Case Study · AI Engineering · Retail Analytics
Most "AI dashboards" are a natural-language search bar sitting on top of charts someone already built. We built something different for a retail analytics platform: a live reasoning layer that composes new answers on real data, cites every number it states against the actual query result, and refuses honestly the moment the data can't back up a claim, instead of guessing.
An "AI-powered" dashboard should mean the model can reason about your data, not that a chat box was bolted onto the same static charts.
Manual vs. AI, Same Screen
This is one real screenshot, not two mockups stitched together. Left, the manual path: open the catalog, find the right table among nine, read that it has 24 columns and a stated grain, then work out the query yourself. Right, same screen: type the question in plain English and get a cited answer in seconds, with the exact number of verified values it drew from disclosed underneath.
Why AI, Not Just a Dashboard: In Plain English
You don't need to know what a "semantic layer" is to see the difference. Here are four real situations, captured directly from this system while it was running, not mockups, not hypotheticals, each with the everyday version first and the proof underneath.
1. It tells you what to do, not just what happened
A normal dashboard shows you a table: 400 products below reorder threshold, sorted by whatever column you clicked. It's still on you to scroll through and decide which ones actually matter. This system has already done that thinking for every single product. It tells you, in a plain sentence: reorder this one, it sold 40 units this week and has two weeks of stock left; skip that one, it's been sitting untouched for three months. What this means for your business: your team stops spending an afternoon a week in a spreadsheet deciding what to reorder, and starts acting on the answer the same day.
2. It answers the question nobody thought to prepare for
Every dashboard you've ever used only answers the questions someone anticipated when they built it. The moment you ask something new, connecting two things nobody thought to link together in advance, a normal dashboard has nothing to show you. Someone has to file a request with IT or an analyst and wait days, sometimes weeks, for a new report. This system answers that same new question in seconds, safely, without a developer writing a line of code. What this means for your business: you stop waiting in a queue for every new business question, and start getting an answer the moment you think to ask it.
3. It says "I don't know" instead of a confident wrong answer
Ask most AI chat tools a question they can't really answer, and many will still hand you something, a number that looks perfectly normal on screen. You'd have no way to know it was made up unless you happened to double-check it yourself. Ask this system the same kind of question when the underlying cost data simply isn't loaded, and it tells you plainly that it can't calculate that, and why, instead of quietly inventing a number that looks fine until someone has already acted on it. What this means for your business: a wrong number never quietly slips into a decision, because the system tells you the moment it isn't sure, every time, regardless of how the question is worded.
4. It knows when AI is overkill, so most of your numbers stay free and instant
Not every question needs a language model to answer it. "Sales by category this month" is the same question every single day. Having a model reason through it fresh each time is slower and costs real money for no real benefit. This system recognizes routine, repeatable questions like that and answers them instantly through simple point-and-click buttons, at zero AI cost, saving the actual reasoning for the handful of questions that genuinely need it (see the Quick Report builder in the product walkthrough below). What this means for your business: your everyday dashboard numbers stay instant and free forever. You only ever pay for AI on the questions that truly require thinking, not on the ones a button could answer just as well.
None of these four are about the model being clever. They're about the system doing something a static set of charts cannot do by construction, and knowing exactly when not to bother, and when to stop cleanly and say so.
The Business Problem
Dashboards show numbers. They don't answer questions. Someone still has to know which filter to click, and when the answer genuinely isn't there, nothing tells you that. Bolting a chat box onto the same pre-built charts doesn't fix this; it just makes the dead end conversational.
A chart can look right and still be measuring the wrong thing: the wrong date range, a silently dropped join, cancelled orders counted in when they shouldn't be. Nothing on the screen tells you which.
Most natural-language BI features answer by matching your question to a chart or topic someone already indexed. Ask something nobody anticipated and it has nowhere to go.
A model that narrates over your data can state a number that sounds right and isn't. Without a verification step, confidence and correctness are the same thing to the reader, until they aren't.
