OpenAI's Internal Data Analyst: How Kepler Transforms AI-Powered Insights in February 2026

When OpenAI's own internal data analysts couldn't quickly answer "How many ChatGPT Pro users are in France?", it wasn't because the data didn't exist. They had 600 petabytes across 70,000 datasets. The issue was getting to it. Every simple question required hunting through 15 tools, reading schema documentation, and waiting for data eng to clarify which table held the answer. Simple queries took hours or days. Talented analysts spent more time searching than analyzing. OpenAI built Kepler to fix this, letting anyone ask questions in plain English and get accurate answers in seconds, without needing to know which of 70,000 datasets to query or how to write SQL.

TLDR:

• OpenAI built Kepler to query 600 petabytes across 70,000 datasets using plain English instead of SQL.

• The six-layer context system validates queries against metadata and business rules before returning answers.

• Natural language data agents cut decision loops from hours to seconds for any team member.

• Index delivers conversational data access that connects to your warehouse and learns your schema instantly.

Why OpenAI Built an Internal Data Agent

OpenAI runs some of the most sophisticated AI systems in the world, yet their analysts were stuck in a familiar trap: drowning in data they couldn't quickly access.

The problem wasn't a lack of information. OpenAI stores hundreds of petabytes across 70,000 datasets, tracking everything from model training runs to user behavior across ChatGPT, API usage, and enterprise deployments. The issue was navigation. Answering a simple business question like "How many ChatGPT Pro users are in France?" required analysts to hunt across 15 different tools, dig through schema documentation, and often wait on data eng to clarify which table held the right answer.

Even with talented data teams, this created bottlenecks. Simple queries took hours or days. Analysts spent more time searching for data than analyzing it. Business decisions stalled while teams waited for numbers that should have been instant.

OpenAI built Kepler to close that gap. The goal was to let anyone ask questions in plain English and get accurate answers in seconds, without needing to know which of 70,000 datasets to query or how to write the SQL.

Technical Architecture: How Kepler Processes 600 Petabytes

Kepler's architecture solves a problem that sounds simple but is brutally hard at scale: how do you let an AI agent accurately query 600 petabytes without returning garbage?

The answer is a six-layer context system that narrows the search space before any query runs. First, Kepler maps metadata across all 70,000 datasets in the data warehouse, cataloging table schemas, column definitions, and relationships. Second, it tracks data lineage to understand where metrics come from and how they're calculated. Third, it applies business logic rules that encode institutional knowledge about what queries make sense. Fourth, it layers in access controls so users only see data they're allowed to touch. Fifth, it maintains a conversation history to understand follow-up questions. Sixth, it validates query results against known benchmarks before returning answers.

This context stack feeds into GPT-5.2, which powers Kepler's reasoning engine. The model generates SQL while determining which of 70,000 datasets are relevant, constructs the query, and sanity-checks the output.

OpenAI also built Kepler using Anthropic's Model Context Protocol, which standardizes how AI agents interact with data sources. MCP lets Kepler connect to different databases, APIs, and tools without requiring custom integrations for each one.

Natural Language Queries Replace SQL Backlogs

Kepler lives inside Slack, where OpenAI employees type questions like "Show me API usage trends for enterprise customers this quarter" or "Which features are ChatGPT Plus users engaging with most?"

No SQL. No schema hunting. No tickets to the data team.

The agent interprets the question, selects the right datasets, runs the query, and returns a chart or table directly in the Slack thread. Follow-up questions work the same way: "Break that down by region" or "Compare to last quarter" flow naturally without restarting the conversation.

Analysts become reviewers instead of query writers. Product managers pull their own metrics. Support leads check ticket volumes without opening a BI tool. Sales ops answers pipeline questions on the fly.

Questions that used to pile up in a backlog now get answered in seconds by the person who asked. Data access stops being a bottleneck controlled by a small team with SQL skills and becomes something anyone can do when they need an answer, democratizing business intelligence (BI) across the organization.

Context and Memory: The Critical Accuracy Layer

The difference between a data agent that works and one that hallucinates comes down to context and memory.

Context is the metadata layer that tells the agent what data exists and how to interpret it. Table schemas, column definitions, and business logic rules in the data catalog give the agent a map of what's queryable. Without this, the agent guesses at table names, misreads column types, or joins data that shouldn't be connected. OpenAI credits context for getting Kepler to 80 or 90% accuracy.

Memory is the learning mechanism. It tracks past queries, corrections, and clarifications. When someone asks "Show me revenue," memory combined with the semantic layer knows whether they mean gross, net, or recurring. When they follow up with "Now by region," memory understands the original question and adjusts the query instead of starting over.

Together, context and memory prevent catastrophic errors. A query that pulls the wrong customer segment or misreads churn rates can sink product decisions, misallocate budgets, or wreck forecasts.

