Databricks CEO Ali Ghodsi says smarter models are not the bottleneck. Missing company context is.
Imagine the AI model froze today. No new versions, no bigger releases, no surprise benchmark jumps.
For most companies, Databricks CEO Ali Ghodsi says, that would not matter much.
In a recent interview he put it plainly: "The models are smart enough, but they just don't have the context that exists inside of any organization."
That one line explains a lot about why enterprise AI feels slower than the headlines promise. And it is good news for data engineers.
Many teams wait for the next model to fix their AI projects. The thinking goes: once the model is smarter, the agent will finally work.
Ghodsi's point is that intelligence is not the bottleneck. "There's no agentic work that's automated the whole enterprise," he said. "That has just not happened."
The models can already write SQL, summarize documents, and reason through steps. What they cannot do is know how your company actually works.
"If the frontier doesn't advance, it doesn't actually matter for the vast majority of organizations on the planet," he said. Most companies are still early in actually automating work and getting value from it.
Every team has one or two people who know everything. Which table is the real one. Why revenue is calculated differently in Europe. Which dashboard nobody trusts.
Ghodsi described them well: "You go tap on their shoulder, and everybody's like, 'Oh my God, what would happen if he or she quits?'"
That knowledge lives in meetings, Slack threads, old tickets, and people's heads. It never reached the public internet, so no model was trained on it.
An AI agent has "not been in every meeting," as Ghodsi says. It does not know your processes. So it guesses, and guesses look like failures.
It helps to separate two layers.
The model layer gives general skill: language, reasoning, and code. You can rent this from many providers.
The context layer gives company-specific knowledge: which data exists, what it means, who may see it, and which rules apply. Nobody can rent this to you. Your team has to build it.
flowchart LR
A[Tribal knowledge<br/>meetings, docs, people] --> B[Governed data<br/>tables, comments, metrics]
B --> C[Context layer<br/>permissions, definitions, lineage]
C --> D[AI model<br/>general reasoning]
D --> E[Useful action<br/>correct answer, safe change]A strong model with no context gives a confident wrong answer. A good enough model with rich context gives a useful one.
Our companion living book, The Context Advantage, is built on exactly this view. It frames the agentic AI era around four Cs.
Context. Give the agent the knowledge it is missing: clear table descriptions, metric definitions, and business rules it can actually read.
Control. Let the agent see only what the person asking may see. Context without permissions is a data leak waiting to happen.
Cost. A smaller model with the right context often beats a larger model with an empty prompt. It is cheaper, faster, and easier to check.
Choice. When your context lives in your own governed platform, you can switch models later without rebuilding your company's memory.
Ghodsi's quote is a short summary of the first C. The book spends its chapters on all four. You can read the first three chapters free, or browse our Context Advantage hub for companion reading.
If context is the bottleneck, the people who organize data hold the key. Here are three practical places to start.
Start with table and column comments in Unity Catalog. Short, honest sentences are enough.
COMMENT ON TABLE sales.orders IS
'One row per customer order. Cancelled orders are kept with status = CANCELLED.';
ALTER TABLE sales.orders
ALTER COLUMN net_amount COMMENT 'Order value after discounts, before tax, in USD.';This works in Databricks Free Edition. Every comment is one less shoulder tap, for a person or an agent.
If three dashboards calculate "active customer" three ways, an agent will pick one at random. Agree on one definition, store it in one governed place, and point everything to it.
Do not paste data into prompts. Let agents reach data through governed tools that respect permissions, such as the Model Context Protocol with Unity Catalog. Then every answer can be traced and checked.
For years, documentation and data quality felt like chores. In the agentic era, they become the main work.
The organizations that win will not be the ones with the newest model. They will be the ones whose data is clear, trusted, and described well enough that a machine can use it.
That is data engineering work. The skills you practice on BricksNotes, building clean tables and reliable pipelines, are exactly what turns a smart model into a useful one.
The frontier may keep moving. Your context is what makes it count.