Vibe Analytics for Exploratory Data Analysis

Most of our analysis work starts the same way. Someone on an operator’s team asks a question nobody has a chart for yet, and the honest answer used to be that we would take it back to a developer and come back in a week or two. By then the question has usually moved on.

Closing that gap is what we went to Atlanta to talk about. On day one of COLLIDE 2025, the Data + AI conference hosted by Data Science Connect, we presented “Vibe Analytics for Exploratory Data Analysis”: what the language model is actually doing in that loop, and where the approach stops being useful.

What We Mean by “Vibe Analytics”

Vibe analytics is an AI-driven, code-based approach that lets someone interact with data through natural language conversation, closer to brainstorming with a colleague than to filing a report request.

The distinction that matters, and the one we spent the most time on: the model is not doing the analysis. We ask it to write the code that visualizes data we have already curated and quality checked. The numbers come out of our data lake. The model builds the view.

In our experience, calculating numbers straight from an LLM does not work well and is open to hallucination. Asking it to generate chart code against a governed dataset is a different job, and it holds up.

Why the Existing Options Fall Short

We framed the case around three limits, drawing on work by Michael Schrage of MIT Sloan and on Plotly’s own writing on the subject:

• BI tools, even with AI bolted on, are a vendor-controlled walled garden. Interoperability is limited, and you build inside someone else’s boundaries.
• Hand-coding is still the gold standard for data apps. Maximum flexibility, but it requires skills and time, and that cost is what kills experimentation. Nobody prototypes six ideas when each one is a two-day build.
• Chatbots are limited by the imagination of the prompter. You only get back what you knew to ask for.

Plotly covered the same ground in What Vibe Coding Gets Wrong (and Why Vibe Analytics is Better), which is worth reading alongside this.

How This Works With Clients

Almost every time we have run this process, it has been with a customer, partner, or prospect in the room. It takes three forms:

• Collaborative analysis. On a complex dataset like utility pole attachments, we can pull up a visual and immediately ask for another angle: most common equipment, average owners per pole, a map of attachment violations.
• Data app mock-up. We mock up a dashboard with the client quickly. They bring ideas, we bring ideas, and it comes together in Plotly Studio. Because the output is code, it can be edited and deployed. It is not a screenshot.
• Analytics portal. A self-service sandbox where clients combine our data with their own.

As I put it in Plotly’s recap of the conference:

Instead of spending weeks talking about a problem or days trying to code something to show it to someone, we narrow that gap where they could be in the same room talking in real time.

None of this works without timely data underneath it. The exploratory work sits on top of our platform: a nightly ETL that pulls from an operator’s systems into our data lake, and a serving layer, running on Plotly Dash Enterprise, that keeps a geospatial and tabular cache four hours fresh. It is fast because the hard part is already done.

What It Has Moved

The operating results we presented, measured across our client work:

• 50% faster installations, from improved scheduling efficiency with real-time data.
• 10+ hours per month saved, through automated data analysis and faster reporting.
• +2% subscriber growth, by identifying high-potential regions with targeted analytics.
• 5% lower operational costs, largely from consolidating data and retiring duplicate software licenses.

Why This Matters for Smaller Operators

The market numbers we put on screen in September 2025 framed the problem:

• 7.2 million U.S. households unserved or underserved by high-speed broadband providers.
• 26% of U.S. homes with access to only one broadband provider.
• $167 billion of capital expenditure projected for fiber to the home over the following five years.

We serve smaller and rural operators, the ones without the data science headcount a national carrier takes for granted. That is why speed matters here. When a question takes weeks to answer, it does not get asked. Our Make Ready Assessment dashboard started as exactly this kind of exploratory work.

Watch the Talk

The full session is on YouTube: Vibe Analytics for EDA, by Patrick Hutto, Broadband Insights at DSC Collide 2025.

Plotly Studio is free at plotly.com/studio if you want to try it yourself. If you would rather see it run against your own network data, reach out to us!