers to slightly different questions. One counts records created; the other counts records closed. One excludes cancellations; the other does not. One draws from a system updated nightly, the other from a system updated on Fridays. Both numbers are correct. Neither is comparable. And because the definitions live in two people’s heads rather than anywhere in the platform, nobody can prove which is which without a meeting.
Dashboards are the last layer, not the first
Analytics work is often approached from the visible end. Someone specifies the dashboard, a tool is procured, and the difficulty is discovered afterwards: the data must be assembled from four systems that disagree about what a customer is, one of which exports on request.
Working from the other direction is less immediately satisfying and considerably more durable. Establish what the organization needs to know. Determine where that information genuinely originates. Fix the definitions and the ownership. Build pipelines that are observable and recoverable. Then build the dashboard — which at that point is comparatively easy, and stays correct.
What a usable data foundation actually requires
Definitions written down
For every measure that matters: what it counts, what it excludes, which system it comes from, and how current it is. This is unglamorous and it resolves most reporting disputes before they occur.
Ownership that is named
Someone has to be responsible for each significant data set — able to decide what a field means and answer when it looks wrong. Governance without named owners is a document, not a practice.
Pipelines that fail loudly
A pipeline that stops without telling anyone is worse than no pipeline: the reports keep rendering, quietly out of date, and people keep trusting them. Observability and clear failure handling are not sophistication, they are basic hygiene.
Access designed rather than inherited
Opening data up and protecting it are usually posed as opposites. With role or attribute-based access, row and column-level controls, masking for sensitive fields and auditable access logging, more people can safely use data — not fewer. What blocks access in most organizations is not policy but the inability to enforce it precisely.
Quality treated as continuous
Data quality is not a cleanup project that concludes. Profiling, validation and reconciliation belong in the pipeline, where problems are caught as they arrive rather than discovered later by whoever happened to notice.
Why this now matters more
Weak data foundations used to produce arguments in meetings. They now produce something worse, because AI systems inherit every inconsistency underneath them and express it with complete confidence. A dashboard with a questionable number invites scrutiny. A model trained on inconsistent data produces a fluent, plausible answer that nobody thinks to question.
The organizations best positioned for artificial intelligence are, with some regularity, the ones that spent the preceding years on integration, definitions, quality and governance. That work was never the exciting part. It turned out to be the part that mattered.