Why Business Users Can't Query Data Warehouses Without SQL (And What To Do About It)

You know the answer is in there. The data warehouse has every order, every SKU, every supplier invoice, every store-level transaction your company has ever processed. But when you need to know which products had the lowest sell-through last month, or which carriers are consistently missing delivery windows, you file a ticket.

L. The problem is structural, and it exists in nearly every enterprise that runs its operations on Snowflake, BigQuery, Databricks, or Redshift. This article explains exactly why the barrier exists, why the most popular fixes don't actually fix it, and what a real solution looks like.

Can business users query a data warehouse without writing SQL?

Yes, but only with a tool that includes a semantic layer. A data warehouse stores information in machine terms: column names like ACCT_STATUS_CD or INV_TRX_DT that mean nothing without context. Business users think in business terms: "active accounts," "inventory transaction date." Text-to-SQL tools translate words into queries. A semantic layer translates meaning, mapping your company's KPIs, custom terms, and business logic to what they actually represent in the database. Without it, even sophisticated AI query tools produce wrong answers on real enterprise data.

The Data Warehouse Was Never Built for You

Data warehouses are architecturally designed for storage efficiency and machine retrieval, not for answering business questions in plain English.

When engineers build a Snowflake schema or a BigQuery dataset, they optimize for query speed, storage costs, and data integrity. Column names get shortened. Tables get normalized and split. Business concepts get encoded in abbreviations that only the person who wrote the ETL pipeline fully understands. The schema is a machine's map of the data: precise, efficient, and completely opaque to anyone who didn't build it.

SQL was created in 1974 for database engineers. It asks users to know not just what they want to find, but exactly where it lives, what it's named, and how it relates to every other table involved in the answer. A supply chain manager asking "which vendors are delivering late" has to know the name of the orders table, the name of the shipments table, how those tables join, and what the date fields are called, before typing a single character of their actual question.

The result is a predictable organizational pattern: a small group of SQL-literate analysts serving a large and growing queue of business users who can't access their own company's data. A 2020 survey of Fortune 1000 executives by New Vantage Partners found that only 24% believed their organization was genuinely data-driven, and access, not quality, was the primary barrier cited.

The data exists. The problem is the translation layer between human questions and machine storage, and for most enterprises, that layer doesn't exist yet. Understanding what self-service analytics actually requires makes it clear why closing that gap takes more than buying another BI tool.

Why SQL Is the Default, and Why That's a Problem

The data team isn't the enemy. They're outnumbered.

In most mid-to-large enterprises, a handful of analysts serve dozens or hundreds of business users across supply chain, operations, merchandising, FP&A, and sales. Each request, "give me sell-through by SKU for the top 20 accounts," requires the analyst to understand the question, locate the right tables, write a correct query, validate the output, and format it for the person who asked. That cycle takes one to three days on average, assuming the analyst isn't already buried.

By the time the report arrives, the meeting where it mattered has passed.

The deeper problem is that even the most diligent data team can't scale. Business users generate questions faster than analysts can answer them. So they filter. They prioritize the urgent over the important. They build standardized dashboards to handle the recurring requests, which leaves the ad-hoc questions, the ones that actually surface new insight, perpetually at the back of the queue.

Lyle O'Banion, a demand planner at Kroger, described this constraint directly when talking about supply chain analysis: "Analyzing this volume of data is very time-consuming. Lumi's ability to de-average and re-aggregate down to store-item impacts makes it far more efficient." The data existed in Kroger's warehouse. The bottleneck was access: getting from a business question to a query result without routing through a three-day analyst cycle.

The ticket-based model isn't a workflow problem you can solve by hiring more analysts. It's a symptom of a system where business users and their data are separated by a language neither side can speak fluently. Understanding why the most obvious technological fixes fail to resolve this is where most organizations get stuck.

