
We are entering a new phase of enterprise AI. In every industry, organizations are moving past simple chat interfaces and racing to put autonomous agents to work. Yet when I meet with data and AI leaders around the world, I keep hearing the same story. Most of those agents are still stuck on the sidelines, unable to make the leap to fully autonomous operation.
Why is this? That simple question is now at the heart of the most pressing question facing AI. Without it, agentic projects cannot reach their full potential.
The reason comes down to a fundamental architectural mismatch. Legacy data platforms were built for a centralized world. Autonomous agents, by contrast, need to operate across dozens of disconnected systems, each with their own data and context. To answer a single business question, an agent has to pull facts from a cloud warehouse, check an operational database, and verify rules in a SaaS application.
So the question becomes, when your data and business logic are scattered everywhere, how do you trust what an agent does? More importantly, how do you build real enterprise intelligence without handing over control of your most valuable assets to a single vendor?
A Tale of two vendors and the myth of centralization
Today, AI is facing two distinct realities overlapping each other at the same time. Charles Dickens spoke of the best of times and the worst of time, and looking at the enterprise data market today, it’s hard not to feel that same sense of duality.
To understand how, you need to go back into the history of data and understand the role that centralization has always played until very recently.
Centralization all the way down
Go back to 2013. Back then, Teradata was pitching a Unified Data Architecture. The hot new workload was big data discovery.
Under the hood, they sold you four things. You got a place to land all your data in Hadoop, a warehouse at the center with the Integrated Data Warehouse, a box for the new workload called Aster, and analytic tools and apps on top.
How much has really changed today?
Jump forward to 2026. Today’s cloud vendors are pitching a Unified Data Intelligence Platform. The workload of today is, of course, AI agents. Yet under the hood, they are selling the exact same four things. You get a place to land your data with Lakeflow and Delta Lake, a warehouse at the center with DBSQL, a box for the new workload with Agent Bricks, and AI and BI apps on top.
Notably, it’s still all on one centralized platform, no matter how much they talk about open formats.
Centralization continues, all the way down. And it comes with the same requirements, and drawbacks. Ingest your data, transform your data, and serve your data inside their walls, and only then can you unlock the value of your business.
Why data centralization fails enterprise AI every time
The problem is that the challenge I heard from customers in 2013 is the exact same challenge I hear today.
First, it’s impossible to get all of your data into one place. For 40 years, companies have been trying and failing to create a single source of truth within a single data platform.
Second, even if you could centralize everything, would you really want to? Is a monolithic data architecture going to give you the flexibility to adapt to a changing landscape, particularly in the era of enterprise AI when technologies are changing so quickly? Furthermore, would it be wise to put all of your company’s most valuable data assets into the hands of one vendor?
Accurate answers don’t earn automatic autonomy
When I meet with customers, time and time again I hear that the real blockers aren’t around centralization at all. Centralization is a distraction. Instead, the real story is around creating defensible answers.
The engines behind leading data platforms today are all highly performant. They can deliver an answer in seconds, yet a person is still checking every single step. Autonomy has not been reached, and instead, we are running the exact same playbook as traditional BI, just with a conversational interface on top.
That’s because accuracy alone is not enough. Answers from agents need to be defensible to the stakeholders who turn those insights into decision-making intelligence. Defensible means you can show how the agent got there. Get defensibility, and people approve outcomes instead of checking steps. Climb further, and agents act within set limits, which is where the real return on AI starts to show.
So what does it take to defend an agent? Three things. Access, meaning, and proof. Did the agent follow the right access patterns? Did it properly interpret the data? And can we prove exactly what it touched and what it did? Where all three overlap, you get an answer you can defend.

Let’s look at each of these one by one, with an example in mind.
Imagine that a fraud agent receives one simple question. Is this transaction fraud? Answering that single question requires data from three separate systems. The agent needs the transaction from a payments database, the customer’s history from your supposedly unified data platform, and the open case from a ticketing system.
So where do you start. Think about access, meaning, and proof.
Access: Who can see the answer?
Data access is the first barrier in a distributed enterprise setting. When an agent crosses system boundaries to answer a question, access controls remain trapped inside individual silos, leaving nobody to govern the task as a whole.
Look at how that breaks down in our fraud example. The payments database enforces its own rules for the transaction, your data platform checks permissions for customer history, and the ticketing system validates access to the open case. Locally, every system checks its own piece, and every check is right. Yet your centralized data platform can vouch for only one hop, its own.
No system steps back to ask the bigger question of whether the analyst is actually authorized to see what all three data points add up to. When security is verified one silo at a time rather than across the entire task, a locally compliant query can still expose sensitive information and undermine trust in the result.

