Gartner predicts that 60% of AI projects lacking AI-ready data will be abandoned through 2026. Separately, Informatica's CDO survey found that only about 12% of organizations say their data is actually sufficient in quality and accessibility to support AI applications. Put those two next to each other and you get an uncomfortable picture: most companies chasing AI right now are building it on top of data they don't actually trust.
I've watched this pattern play out from the inside at more than one company. Leadership greenlights an AI initiative, a vendor gets selected, a pilot gets built — and it works, in the demo, on the sample dataset someone hand-picked. Then it hits production data and starts hallucinating numbers, missing records, or confidently reporting something that was wrong three systems upstream. Nobody blames the data. Everybody blames the model.
Here's the thing: the model was never the problem. The foundation was.
Three terms people use interchangeably that aren't the same thing
Data quality is whether the data itself is correct — accurate, complete, consistent, not duplicated, not stale. This is the "is this number actually true" layer.
Data observability is whether you'd know if it stopped being true. Freshness monitoring, volume anomaly detection, schema drift alerts, distribution checks. Quality is a snapshot; observability is the alarm system that tells you the snapshot changed.
Data catalog is whether people (and systems) can find what exists, understand what it means, and know who owns it. Metadata, lineage, business glossary, access controls.
All three matter. But they don't matter in the same order, and most organizations get the sequence backwards.
Why catalog is usually the wrong first move
Catalogs are the easiest of the three to justify in a budget meeting. They're visible. You can demo a slick UI, show a search bar, point to lineage diagrams, and call it "data governance progress." I understand the appeal — I've sat in those budget meetings.
But a catalog doesn't make your data more correct. It makes it easier to find data you still can't trust. If your customer table has 8% duplicate records and nobody's watching for when a pipeline silently breaks, a beautiful catalog entry just means more people can quickly discover the same bad data. You've made the problem more discoverable, not less real.
AI systems make this worse, not better. A human analyst browsing a catalog might squint at a number and think "that looks off." An agent or model doesn't squint. It ingests whatever's there and reports it with full confidence. The audience for a catalog is mostly human judgment; the audience for quality and observability work is the machine that has none.
The order that actually works
- Observability first, on your critical paths only. Don't try to instrument your entire warehouse on day one. Pick the 3-5 data sources feeding your top AI use case and put freshness, volume, and schema-drift monitoring on those specifically. You want to know within hours, not months, when something upstream breaks.
- Quality rules on the same critical paths. Define what "correct" means for those specific tables — completeness thresholds, validity rules, dedup logic — and enforce it before data reaches anything AI-facing. This is unglamorous work. It's also the work that actually determines whether your AI output is trustworthy.
- Catalog, once there's something worth cataloging. Once your critical data is monitored and held to a quality bar, a catalog becomes genuinely useful — it's now helping people find and understand data that's actually reliable, rather than dressing up a mess.
- Then the AI initiative. Not before.
This isn't a purity argument. You don't need to reach 100% quality and observability coverage across the whole organization before touching AI — that's its own kind of paralysis, and it's how "we're still doing data governance" becomes the answer to "where's the AI project" for two years running. The goal is narrow and specific: get the handful of data sources your AI use case actually depends on into a trustworthy, monitored state first. Everything else can catch up in parallel.
Where to start this week
If you're a data or business leader with an AI initiative already underway, don't start by asking "do we have a catalog." Ask: for the specific data feeding this AI use case, would we know today if it broke, and do we actually know it's correct? If the honest answer is no to either, that's the project — before the next model evaluation, before the next vendor pitch. The AI headline is more exciting than a data quality audit. The data quality audit is what makes the AI headline true.