The Promise (and Real Limits) of Text-to-SQL

The logical fix sounds obvious: use AI to translate plain English into SQL automatically. Type your question, the AI writes the query, the warehouse runs it, you get your answer. This is called text-to-SQL (or NL2SQL), and it's what most vendors selling "AI analytics" are actually describing.

On controlled academic benchmarks, it works remarkably well. The Spider 1.0 benchmark, a widely used text-to-SQL evaluation, shows state-of-the-art models exceeding 85% accuracy. That number gets cited in a lot of vendor pitch decks.

It does not reflect what happens on your database.

In April 2025, researchers evaluated leading text-to-SQL systems against Spider 2.0, a benchmark built on realistic enterprise schemas with queries over 100 lines long, tables with more than 1,000 columns, and the kind of complexity found in actual production environments. The best model achieved 31% execution accuracy. Uber built an internal NL2SQL tool for its own data and reported 50% overlap with ground truth tables on its internal evaluation set.

The gap between academic and enterprise performance has a specific cause. Enterprise schemas are optimized for storage, not retrieval. They contain nested tables, multi-dimensional structures, deprecated fields sitting alongside active ones, and columns with names that encode meaning only engineers remember. When an AWS and Cisco team published their analysis of enterprise-grade NL2SQL challenges, they identified complex schemas as the primary failure mode, not the AI's language understanding, but its inability to navigate the structural complexity of real production databases.

There's a second failure mode that's harder to catch: the query runs successfully and returns data, but the data is wrong. The AI interpreted "active customer" as any account with a non-null status field, when your company defines "active" as any account with a purchase in the last 90 days. The SQL executed. The result looked correct. The insight was false. This is the core problem that navigating the data bottleneck with semantic parsing is specifically designed to solve.

Text-to-SQL solves the syntax problem. It doesn't touch the semantic problem, and in enterprise analytics, semantics are where the actual work lives.

Why BI Dashboards Don't Solve It Either

The other standard answer is dashboards: invest in Tableau, Power BI, or Looker, build a library of reports, and let business users serve themselves. It sounds right in a budget meeting and breaks down in practice within six months.

Dashboards answer the questions someone anticipated. They work well for recurring reports on known metrics: weekly sales by region, inventory levels by warehouse, on-time delivery rates by carrier. For those use cases, a well-built dashboard is genuinely useful.

The problem is that the questions business users actually need answered in real time are rarely the ones anyone anticipated. "Which of our top-20 accounts by revenue has the highest rate of unfulfilled demand at the store-item level this month?" is not a question someone built a dashboard for. Neither is "Which raw material categories are driving the unexpected cost increases in procurement this quarter?" These are the questions that lead to decisions worth millions of dollars, and they live permanently in the backlog.

When users hit the edge of a dashboard, they have two options: file a ticket or try to customize. Customization in Tableau or Power BI requires either dragging pre-defined fields (which doesn't help with novel questions) or writing SQL (which returns the original problem). ThoughtSpot, which pioneered search-based analytics and is often cited as the solution to this problem, is consistently described as having a high learning curve for non-technical users, a direct comparison worth examining if your team is currently evaluating it against more purpose-built alternatives.

The deeper issue is what happens over time. Organizations accumulate dashboards. Analysts build them in response to requests, business contexts shift, and the dashboards stop reflecting reality while remaining in the system. Companies end up with hundreds of reports, most unused, none of which answer the question someone is asking today. Lumi AI's investors and the press covering its seed round called this "dashboard anarchy," and the term resonates with anyone who has tried to find the authoritative version of a metric across six slightly different reports.

Dashboards are a storage mechanism for anticipated questions. Business users need something that can answer questions that haven't been asked before.

The Missing Layer: Business Context, Not Just Query Translation

The structural fix is not a better SQL generator. It's a layer that sits between the AI and the database, one that encodes what your business actually means by its own terminology.