Meaning: Did the Agent Properly Interpret the Data?
The second barrier is meaning, which is often the hardest to solve. Across a large organization, business terminology changes from department to department. Look at how the word fraud is treated in our example. In the payments database, fraud means a predictive model scored a transaction above a specific risk threshold. In the ticketing system, fraud simply means a customer called in and filed a claim.
One is an algorithmic prediction, while the other is a customer claim, and your centralized data platform knows neither. If an agent guesses between them, it will bring the wrong answer to the business. To provide a truly defensible outcome, agents need to know exactly which definition they are using.

Proof: What did the agent read?
The final barrier to defensibility is proof. When data is scattered, disconnected systems generate disconnected logs. When an internal auditor asks what the agent read and in what order, they are left staring at a fragmented history with no single trace tying the task together.
Historically, organizations tried to solve this by copying all of their payments and ticketing data into one central platform. That approach just builds a massive second estate that you have to secure, govern, and pay for. True defensibility means proving exactly how an agent arrived at its conclusion across multiple systems, without forcing you to copy your data just to govern it.

The fight against (and need for) context gravity
Next, let’s talk about two types of gravity. Since day one at Starburst, we have fought against the first type, data gravity, that locks you in and prevents you from accessing all of your data. Today, AI is creating a new problem, context gravity. Vendors are racing to trap your business definitions inside their own tools, aiming to own your context for themselves rather than for you.
The limitations of centralized, converged data platforms
Centralized data platforms miss any data and context that live outside their boundaries. To make effective use of them, you are forced to migrate and centralize everything. At that point, you lose control over your data, trap your context, and lock your business into a single vendor.
Why catalogs and isolated AI apps fall short
What about catalogs as a source of truth? Standalone catalogs attempt to centralize context without moving your data, but they also separate definitions from the query engine. When context is divorced from underlying data execution, definitions inevitably drift out of sync, leaving no way to enforce unified access or trace answers. Meanwhile, the proliferation of individual LLM sessions and AI apps creates even more centers of gravity, trapping valuable context inside isolated, non-scalable silos.
How the Enterprise Intelligence Platform sits above every platform
There is only one architecture that solves for access, meaning, and proof without vendor lock-in, and that is an architecture that sits above every underlying platform.
Instead of forcing you to centralize data or maintain separate context silos, Starburst brokers a single, secure view across your entire estate. By delivering governed data and certified context together at query time, we give your teams complete visibility into who accessed what and the exact logic behind every answer. That unified intelligence can then power Starburst AIDA or connect to any external agent via MCP.
Own your data, context, and infrastructure
So where does this leave us, and where does it leave the future of AI? Beyond any single architectural diagram, there is a deeper principle at stake in the AI era, and that is ownership. From day one, Starburst has been built on optionality so you never have to hand control of your most valuable assets to a single vendor. Today, we are extending that promise across your entire stack so you can put agents to work on your own terms.
In practice, this means three things.
Own your meaning and answers
Context Studio harvests your existing business definitions from across your tools and delivers them as governed context alongside your data at query time, completely eliminating the need for a rebuild. That shared context directly powers Starburst AIDA, whose coordinated agents answer questions across distributed data products, build dashboards, and propose fixes that humans approve.
Own your operations
True ownership also means having clear visibility and control over how your platform runs. Mission Control gives teams an immediate view into usage, costs, and failures, paired with an expert agent that recommends what to change next. Meanwhile, LakeOps automates Iceberg data lakehouse maintenance by inspecting each Iceberg table and executing only the exact optimizations it needs.
Own your infrastructure
Finally, you should never have to sacrifice deployment control or execution speed. With Starburst Galaxy BYOC, you can keep control by running Galaxy inside your own AWS account while Starburst manages platform operations. And when workloads demand maximum speed, Torin accelerates your existing analytical SQL queries on GPUs without rewrites, delivering up to 7x faster performance in Starburst benchmarks.
Starburst Launch Point 2026
For more on this, watch my keynote and the full Launch Point 2026 keynote and product demos on demand to see how Starburst delivers access, meaning, and proof across every platform.
Watch the Launch Point 2026 keynote now

