How a Major Retailer Used NL2SQL to Let Teams Ask Data Questions Without SQL

Overview

A large department store wanted merchandisers, planners and store teams to get reliable answers from their data without learning SQL or waiting on the data team.

Instead of tickets, spreadsheets and delays, they wanted people to ask questions in plain German (or any language they chose) and get clear, trustworthy results straight away.

We built a secure AI-powered NL2SQL assistant: a small web app where users log in, type a question and receive:

• a clean data table
• a simple interactive chart
• a short explanation of the result

The application runs entirely in the retailer’s own cloud environment, keeping data secure and under their control.

The MVP went live on 30 October 2025. Phase 2 focuses on performance, security and scalability.

The problem: great data, trapped behind SQL

This retailer had strong underlying data. The issue wasn’t quality; it was access.

To answer simple questions like:

• sales by store
• promotion performance
• stock levels

teams had two options:

1. Raise a ticket with the data team
2. Learn SQL themselves

Both slowed decision-making. Business teams waited days for answers, analysts became a bottleneck, and opportunities were missed.

This is a common problem in retail analytics: data exists, but only specialists can reach it.

What they wanted: instant, trustworthy answers

They weren’t looking for dashboards or flashy AI demos.

They wanted:

• self-service access to data
• answers in plain language
• results they could trust and verify
• something their teams would actually use

In short: conversational analytics that worked in the real world.

How we helped: a practical NL2SQL AI assistant

We deliberately kept the scope small and focused.

1. We started with real questions

We listened to merchandisers, planners and analysts and mapped the everyday questions they were already asking.

2. We built an agentic AI web app

In simple terms, this is an AI data assistant that:

• understands a question in natural language
• determines how to query the data
• generates the SQL
• returns the result as a table, a chart and a short explanation

Crucially, every answer includes the SQL used, so teams can trace and validate the output.

3. We tested with real users

Business users tested the tool, shared feedback and helped refine the experience. This ensured the system worked for real decision-making, not just demos.

4. We handed it over cleanly

The solution runs in the client’s cloud and was delivered with:

• a short support window
• clear handover documentation
• no major platform changes or rewrites

The retailer owns the solution and can scale it at their own pace.

What we delivered (MVP)

• A working NL2SQL conversational analytics app
• Questions asked in plain German (or other languages)
• Clear outputs: data table, interactive visual and reasoning
• Full SQL traceability for every answer
• Handover documentation and support
• MVP launched on 30 October 2025

What success looked like

Before going live, we agreed clear, practical targets:

• ~95% successful execution for valid queries
• Response times of 30–90 seconds (reducible with more SQL compute)
• ~99% service uptime
• User acceptance in UAT (User Acceptance Testing) – real users confirmed the outputs were useful and trustworthy

Testing captured role-based feedback to prove the tool worked across teams.

Early wins

• Teams stopped waiting on analysts for simple answers
• Merchandisers checked promotions and stock in minutes, not days
• The business had a clear path to production rollout

Small changes, but meaningful impact: faster decisions, fewer errors, less friction.

What’s next: Phase 2

Phase 2 focuses on turning the MVP into a production-ready service:

• improved performance and reliability
• stronger security controls
• better chat history and usability
• lower cloud running costs

Planned enhancements include:

• domain-specific AI agents and orchestration
• performance monitoring
• structured user feedback
• fine-tuned language models

Once complete, the retailer will have a scalable AI data assistant ready for wider rollout.

Why this matters

This isn’t AI theatre.

It’s about giving business teams direct access to data they already have, with proof the answers are right.

The result:

• quicker decisions
• fewer mistakes
• better use of existing data

Want to try this for your team?

If your people are stuck waiting on SQL for simple answers, we can build a low-risk NL2SQL MVP, test it with real users and hand over a secure, usable system.