This is called a semantic layer. It maps business terms to database columns, defines KPIs in precise and consistent terms, encodes business logic (what "fiscal week" means, how "top accounts" are ranked, what qualifies as an "active" customer), and stores all of this as shared, queryable knowledge. The AI doesn't have to guess what "sell-through" means at your company. It knows, because someone defined it.

Without a semantic layer, the AI is a skilled translator working without a dictionary. It can handle the grammar. It can't handle the vocabulary.

With a semantic layer in place, a supply chain manager can ask "which vendors have the highest rate of short shipments against PO quantities this quarter" and get a correct, auditable answer, because the system knows what "short shipment" means, how PO quantities are stored, and which field represents the current fiscal quarter. The SQL the AI generates is grounded in definitions your team provided, not in guesses about your schema.

Jordan Kuhns, Director of Wholesale Technology Portfolio at GROWMARK, one of North America's largest agricultural cooperatives, described what happened when his organization deployed Lumi: "Lumi gave us a tangible reason to invest the time in documenting our data. Every time we add something, we should be that much better at answering the next question." That's the semantic layer in practice. Each business definition added to the system makes every future query more accurate.

Lumi AI's Knowledge Management feature is the implementation of this concept. Business teams define their custom metrics, field aliases, and KPI logic directly in the platform. The conversational AI uses that context in every query, so the answers reflect the company's actual business definitions, not the database engineer's column naming conventions.

The semantic layer also enforces consistency at scale. When everyone queries "churn rate" and gets the same answer because the definition is codified and shared, you eliminate the version-of-truth problem that plagues most analytics environments.

What Actually Works: A Framework for Business User Data Access

There are three meaningfully different approaches to giving business users access to warehouse data, and they're not interchangeable.

Tier 1: Dashboards

Dashboards handle recurring reports on pre-defined metrics. They are the right tool for questions that repeat: weekly revenue by region, monthly inventory turns, daily order fill rates. They require analyst involvement to build and maintain, and they break the moment someone needs a question answered that falls outside the pre-built views. SQL is still required for customization.

Best fit: teams with stable, predictable reporting needs and dedicated analyst support.

Tier 2: NL-to-SQL tools (without a semantic layer)

Generic text-to-SQL tools let users type questions in plain English and receive generated SQL in return. They work reasonably well on simple schemas with consistent naming conventions, common in early-stage startups and smaller data environments. On complex enterprise schemas (ERP data, multi-system warehouses, large Snowflake or BigQuery implementations), accuracy degrades to levels that require constant analyst oversight to validate.

Best fit: small teams with simple schemas who are comfortable reviewing AI-generated SQL before acting on results.

Tier 3: Conversational AI with a semantic layer

This approach layers business context on top of NL-to-SQL. A knowledge base or semantic layer is configured with the company's terminology, KPI definitions, and business logic. The AI generates queries grounded in that context rather than guessing from schema alone. Business users get accurate, auditable answers in plain English with visualizations, without writing SQL or waiting for analyst involvement.

Best fit: supply chain, operations, FP&A, and merchandising teams at mid-to-large enterprises running ERP and warehouse systems like Snowflake, BigQuery, Databricks, or SAP.

Tier 1: DashboardsTier 2: NL-to-SQLTier 3: Conversational AI + Semantic LayerSQL required for ad-hoc queriesYesNoNoHandles novel questionsNoPartiallyYesBusiness context awareNoNoYesAccuracy on enterprise schemaHigh (for pre-built)Low (10-31%)High (with configured knowledge base)Time to answerHours to daysSeconds (with validation needed)SecondsAnalyst involvement requiredAlwaysOftenRarely

The right tier depends on the complexity of your schema, the variability of your questions, and how much analyst oversight your team can realistically provide.

What This Looks Like in Practice

The difference between Tier 2 and Tier 3 becomes concrete when you look at what enterprise teams actually need to know.