Human review still matters. The agent handles the mechanics while people judge whether the answer makes business sense.

Security Controls for Autonomous Data Access

Allowing AI agents to roam freely through enterprise data poses real risks. 48% of security professionals call agentic AI the top attack vector for 2026.

OpenAI solves this with layered access controls. Kepler inherits the credentials of whoever asks the question. Users only see data they already have permission to touch. If you can't access revenue data through SQL, you can't access it through Kepler either.

Audit logs track every query the agent runs. Who asked it, what data it accessed, and what results it returned. Teams can review agent behavior retroactively, spot anomalies, and debug incorrect answers.

Quality checks run before results surface. Kepler validates outputs against known metrics and flags queries that return unexpected values or contradict business logic rules.

The tradeoff is speed versus control. Autonomous agents move fast. Mistakes can propagate before someone catches them. Guardrails slow things down but prevent bad data from driving decisions.

Human oversight remains the final layer. The agent accelerates analysis, but people decide whether to act on the answer.

What This Signals for Enterprise Data Analysis

OpenAI's internal use of Kepler follows a pattern: AI companies test their own tools before releasing them. GitHub Copilot ran inside Microsoft for months. Anthropic's Claude tools tested internally first. When the builders need it badly enough to build it before customer-facing features, the tech is getting real.

The bigger shift is from dashboards to conversational BI. Static BI reports freeze assumptions at design time. Someone decides which metrics matter, how to slice them, and what visualizations to show. Questions outside that framework require new dashboards or analyst time.

Conversational data agents flip that model. Users start with a natural-language query, not a pre-built chart. Follow-ups flow naturally. The interface adapts to what you're trying to learn instead of forcing you into fixed views.

66% of organizations deploying enterprise AI report measurable productivity gains. The wins come from collapsing decision loops, not automating entire workflows.

Data agents won't replace BI tools overnight. But they change who can ask questions and how fast answers arrive.

Bringing AI Data Analysis to Your Team

You don't need OpenAI's resources to deploy a working data agent.

Index delivers conversational data access without requiring an AI research team or custom infrastructure. Ask questions in plain English. Get charts and tables in seconds. No SQL required.

Index, one of the leading AI-powered business intelligence tools, connects directly to your existing data warehouse (Snowflake, BigQuery, Redshift, Postgres, ClickHouse) and learns your schema. The AI learns your metrics, business logic, and table relationships, so queries hit the right data every time. Follow-up questions flow naturally.

Teams of 20 or 500 get the same capabilities. Product managers pull retention cohorts. RevOps tracks pipeline velocity. Support leads monitor ticket trends. All without waiting on data eng or learning query languages.

The gap between demos and production is in accuracy under real conditions. Index combines language understanding with validation layers that check results against known metrics before surfacing answers. Fast queries matter less if the numbers are wrong.

Final Thoughts on Conversational Data Access

OpenAI's internal Kepler builds prove that AI data analysts solve a real bottleneck beyond demos. The shift from static dashboards to natural language queries changes who can ask questions and how fast decisions move. Your analysts should review answers instead of writing SQL, and product managers should pull retention cohorts without waiting on data eng. Book a demo to see how Index handles your warehouse schema and business logic without months of setup or custom integrations.

FAQs

How does Kepler achieve high accuracy across 70,000 datasets?

Kepler uses a six-layer context system that maps metadata, tracks data lineage, applies business logic rules, enforces access controls, maintains conversation history, and validates query results against known benchmarks before returning answers. This context stack feeds into GPT-5.2, which assesses relevance before constructing and executing queries.

What makes conversational data agents different from traditional BI dashboards?

Conversational agents let users start with any question and follow up naturally, adapting to what you're trying to learn in real-time. Dashboards freeze assumptions at design time, forcing you into pre-built views where questions outside that framework require new builds or analyst time.

Can non-technical teams really get accurate answers to data questions without SQL?

Yes, if the agent has proper context and memory layers. The agent needs access to your schema, business logic, and metric definitions to correctly interpret questions. Product managers, support leads, and sales ops can pull their own metrics when the underlying structure is sound and validation checks run before surfacing results.

What security risks come with letting AI agents access enterprise data?

48% of security professionals call agentic AI the top attack vector for 2026. The main risks are unauthorized data access, incorrect queries propagating bad decisions, and audit gaps. Proper controls include credential inheritance (users only see data they already have permission to access), query audit logs, and validation checks before results are returned.

How long does it take to deploy a conversational data agent?

Index connects to your data warehouse and learns your schema in minutes, versus the weeks or months traditional BI tools require. The faster timeline comes from direct warehouse connections (Snowflake, BigQuery, Redshift, Postgres, ClickHouse) and automated schema learning instead of manual configuration and custom integrations.