I keep hearing the same quiet panic from people who have spent a decade getting good at analytics.
The company wants an AI platform. The job reqs say "AI engineer." Someone in leadership asks whether we still need a BI team. The analyst who can explain why revenue moved last Tuesday starts to wonder if that skill still counts.
It does. It is the whole job.
Traditional analytics is not the old world. It is the training that AI platforms actually run on. Metrics. Modeling. Pipelines. Quality. Stakeholder consulting. Storytelling. If you can do that work, you are not behind. You are standing on the only part of this stack that is currently failing.
The failure is not the model
Gartner predicts that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data. Precisely and Drexel University's LeBow College of Business found that only 12% of organizations report their data is of sufficient quality and accessibility for AI.
Most companies chasing AI are building on data they do not trust.
I have watched this from the inside more than once. Leadership greenlights an AI initiative. A vendor gets picked. A pilot looks great on a sample dataset someone cleaned by hand. Then it hits production data and starts missing records, or reporting something that was already wrong three systems upstream.
Nobody blames the data. Everybody blames the model.
The model was never the problem. The foundation was.
A catalog is not a strategy
When the foundation is weak, companies reach for the thing that is easiest to demo. That is usually a data catalog.
I understand why. You can show a search bar. You can point at a lineage diagram. You can call it governance progress in a budget meeting.
A catalog does not make data more correct. It makes untrustworthy data easier to find. If the customer table has duplicates and nobody is watching when a pipeline silently breaks, a nice catalog entry just means more people can discover the same bad data faster.
AI makes that worse. A human analyst might squint at a number and think it looks off. A model does not squint. It ingests what you give it and reports it with full confidence.
The order that works
- Observability first, on the critical paths only. Pick the few sources that feed the use case. Watch freshness, volume, and schema drift. You want to know in hours, not months, when something upstream breaks.
- Quality rules on those same paths. Define what "correct" means for those tables. Completeness. Validity. Dedup. Enforce it before the data reaches anything AI-facing. This is unglamorous. It is also the work that decides whether the output is trustworthy.
- Catalog, once there is something worth cataloging. Now the catalog helps people find data that is actually reliable.
- Then the AI initiative. Not before.
You do not need 100% coverage across the company before anyone touches a model. That is its own kind of stall. Get the handful of sources the use case depends on into a monitored, trustworthy state. Everything else can catch up in parallel.
A model is not a strategy
Buying a model does not make an organization smarter. Without data quality, context, and training, it is a faster search engine.
True AI work needs governance, observability, and quality. That does not take forever. It does take commitment, and it takes someone who knows how to do it.
That someone is usually the person who has been doing traditional analytics.
They already know how a metric gets defined. They already know which stakeholder will fight the definition. They already know which pipeline breaks on month-end. They already know how to tell a true story with a number, and how to stop a false one.
Do not replace those people with AI. They are more qualified to make AI actually work than a model sitting on top of a mess.
Letting your best analytics people go in favor of an AI hire can cost you months of production and quality. You are firing the people who understand the context, then asking a system with none to take their place.
What "stay relevant" actually means
If you are an experienced data person, staying relevant in this push is not learning a new trick for the resume. It is putting the skills you already have on the path the AI work depends on.
You already know how to:
- Define a metric so two teams stop arguing about it
- Model a process so the number means one thing
- Build a pipeline that can be trusted on Monday morning
- Catch bad data before a dashboard ships it
- Sit with a stakeholder and find out what they actually need
- Tell the story without dressing it up
That is observability, quality, semantics, and consulting. That is the on-ramp to an AI platform. The people who skip it ship demos. The people who do it ship systems someone will still use in a year.
Start with one path
If you have an AI initiative already underway, do not start by asking whether you have a catalog. Ask two questions about the specific data feeding that use case.
Would we know today if it broke?
Do we actually know it is correct?
If the honest answer is no to either, that is the project. Before the next model evaluation. Before the next vendor pitch.
The AI headline is more exciting than a quality audit. The quality audit is what makes the headline true.
If you are in the middle of this, hit me up. I would rather talk through the actual mess than add another slide to the pile.