Poor data quality costs the average organisation around $12.9 million a year according to Gartner's cross-industry research. That figure predates the AI investment surge, and AI does not tolerate what reporting tolerates.
The connection is now visible in project outcomes. Gartner forecasts that organisations will abandon 60% of AI projects through 2026 because of inadequate data, and only 37% of organisations report confidence in their data practices. As AI spend accelerates past $2 trillion, the cost of bad data scales with it.
Why governance documents do not work
The standard organisational response to data quality problems is a policy. A document is written, circulated, approved, and filed. Meanwhile the actual behaviour of the data continues to be determined entirely by what the pipelines do.
If a data quality rule is not an executable test, it is a preference that someone wrote down.
Treating governance as an engineering concern changes the outcome. Quality rules become tests that run on every load. Ownership becomes metadata attached to the dataset. Access control is enforced in the warehouse rather than requested by email.
Start where the damage is
A common failure is attempting comprehensive coverage — profiling every table, testing every column. These programmes stall, because the work is enormous and the early returns are invisible.
Prioritise by consequence instead. Which datasets feed decisions that would be expensive to get wrong? Those get thorough testing first. A dimension table nobody has queried in a year can wait.
This usually means starting with a small number of tables — the ones behind revenue reporting, regulatory submissions, and any model in production.
Ownership needs a name
Datasets owned by a team are owned by nobody. When quality alerts route to a shared inbox, they get triaged by whoever has time, which is a category of person that does not reliably exist.
Assign a named steward per critical dataset. The steward is usually not an engineer — it should be whoever understands what the data means in business terms. Defining correct is their job. Automating the enforcement is ours.
Lineage turns audits from weeks into hours
Two questions arrive during audits, and both are painful without lineage: who accessed this data, and how was this figure derived?
With query logging and column-level lineage, both are lookups. Without them, both are reconstruction projects involving several people and a lot of guessing. We have seen organisations move audit response from weeks to hours purely by having the lineage graph already captured.
The regulatory driver makes this urgent in some sectors, but the everyday benefit is broader — the same lineage that satisfies an auditor is what lets an analyst answer where did this number come from without opening a ticket.
Classify sensitive data before you scale
PII scattered untagged across a warehouse becomes progressively harder to control as the warehouse grows. Classification is far cheaper at fifty tables than at five hundred.
Once fields are tagged, masking and retention policies can be enforced automatically rather than depending on every analyst remembering which columns are sensitive.
What changes when it works
The visible outcome is not a dashboard. It is that people stop hedging. Nobody prefaces a number with assuming the data is right. Investigations start from the assumption that the figure is correct and something interesting happened, rather than from suspicion of the pipeline.
Trust is the actual deliverable. Compliance is a by-product.