Knowing which metric, which filter, which join to use is schema knowledge most business users don't have and shouldn't need, just to ask what changed last month.
Our Approach
The model never touches raw SQL and never sees the database. It picks from a closed, reviewed vocabulary of metrics, dimensions, and filters, declared once, in code, and every plan it builds is validated at eight separate stages before a single query executes. The result is checked against what actually ran before it's allowed to speak.
We don't let the model improvise SQL, and we don't let it state a number it can't prove. Those two rules are the entire trust model: everything else in the architecture exists to enforce them without slowing the product down.
What "revenue" means, which joins are safe, and what the data genuinely can't answer (and why) are declared in one semantic layer, read by the agent, the report builder, and every dashboard card alike, so none of them can drift apart.
The model only ever emits a typed, strictly-validated query plan, never free text SQL. A dedicated compiler is the one and only thing that ever generates SQL, so the attack surface for a bad query is a small, reviewed piece of code, not a language model.
The model's final answer must cite the exact row and column each number came from. A deterministic check confirms every citation against the real result set: a mismatch withholds the sentence and shows the raw data instead.
See the Actual Plan, Not Just the Answer
"Bound what the model can do" isn't a claim you have to take on faith. Every answer has a "How this was calculated" panel that expands into the exact typed query plan the model produced, the real tables and columns it touched, and the row count it scanned, the same object the compiler and validator actually ran against, not a simplified explanation written after the fact.
Reference Architecture
execute_sql is never exposed to any adapter, including the model.Product Walkthrough
The interface is built for someone who has never heard the words "dimension" or "semantic layer." Every capability below is reachable by clicking, not by knowing the right question to type.
Real Answers, Real Charts
A chart here is never a separate "visualization layer" guessing at intent: it's the literal rows the query returned, rendered with the form that fits the data's own shape. A ranked comparison reads as a bar; a share of a fixed whole reads as a donut. Every number on both charts below is real output from this system, not a mockup.
Business Outcomes
Citations are checked against real query output before an answer ships. A confident-sounding wrong number isn't a risk the reader has to catch themselves.
The agent can compose a new answer live, including joining data sources with no pre-declared relationship, instead of only surfacing what an analyst already anticipated.
When the data genuinely can't answer a question (no cost data to compute margin, for instance), the system says so plainly, explains why, and offers what it can answer instead.
Daily numbers and recommendation cards run through a fixed, deterministic path with zero model calls, no latency, no token cost, for the questions asked every single day.
The chat agent, a no-AI report builder, and the pinned dashboard cards all read the same declared metric and join definitions: "revenue" can't quietly mean something different depending on which screen you're looking at.
Every request is authenticated before it can spend a token, and a per-user query quota is enforced atomically, proven under real concurrent load, not just assumed safe.
Technical Highlights
How This Compares
Modern AI-in-BI tools already follow roughly this shape: a question, an LLM, a declared semantic layer, a generated query, and a written explanation on top. That's not a difference worth pretending doesn't exist. What's provable is what happens on the last mile: the explanation itself.
We never claim to know the internals of a specific competitor's product we haven't verified firsthand. That's not a standard we'd want held to the reverse way either. What's above is what we built and tested, not a claim about what anyone else's tool can't do.
Why Organizations Work With Us
We built the query engine and validated it against its own worked examples before a single model call was ever wired in, so when the agent produces a plan, it's targeting a pipeline already proven correct, not debugging the model and the compiler at the same time. Every trust guarantee in this platform is backed by a real test against real infrastructure: a genuine 25-concurrent-connection race condition proving the quota system can't be gamed, a live model call proving a citation mismatch actually degrades to the raw table, a forced credential in a log line proving it never appears in output.
If a claim about this system can't be demonstrated against real data, live, on request, we don't make it. That standard is slower to build to and it's the only one worth building an AI product on.
Whether you're evaluating an AI-in-BI vendor or want a copilot built on your own data model, we can walk through this architecture live, including the moment it says "I don't know."