The Shift I’m Seeing

When I joined Dremio in 2023, our data and analytics approach was what most have done for 30+ years. An ask from an executive would eventually get to a data analyst or FP&A leader, who would then look at Tableau or PowerBI to see if they have dashboards available to answer my business question. If the dashboard didn’t exist, an analyst would spend a few days creating the right dashboard. If the dataset didn’t exist in the right form, now the request went to a data engineer and even an application engineer. Thus, a simple request - even at a company of our size - could move from a few hours to a few days.

When I speak with business leaders and data team leaders at large organizations, the above process occurs at a much larger scale with additional approvals and sequencing. But, time scales expand exponentially - from a few weeks to months. If you are an executive, I’m certain your questions are not relevant a few weeks later.

Self service has always been the ethos of Dremio, but it was focused on self-service for the analyst. While that continues today, self service has clearly expanded to anyone who can wield an AI client - be it Dremio’s own agent, Anthropic’s Claude, OpenAI’s ChatGPT or Gemini.

Today, this pattern has changed for me at Dremio. I no longer go to anyone at the company for data, and rather I leverage Dremio’s own embedded agent or Claude to query our internal data lakehouse and not only search cross functional data sets, but build and answer questions in minutes, not days or weeks. With the most recent advances in these agents, now we can visualize at a level not found with traditional BI tools.

The Problem with “Just Chatting with Your Data”

The notion of chatting with your data is a common trend in the market today. Every lake house platform, data platform, application, or BI tool in the market is claiming some version of it. While all provide great demos, I can tell you from personal experience that the demos will fail when querying your data. Not because the tools are horrible - but rather because they are based on incorrect, incomplete, and delayed context. This context is often provided by the industry layer called the “semantic layer”. In your modern data architecture, this is an important area of focus going forward.

If your agent queries raw data without a semantic layer, without context defined as consistent definitions, without information on what your metrics actually mean, you have not accelerated analytics. You have changed the interface and provided a more user friendly model, but the results will not be trusted. The lack of trust means you fall back to your teams as mentioned previously. The underlying bottleneck, which is the correctness and accessibility of the data itself, remains exactly where it was.

The real capability I have built is the ability to reach across a relational database and a separate operational data source simultaneously, resolve the query through a semantic layer that understands my business context, and receive a result without opening a data engineering ticket or sending a Slack to a data analyst. That is the difference between a new UI and a genuine productivity shift that I’ve been able to experience.

What Your Peers Are Doing

According to Gartner’s 2025 Agentic Analytics research, 65% of data leaders have prioritized agent-based analytics capabilities for 2026. And yet 70% of those same leaders identify siloed data and weak governance as their primary barriers to realizing those capabilities.

Organizations that are moving fastest are not those that purchased the newest AI tool. They are the ones whose data infrastructure was already structured to support consistent, governed, federated access before the agent was introduced.

Leveraging AI You Can Trust

When we first started leveraging AI at Dremio, it was a fast uptake, but we started to get incorrect answers from our data. The agent would hallucinate and provide poor responses. It would also exhaust massive tokens doing re-discovery of the data sets, systems, and information. In May we moved into early preview of our “AI Semantic Layer” - not just a generic manual semantic layer, but one auto-built from your own team’s query activity. Namely, the system captured the queries being run by existing dashboards and ad hoc queries, found high frequency high confidence relationships and automatically built a semantic layer. This layer can be done fully autonomously, or more commonly with a human in the loop to approve the implicit relationships Dremio was able to find.

The net result? A 50% reduction in token usage, a 70% improvement in answer quality. This is the power of a semantic layer on AI experiences. This was the only way I could leave dashboards as it allowed me to actually get AI results that I could trust. In addition, I would ask the agent to prove to me its analysis was correct - which resulted in the agent providing evidence from the semantic layer.

Your Perspective: A Quick Poll

If this describes a shift you are ready to make, I would encourage you to see what it looks like in practice. You can book an executive briefing at dremio.com, or reply directly to this message if you want to discuss what the right architecture looks like for your organization.

Sendur Sellakumar

CEO, Dremio