Every larger museum has a device on the wall that nobody photographs. An unremarkable box, often close to the floor, recording temperature and humidity. Thoroughly boring, and the reason a 150-year-old oil painting still looks like an oil painting.
No museum puts its collection in order once and is then finished. It maintains a state: measured, with named responsibilities, with a line in the budget.
In companies, data quality almost always runs as a project instead. Six months of clean-up, a final report, a relieved team. Two years later it looks the way it did before.
Someone recently put it perfectly in a comment under one of my posts: treat data quality as a one-off clean-up project and you repaint the picture while ignoring the cataract. That is exactly it.
An operating state does not come from cleaning harder. It comes from four parts that have to work together. Three of them come from data governance. The fourth comes from your data and AI strategy, and without it the other three quietly fall apart again.
The target value is written on the catalogue card
Every object in a collection has a catalogue card: what it is, where it came from, what it is made of, what condition it should be in, how often that gets checked. Without that card a conservator cannot judge anything at all. They see a dark painting and cannot tell whether it was painted dark or has become dark.
That is what metadata does. Data governance is often misread as a rulebook, as something that mainly forbids. Its most useful part is a different one: it records what a data object means in business terms and how good it has to be.
Without a target value there is no quality verdict, only opinions. „Is our data good?” is therefore not too demanding a question, it is an unanswerable one. „Is this field populated the way the forecast needs it in 98 per cent of cases?” is a measurement, and somebody can repeat it tomorrow.
The same record can carry two verdicts, and both are correct. A customer address that is 95 per cent accurate is perfectly fine for the newsletter. For delivery it is a cost problem. Quality is not a property of the data – it is a promise made to a specific use case.
No name on the wall, no responsibility
In a museum it is not „conservation” that looks after a work, it is a person with a name. That is precisely what data ownership provides.
A data engineer can sort out the syntax: formats, duplicates, mandatory fields, completeness. Whether a number means the right thing in business terms is not something they can know, and that is not a criticism, it is division of labour. Whether „revenue” is gross or net, whether a cancelled order still counts, whether a customer is one entity or two after a merger: the business decides that, or nobody does. Business leads, technology enables, here as everywhere else.
And now the asterisk: data ownership rarely fails because of unwillingness. It fails because of time. An owner carrying the role on top of a full calendar is a name on a slide. Naming ownership without freeing up the capacity for it does not settle responsibility, it merely documents it.
When something breaks, only one question matters: who is affected?
Museums track provenance and loans. If something happens to a work, it takes minutes to establish which exhibition, which catalogue and which institution are affected.
Data lineage answers the same question: which report, which model, which process depends on this field? Without it, every change is a blind flight. An interface is switched over, three months later a forecast stops adding up, and nobody connects the two. That is exactly how the slow decay happens that never feels like an outage.
With AI this question becomes more urgent, for two reasons. The first: a dashboard with a wrong number is read by people who eventually get suspicious. An agent that uses the same data to set order quantities or prioritise offers does not get suspicious. It acts – immediately and at scale.
The second reason carries more weight and is readily overlooked: metadata is what gives an AI any context at all. A model cannot tell whether a column holds gross or net figures, whether a time series is calendar-adjusted, or what a status code means in practice. Unless somebody wrote it down. What passed for tedious documentation duty for years has become a functional component.
Not every painting hangs in a climate case
Museums have storage, and a great deal of material sits there under considerably less effort. That is not negligence, it is prioritisation.
The same holds for data. Anyone setting out to catalogue four thousand tables, assign owners and monitor all of them will be finished in about eighteen months, having improved not a single decision in the meantime. The way in is the data that carries a decision that moves money. The rest follows when it is needed.
The same measure applies to tooling. Automated checks measure reliably against a target value, but they do not set one. Whoever automates first and only then works out what „good enough” means ends up scaling their own blind spot – with a nicer dashboard.
What the data and AI strategy adds
Metadata, ownership and lineage are structure. They hold up what already stands, and on their own they trigger not one single check. The continuity comes from your data and AI strategy, and it contributes three things: the target picture that decides which decisions actually need to be sharp; the prioritisation that determines where quality matters first; and the place in the operating model where checking, thresholds and correction happen on a regular beat.
The difference from a clean-up project comes down to three properties. A project has an end, an operating state has a rhythm. A project measures progress in records cleaned, an operating state measures it in decisions that can be made without debate. And a project needs a budget, an operating state needs a line in the operating model.
None of that is spectacular. It is also why it gets skipped: a clean-up project produces a final report you can present. A rhythm produces a reading that somebody has to look at.
Back to the museum
Monet’s failing eyesight did not arrive overnight. It came over years, in small steps, with nobody raising an alarm, until he was operated on late in life and saw what he had been painting. Data estates degrade in exactly the same way: never with a bang, always in small steps.
His painting hangs today in a building where an unremarkable box on the wall notes temperature and humidity every hour. Not because anyone suspects a problem. So that there will not be one.
Data quality is not a state you establish once. It is one that somebody maintains.