Kroger's demand planning team needed to analyze unfulfilled demand at the store-item level, a granular query that required de-averaging aggregate data down to individual store-SKU combinations. This is exactly the kind of multi-step analytical question that breaks generic NL-to-SQL tools: too much schema complexity, too many joins, too many business-specific definitions embedded in the columns. With a semantic layer, the system understood what "store-item level" meant in Kroger's data model and surfaced the unfulfilled demand analysis directly, a task that had previously taken significant analyst time to structure correctly.

GROWMARK, a leading agricultural cooperative managing complex procurement data across member companies, faced a different version of the same problem. Their data lived in Amazon Athena. Their business users were not data scientists. Before deploying Lumi, teams spent time "looking for the right folder or report," in the words of Kuhns, navigating a fragmented reporting environment that hid the answers they needed. After configuration, those same users could ask statistical questions in plain language and get results that previously would have required dedicated data science involvement.

Chalhoub Group, the largest luxury retailer in the Middle East, used conversational querying against its sales data to identify $60 million in additional revenue opportunities, a finding that had been invisible to its traditional BI stack. The insight wasn't buried deep in the data. It was a question no one had thought to ask, answered in seconds once business users had direct access to query the warehouse.

The pattern across these cases is consistent: the data was always there. The bottleneck was the layer between the business question and the warehouse, and removing that bottleneck changed what the business could see.

Frequently Asked Questions

Can business users really query a data warehouse without writing SQL?

Yes, with the right tool. A conversational AI platform that includes a semantic layer lets business users ask questions in plain English and receive accurate answers with visualizations, without writing any SQL. The critical requirement is that the platform must understand your company's specific business terms, not just the raw database schema. Without that business context layer, accuracy degrades significantly on real enterprise data.

What is the difference between text-to-SQL and a semantic layer?

Text-to-SQL converts a natural language question into a SQL query. A semantic layer maps business concepts, your KPIs, custom terminology, fiscal conventions, and business rules, to what they actually mean in the database. Text-to-SQL handles the translation of words into syntax. The semantic layer handles the translation of meaning. On academic benchmarks with clean schemas, text-to-SQL alone achieves 85%+ accuracy. On realistic enterprise databases, that number falls to 10-31% without a semantic layer providing business context.

Why do tools like Tableau and Power BI still require SQL for custom queries?

Tableau and Power BI were designed for analyst-authored dashboards, not ad-hoc business user queries. Drag-and-drop functionality works for pre-defined metric combinations. The moment a user asks a question that falls outside the pre-built report, a novel filter, a new dimension, a calculation that wasn't anticipated, they hit a wall that requires SQL or analyst involvement to get through. These tools are built to display data, not to answer open-ended questions about it.

How do I know if the AI's answer is actually correct?

Any platform you deploy for business user querying should surface the SQL it generated, show the logic steps it followed, and allow validation before anyone acts on the result. Lumi AI includes a human verification workflow that lets users flag results for expert review, keeping the speed of AI with the accuracy checks that high-stakes decisions require. Platforms that return answers without showing their reasoning are not suitable for enterprise use.

The SQL Barrier Is a Design Problem: Here's Your Next Step

Business users shouldn't have to learn SQL any more than they should have to learn to build their own database infrastructure. The barrier isn't a skills gap to be trained away. It's an architectural mismatch between how enterprise data is stored and how humans think and ask questions.

The fix follows a specific sequence: configure a semantic layer that encodes your business terminology, then layer conversational AI on top of it. The AI handles the query. The semantic layer ensures the query is grounded in what your business actually means. The business user gets an answer in seconds rather than waiting three days for a ticket to clear.

If your organization runs supply chain, retail, or CPG operations on Snowflake, BigQuery, Databricks, or an ERP like SAP or Oracle, book a demo with Lumi AI to see question-first analytics working against your actual data, your schema, your terminology, and your business questions.

For teams still evaluating the landscape, Lumi's Foundations of AI-Native Analytics whitepaper covers the architectural concepts in more depth.

The data your team needs to make better decisions is already in your warehouse. The only question is whether your current setup lets anyone actually use it.