“LLMs thrive on context. That’s it. You know, context is the king.”
At this year’s Big Data London, Keboola CEO Pavel Doležal sat down with data expert Christina Stathopoulos on EM360Tech’s Don't Panic, It’s Just Data podcast to explore what’s next in enterprise data automation.
The conversation centered on a quiet but powerful shift in how AI, particularly large language models (LLMs), is changing the way organizations interact with their data platforms - not just by querying them, but by acting on them.
It’s a shift driven by one thing: context.
Today’s AI models can write SQL, generate dashboards, and even suggest pipelines. But in most companies, the friction isn’t in the model - it’s in the environment the model operates in.
“We've all read the MIT study which said 95% of POCs actually don't go into production.”
The failure rate of AI projects isn’t due to lack of ideas. It’s because real enterprise environments are messy. Data is siloed, tooling is fragmented, and no single system understands the full picture.
According to Doležal, most AI tooling falls short because it lacks the very thing that made BI succeed: embedded awareness of where the data is, how it flows, and why it matters.
“Now the companies have over 300 sources of data and they have 80, 90 departments which need to work with the data.”
In that kind of landscape, traditional self-service BI starts to break down. Contextual automation - the ability for AI agents to not only answer questions, but understand the structure and take the next step - becomes not a luxury, but a necessity.
That’s the space Keboola’s “Agent” now inhabits: part developer assistant, part automation driver, part chat-based interface. It's designed to operate with built-in metadata, lineage tracking, and transformation context.
But its core function is clear: to help users take action without losing control.
“By default, the chat has a context, knows what to do, knows where not to go.”
The most interesting insight from the interview isn’t technical - it’s cultural.
Doležal describes how AI tools have begun bridging the gap between engineers who maintain data stacks, and business users who consume insights. Both groups benefit from AI, but in different ways.
“What we actually do in the platform is assist the data engineers with the build of the pipeline… But then we started to apply that to the data marts where businesses actually work.”
Engineers get assistance writing and orchestrating complex logic. Business users get help interpreting tables and adjusting dashboards. Both interact with the same system - just from different ends.
The “Agent” in this scenario doesn’t just suggest answers. It operates inside a framework with context-aware tooling, lineage, and limits.
The most compelling moment in the interview comes when Doležal describes what users are asking for next:
“What people are asking for is like, ‘Hey, we ask the question, you have all the context for it. Can you press automate?’”
This is the true endgame for contextual automation: agents that don’t just answer questions, but act on the answers. Create the flow. Launch the job. Deploy the dashboard. With just enough oversight to keep things safe.
In the final moments of the podcast, Doležal sketches an idea that feels more like science fiction - for now.
“Can we actually have agents not only running internal processes, but kind of like doing business with external processes?”
Imagine agents that negotiate between systems, interact across company boundaries, and maintain secure, structured communication. The infrastructure isn’t all there yet - but the intent is forming.
The core promise of contextual automation isn’t just scale or speed. It’s relief. From the endless clicking, scripting, and chasing down pipelines.
“It’s about context. It’s about knowing the process and what you want to do. The technical skills… they just got a hundred times easier.”
Whether that’s hype or reality will depend on execution. But for teams buried in pipeline maintenance, reporting requests, and “just one more dashboard,” the direction is promising.
And if you’ve ever lost a weekend to debugging Airflow, it might sound like the future can’t come soon enough.
Watch the EM360 Podcast or head to our Keboola Agent page to see how it works and request a demo.
.png)
