Plata's One-Hour Warehouse and the End of Tableau

Any platform lead who has run analytics for a retail bank knows the daily rhythm: the warehouse refreshes overnight, fraud spikes get spotted at 10am, and the ops team spends the afternoon cleaning up damage done at 3am. Plata, Mexico's fastest-growing fintech bank, decided that lag was unacceptable. The data warehouse now refreshes within one hour and underpins 80 to 90 percent of every data process in the company.

That single design choice, made three and a half years ago when the team was eight people, is why the department is now over fifty. It's also why Plata thinks it can walk away from Tableau entirely.

The Numbers

Start with the growth curve. The data warehouse team went from 8 people to more than 50, organised across roughly 10 specialist teams, in three and a half years. That's a 6x headcount expansion in a discipline where most fintechs are still arguing about whether analytics engineering is a real job title. As FinTech Magazine reported, Head of Data Warehouse Ivan Shchukin built the function from zero after being headhunted by CTO Andrey Shelekhin, who told him bluntly: "we need to create the best product."

The operational numbers matter more than the headcount. One-hour refresh latency, versus the 24-hour lag that still defines most bank warehouses, changes what analytics can be used for. At 24 hours, you're doing forensics. At one hour, you're doing operations. A spike in failed card authorisations at 11am becomes an incident you fix by lunch, not a Jira ticket the analytics team files tomorrow.

The 80 to 90 percent figure is the one I'd underline for any CTO reading this. When a single warehouse underpins that share of your data processes, with automated quality checks and alerts on every pipeline, you've centralised your source of truth in a way most banks can only draw on a whiteboard. Production incidents I've seen at fintechs of similar size almost always trace back to two or three parallel data stores drifting out of sync. Plata has designed that failure mode out.

Then there's the full-versioning decision. Every change to every record is logged and preserved. Shchukin is candid that "most players in the market are afraid of it." He's right to flag that. Full versioning multiplies storage costs, complicates GDPR-style deletion requests, and makes schema evolution harder. Doing it anyway, from day one, at a bank, is either brave or reckless. The 80-90 percent process coverage number suggests it's paid off.

Fifty people is not cheap. On a lean fintech, that's a material line item, easily two engineering squads worth of budget. The counterweight is that Plata plans to seed each new Latin American market with just two to four analytics engineers, riding on the tooling the central team has already built. That's the whole thesis: invest heavily in the platform, then scale geographies with tiny cells.

What's Actually New

The one-hour warehouse isn't new in principle. Streaming CDC into Snowflake or ClickHouse has been technically possible for years. What's new is a Mexican retail bank actually operating on it as the default, not the exception, and building 80-90 percent of its data flows around that assumption.

The genuinely new bet is the plan to phase out Tableau and Apache Superset entirely. Not augment them. Not integrate an LLM on top. Retire them. Shchukin's framing: "within the next year, every analyst will write their own small BI tools using AI agents." The replacement is AI-generated Streamlit apps, bespoke Python visualisations produced on demand, not dashboards curated by a central team.

That's a genuine departure. The dashboard-as-artefact model has been the industry standard for fifteen years. Every enterprise data team has a graveyard of stale Tableau workbooks nobody remembers commissioning. Replacing them with disposable, generated Streamlit apps flips the ownership model. The analyst is no longer a dashboard consumer requesting changes from a BI team. They're the author, with an LLM as pair programmer.

The other new element is the internal IDE. Plata's staff currently touch about 10 different tools for data work: Airflow for scheduling, Snowflake worksheets for SQL, Tableau and Superset for dashboards, and the rest. Shchukin wants "just one window for any task." Consolidating those into a single AI-assisted environment is on the year's roadmap.

My take: the IDE consolidation is the harder problem and the more valuable one. Anyone can prototype a Streamlit app with Claude. Building an integrated environment where scheduling, querying, modelling, visualisation, and lineage live in one pane, with an agent that understands your specific schema, is a multi-year platform investment. Plata is signalling they'll build it themselves rather than wait for a vendor. Given how badly the current BI vendors have handled the LLM transition, that might be the right call.

What's Priced In for Data Teams

Most senior data leaders already expect analytics engineering to be the default hiring title for anyone sitting between raw pipelines and business users. Tools like dbt made that role legible. Plata claims to have been "among the first" to hire with the analytics engineer title, and that timing looks right. What's priced in: the discipline itself. What isn't priced in: staffing an entire fifty-person department around it and treating data engineering proper as a supporting function.

The one-hour refresh cadence is also increasingly expected among fintechs that take fraud and risk seriously. Payments teams already run streaming stacks for authorisation flows. What surprises here is the scope: not one hot path, but 80 to 90 percent of company data processes on the same near-real-time substrate.

The bit that will catch data teams off guard is the honest limitation Shchukin admits about LLMs on top of warehouses. He rejects the vendor pitch that you can "integrate an LLM with any database and it becomes a simple computer you can question about anything." His counter: "you need to describe your data first, and that's where we are working now." That's the semantic layer problem, and it's the actual work nobody wants to fund. Every team betting on text-to-SQL as a shortcut is going to hit this wall.

The uncomfortable read: most fintechs have spent 2025 buying AI features from their BI vendor and calling it a strategy. Plata is doing the unglamorous work of documenting its schema well enough that an LLM can reason about it. That's the moat.

Contrarian View

The contrarian read is that killing Tableau is a mistake, or at least premature. Dashboards exist for a reason. They enforce a shared view of the numbers. When every analyst is generating their own bespoke Streamlit app on demand, you get the analytics equivalent of shadow IT: forty slightly different definitions of "active customer," each blessed by an LLM that hallucinated a JOIN.

Governance is the counter-argument. A curated Superset dashboard has an owner, a change log, and, ideally, a certified data model behind it. An AI-generated Streamlit app has none of that unless Plata builds serious guardrails into its internal IDE. The one-hour refresh doesn't help you if two analysts draw different conclusions from the same data because their generated queries disagreed on filter logic.

There's also a hiring risk. Plata's plan to seed new LATAM markets with two to four analytics engineers assumes the platform is so good it carries the load. If the platform slips, those tiny cells become bottlenecks and the central team ends up firefighting across five countries. The lean-cell model works beautifully until it doesn't.

Key Takeaways

• One-hour latency is now table stakes for serious fintechs. If your warehouse still refreshes on a 24-hour cycle, you're doing forensics while competitors do operations.
• Full versioning is expensive and worth it. Plata built it from day one and now runs 80-90 percent of data processes on that foundation. Retrofitting later is brutal.
• The BI vendors are on notice. If AI-generated Streamlit apps replace curated Tableau workbooks at scale, the license revenue model behind traditional BI collapses within a few years.
• Analytics engineering is now the correct hiring title. Plata seeds new markets with cells of two to four. If your org still hires "data analysts" who can't model or ship, you're behind.
• The semantic layer is the real AI moat. Text-to-SQL doesn't work without rigorous data description. That's boring, expensive documentation work, and it's what separates the demo from production.