Technology, People and AI

Telmo Silva sits down with Dan Stelljes, a 30-year veteran of data management in healthcare, for a grounded conversation about what it really takes to make data projects succeed. From the hidden human layer inside structured data to the trust problem at the heart of AI adoption, this episode is a reality check on where technology ends and judgment begins.

Questions We Asked

• Can AI-generated visualizations replace the analyst, or do we still need a human in the loop?
• How do you build trust in unstructured data when decision-makers have spent decades relying on hard numbers?
• Is the speed at which organizations are adopting AI outpacing their ability to use it responsibly?

What We Learned

• Structured data only tells part of the story. The operational, human context hidden in unstructured sources reveals why problems happen, not just that they happened.
• Triangulating AI responses across multiple questions and models is the most practical way to validate results today.
• AI is one tool in the workflow, not the entire workflow. It works best when embedded alongside other data tools and human expertise.
• The experience gap is real. As junior analysts rely on AI to do the hands-on work, the next generation may lack the domain instincts needed to know when an answer is simply wrong.

From Healthcare Claims to a Broader Truth About Data

Daniel Stelljes has spent 30 years wrangling some of the messiest data in any industry: healthcare revenue cycle. In that world, structured data can tell you there was an underpayment. What it cannot tell you is why. That requires a third layer: the operational, human context captured in unstructured sources like PDFs, reports, and workflow logs. Daniel’s approach is to read all three together, treating them as complementary signals rather than competing ones. When structured and unstructured data tell the same story, confidence increases. When they diverge, that gap is often where the real problem lives.

The Trust Problem: Why Hard Numbers Still Win

Despite the promise of AI-generated insights, many stakeholders still anchor their trust in hard numbers. Asking them to rely on softer, inferred signals without visible evidence is a bit like the early days of ATM cards: people would withdraw money at the machine, then walk into the bank to verify the same transaction. The challenge for data teams today is not just surfacing the right answer but making the reasoning behind it visible enough to be believed.

I use AI like an intern. Go off and do this set of tasks for me and then I’m going to judge whether that seems to be accurate or not.
Dan Stelljes, Principal, Beastbook Solutions

That healthy skepticism, Dan argues, is not a weakness. It is the appropriate posture for anyone working with systems that can be confidently wrong.

Triangulating AI: A Practical Method for Validating Outputs

Telmo and Daniel share a common frustration: AI models are inconsistent. Ask the same question twice, across different days or different model versions, and you may get different answers. Daniel’s response is to triangulate: break a problem into its component parts, run each separately, then compare the outputs for internal consistency. If results converge, that is a reasonable signal of reliability. If they diverge, that is the prompt to dig deeper. It is, as Telmo puts it, doing statistics on statistically generated answers.

Visualization Is Not Going Away

The idea of replacing dashboards with a chat interface is appealing in theory. In practice, it falls short the moment real exploration begins. AI can generate a chart quickly, but verifying that it handled null values correctly, applied the right filters, and aggregated data as intended still requires a trained eye. Interactive visualization, the kind that lets analysts drill down, zoom in, and pivot on the fly, remains faster and more reliable for genuine analytical work than iterative prompting. As Telmo notes from his own experience building at ClicData, prompting works well for summaries and narrative questions. For deep visual analysis, it adds friction rather than removing it.

AI Is One Tool. Not the Only One.

Organizations are moving fast on AI, sometimes faster than their understanding of what it actually does. Daniel draws a clear distinction between using AI as an accelerator for known, bounded tasks versus expecting it to handle complex, multi-variable analysis end to end. The latter requires the same rigor as any other data project: scoped requirements, tested outputs, and human review at every stage. AI has its place in the ETL stream. It does not replace the stream.

The Experience Gap No One Is Talking About

Perhaps the most thought-provoking moment of the episode arrives near the end. As senior developers shift into validator roles and junior analysts rely on AI to produce the work, a quiet question emerges: where does the next generation of experienced practitioners come from? Experience is earned by doing, by making mistakes, by developing the instinct to recognize when an answer does not make sense. If AI absorbs that process entirely, the field may eventually face a gap between the tools it uses and the judgment required to use them well.