Context Is Becoming the Most Important Data Skill

A regional retail chain rolls out an internal AI assistant. On Monday morning, the new VP of merchandising asks it a simple question: how many active customers did we add in the last quarter?

The assistant answers in two seconds. The number is wrong. Not by a rounding error — by almost forty percent.

An hour later, the analytics team is in a meeting room with the marketing team and the loyalty team. Marketing defines an active customer as anyone who opened an email. Loyalty defines them as anyone who made a purchase. The warehouse has a column called is_active that was last updated by a contractor who left in 2022.

Three teams. Three definitions. One AI. The model was not wrong. The meaning was.

In the dashboard era, humans absorbed this kind of confusion. A senior analyst would read the chart, frown a little, and apply the missing context in their head. They knew which definition was used. They knew which join dropped which rows. The context lived in people.

Agents do not have that memory. They only have what you wrote down. Whatever is missing from your semantic layer, glossary, or ontology, the model will quietly invent. Confidently.

This is the new failure mode of enterprise AI. The models are getting smarter. The meaning underneath them is not.

Of the four C's — Context, Control, Cost, Choice — Context is the one most teams under-invest in, because it does not look like engineering. It looks like meetings. It looks like glossaries. It looks like arguing about what 'customer' means for the eighth time this year.

But Context is the foundation. Without it, the other three C's cannot do their job. Control cannot govern what it cannot define. Cost cannot be optimized for queries no one understands. Choice cannot survive a migration if the meaning is locked inside one vendor's UI.

The teams pulling ahead right now treat context as a first-class engineering discipline. Owned. Versioned. Reviewed. Improved on a schedule, like any other production system.

A semantic layer is a single place where every important business term is defined in code. Active customer. Net revenue. Fulfilled order. Each one has a definition, an owner, a test, and a version history.

An ontology goes one step further. It describes how the business is shaped — what an order is, what a customer is, how they relate, what rules apply. Think of it as a map of the business that both humans and machines can read.

A context layer wraps both of these together and exposes them to your agents through a clean interface. When the agent answers a question, it pulls from this layer instead of guessing.

Pick your three most-asked business questions. For each one, list every noun in the question. Then ask three different senior people on three different teams to define each noun. If you get three different answers, you have just found the work that matters most.

Write down the agreed definition. Put it in code. Wire it into your dashboards and your AI features. Make the agreed definition the only one the system can serve. Retire the rest.

That is the entire job. It looks small. It is the highest-ROI engineering investment in your stack right now.

If you are a data engineer, your job description is quietly expanding. You still own pipelines, but you also own the semantic layer that sits on top of them. If you are an analytics engineer, this is your moment — the work you have been doing in dbt models is now the front door to every AI feature in the company. If you are a BI developer, your metric definitions are no longer just for dashboards; they are for agents too.

Architects and data leaders: this is the new design conversation. Where does business meaning live, who owns it, how is it tested, and how does it reach the model at runtime?

Most teams treat the glossary as documentation. A wiki page. A spreadsheet. Something someone in the data office updates once a quarter, if anyone remembers.

Documentation rots. Code does not. The moment your glossary lives outside the systems that use it, it is already drifting from reality.

Treat your semantic layer like a product. Give it an owner. Give it a backlog. Give it tests that fail loudly when a definition changes. Review it the same way you review schema changes — because in the agentic era, it is a schema change.

Then plug your AI features into it as the only source of business meaning. Not the prompt. Not the model card. The layer.