
LLMs for Analytics is more than ‘text2sql’

Category: Trends
Want to try an analytics-as-code solution? ⭐ Check out Preswald on GitHub to get started in under 4.7 mins.
In 2022, the analytics world was dazzled by “chat-with-your-data” tools. These applications promised magic: type a question in plain English, and AI would write SQL for you. Typing natural language questions to retrieve SQL-generated insights felt like a magic trick, and it was. But now it’s 2025, and it’s increasingly obvious that SQL generation is just scratching the surface of what LLMs can actually bring to analytics.
Analytics is NOT regurgitating the same KPIs every month. It’s about asking new questions, refining your hypotheses, and uncovering patterns you didn’t even know to look for. If you’re handing over the keys to an LLM without interrogating its output, you’re outsourcing the very thing that makes analytics valuable: human curiosity and creativity.
Writing SQL Isn’t the Hard Part
Writing SQL is rarely the bottleneck in building great analytics. Often it's the ecosystem.
-
System Tradeoffs. Picking between Snowflake, BigQuery, or something else is more than a query performance or cost-per-TB conversation. It’s about tradeoffs:
-
How does the system integrate into your tech stack?
-
What operational overhead does it impose?
-
How does it scale under heavy, concurrent workloads?
-
LLMs can contextualize these tradeoffs, providing tailored recommendations based on usage patterns, schemas, and SLAs.
-
-
Data Modeling.
-
Designing a resilient data model is one of the most intellectually demanding parts of analytics.
-
A schema that works for today’s queries might crumble as data grows or as teams ask increasingly complex questions.
-
LLMs can automate modeling recommendations by analyzing query logs, spotting anti-patterns, and predicting future requirements.
-
-
Data Transformation.
-
ETL/ELT isn’t just “clean the data.” It’s a minefield of dependencies, schema drift, and idempotency challenges.
-
LLMs can draft robust transformation pipelines, auto-generate dbt models, and proactively suggest data quality tests based on historical failures.
-
-
Orchestration.
-
A failed job in your pipeline at 3 a.m. isn’t just an inconvenience—it’s a blocker for downstream systems.
-
Tools like Airflow or Dagster help, but managing interdependencies at scale requires intricate DAGs that are brittle under change.
-
LLMs can dynamically reconfigure pipelines, adapt schedules based on query priorities, and even predict failures before they occur.
-
-
Data Storytelling: Designing dashboards that resonate with users and communicate insights effectively is as much an art as it is a science.
Sure, LLMs can write a query faster than you can say "SELECT *," but solving these larger challenges is where the value is actually added.
Rethinking the Interface: The Human-in-the-Loop
Analytics is inherently human. While LLMs can suggest, summarize, and transform, the user’s intent, often nuanced and domain-specific. remains central. A system that’s too prescriptive risks missing the point entirely. That said, LLMs can be a great companion for brainstorming and report planning.
Before you ever write a query, you need to know what you’re looking for. LLMs can:
-
Help draft report plans, outlining potential metrics, dimensions, and comparisons.
-
Brainstorm unexplored opportunities by suggesting alternative data angles or KPIs.
-
Guide analysts to think critically about their hypotheses, rather than just automating the first query they write.
This planning phase is where the seeds of meaningful insights are sown.
-
Narrative Control. LLMs can propose insights, but the overarching story needs to come from a human. It’s the difference between data exploration and insight curation.
-
High-steer systems often fail because they can’t account for the iterative nature of analysis. A human-in-the-loop approach lets analysts refine, redirect, and reframe their questions dynamically.
The Rise of Development Agents
For the past decade, SaaS tools have dominated the analytics landscape. They abstracted away complexity, sure. But in doing so, they also boxed us in. Each tool came with its own interface, its own quirks, and its own silos. Need orchestration? Use Tool A. Need transformation? Tool B. Reporting? Tool C. Managing this patchwork stack became its own full-time job.
With development agents, we’re seeing the rise of intelligent assistants that don’t just write SQL but also:
-
Draft Python/dbt transformation logic.
-
Propose schema optimizations.
-
Automate orchestration and error handling.
Suddenly, the need for a sprawling SaaS stack starts to look... optional. The question isn’t, “Which tool should I buy?” It’s, “What’s the minimum viable stack I need if code becomes this frictionless?”
Dynamic Data Dictionaries and Contextual Knowledge Bases
Data dictionaries are supposed to be the Rosetta Stone of analytics. In reality, they’re stale, static, and ignored. How many times have you seen a dictionary that doesn’t match what’s in the warehouse? Or one that hasn’t been updated in months (or years)? LLM-powered knowledge agents go beyond being reference guides. They’re context-aware agents that can:
-
Answer nuanced questions about data definitions.
-
Surface related metrics or columns based on your analysis.
-
Automatically update themselves to reflect schema changes.
The static dictionary is dead. It’s time to embrace tools that evolve as fast as your data does.
Making Reports Dynamic, Not Static
Traditionally, reports have been static artifacts: the same KPIs, refreshed periodically. But in a world where LLMs make data more accessible, this paradigm doesn’t hold up. Static reports can give way to dynamic exploration:
-
Imagine a dashboard that updates not just with new data, but with new questions based on changing business conditions.
-
Reports that morph to fit the context of their audience, focusing on what matters most at that moment.
Unlocking a New Era of Exploration
Perhaps the most transformative potential of LLMs is the explosion of exploratory data analysis (EDA). Historically, the high cost of building models, maintaining pipelines, and interpreting results has limited how much exploration teams could afford. But if those costs plummet:
-
Analysts can ask 100x or 1000x more questions in the same amount of time.
-
Businesses can surface insights that were previously buried under the constraints of bandwidth and tooling.
-
Data exploration can become a routine part of decision-making, not just a high-effort, one-off task.
The Future
Turn analytics from a linear, tool-driven process into a fluid, dynamic interaction between humans and machines. The flashy demos of “chat-with-your-data” apps were just the opening act. Human creativity and machine intelligence should work together to uncover insights that were previously unimaginable.
Try Preswald today!