ADA: The Open-Source AI That Shows Its Analysis Calculations

When your AI gives you an answer without explaining why

One of our clients — the head of an SME distribution company in Normandy — got in touch with us after using ChatGPT to analyze his sales data. The result looked clean, well presented, apparently coherent. Except that when he dug deeper, there was no way to know how the tool had handled the missing values in his file. Or why certain stock anomalies didn’t appear in the report.

He had an answer. But no trust.

That’s exactly the problem ADA solves — Autonomous Data Analyst — an open-source data analysis tool that, unlike classic LLMs, shows its calculations. No black box. No result pulled out of thin air. Every step of the analysis is traceable, reproducible, verifiable.

In a context where companies are starting to entrust their business data to AI tools, transparency is no longer optional. It’s an operational requirement.

What classic LLMs don’t tell you

Large language models — GPT-4, Claude, Gemini — are exceptional at generating text, summarizing documents, drafting emails. But when you ask them to analyze numerical data, a structural problem shows up.

These models reason through probabilities, not deterministic calculations. Ask the same question twice about the same CSV file, and you can get two slightly different answers. Ask them how they calculated a growth rate, and the answer will often be vague or circular.

Something agencies rarely tell you: most “data analysis” AI tools actually do text generation about data, not analysis in the statistical sense of the term.

The difference is huge for a business owner making decisions based on those numbers.

Opacity + decision = risk.

ADA breaks with this logic. The tool generates executable Python code to answer your analytical questions. You see the code. You see the execution. You see the result. If something’s off, you can pinpoint exactly where.

ADA in practice: what the tool actually does

ADA is an AI agent specialized in data analysis. It works on a simple principle: instead of answering your question directly in natural language, it writes and executes code to answer it.

You ask a question in English (or French). ADA generates the matching Python code — pandas, matplotlib, scipy depending on what’s needed — runs it in a secure environment, and returns both the result AND the code that produced it.

Concretely, here’s what you can do with it:

On-demand dashboards. Load a data file (sales, stock, HR, finance) and ask for a visual summary. ADA generates the charts along with the source code. You can tweak the parameters, rerun, iterate — the same logic we deploy in our automated AI reporting solutions.

Anomaly detection. “Which rows in this file look off?” ADA applies statistical methods (z-score, IQR, isolation forest) and explains why each point is flagged. Not a magic list — a documented method.

Reproducible reports. Today’s analysis can be rerun on next week’s data with exactly the same processing. That’s what’s called reproducibility — the standard for serious data analysis.

“Trust in an AI result doesn’t come from the result itself, but from the ability to understand how it was obtained.” — a core principle of algorithmic auditing

This isn’t a tool reserved for data scientists. The conversational interface lets a non-developer ask questions in natural language. But unlike classic chatbots, the answer is grounded in verifiable code.

Why transparency is becoming a business issue

Ask yourself this: if your analyst handed you a report without showing you the sources or the method, would you sign off on it?

Yet that’s what most AI analysis tools do today.

This isn’t a theoretical problem. It’s a concrete operational risk.

Take a real case: a company uses an LLM to analyze its customer data and identify high-potential segments. The model produces a segmentation. The sales team adapts its strategy. Six months later, the results aren’t there. There’s no way to know whether the problem came from the strategy, the initial data, or the AI processing.

No traceability, no learning.

With ADA, every analysis is a documented artifact. You know exactly which columns were used, which filters were applied, which statistical method was chosen. If the result is bad, you can diagnose why.

It’s also a solid argument for GDPR compliance. When your analyses involve customer data, being able to explain the processing logic isn’t a luxury — it’s a legal obligation in certain contexts.

How to get started with ADA: a practical guide

ADA is open-source. That means you can install it, audit it, modify it. No dependency on a vendor who changes their pricing or shuts down their API overnight.

Here’s how to approach it if you’re an SME or a freelancer who wants to test it out:

Step 1 — Prepare your data

ADA works with CSV files, Excel, or database connections. Before launching anything, clean your data: name your columns clearly, check the types (dates in date format, amounts as numeric). A clean file means a clean analysis. That said, data that looks clean can still be wrong, as shown in our experience report on a scraper that returned clean but incorrect rows.

Step 2 — Ask precise questions

The classic trap: asking vague questions. “Analyze my sales” won’t give you anything useful. Ask operational questions instead: “How has my revenue evolved month by month over 2024?” or “Which customers haven’t ordered in more than 90 days?”

The more precise the question, the more relevant and reusable the generated code will be.

Step 3 — Read the code, not just the result

This is where ADA changes the rules of the game. Get in the habit of reading the generated Python code, even if you’re not a developer. You don’t need to understand it line by line — just check that the right columns are being used, that the filters match your intent.

Step 4 — Iterate and save

Every validated analysis becomes a reusable template. Built a monthly stock anomaly report? Save the code. Next month, reload your new data and rerun it. Same method, zero manual rework.

This workflow — question → code → validation → template — is what turns a one-off analysis into an industrialized process.

The limits you should know before diving in

I’m not selling a miracle solution. ADA has real constraints.

The setup curve isn’t zero. It’s an open-source tool that you install locally or on a server. If you don’t have someone technical on your team, you’ll need help with the initial setup. This isn’t a plug-and-play SaaS app.

The quality of results depends on the quality of the data. ADA is transparent about its calculations, but it doesn’t fix fundamentally bad data. Garbage in, garbage out — that still holds true even with the best tool in the world.

Interpretation stays human. ADA gives you statistically solid, traceable results. But deciding what to do with those results — that’s still your job. The tool doesn’t replace business judgment.

On the projects we’ve run, the real ROI of a tool like ADA shows up after 2 to 3 months of regular use, once the templates are built and the team has gotten into the habit of asking precise questions. Not in the first week.

“A good AI analysis tool doesn’t tell you what to decide — it gives you the elements to decide with confidence.”

What this actually changes for an SME

After 15 years supporting Normandy businesses with their digital presence, here’s what I observe: most SMEs have data. They just don’t have the tools to make use of it without depending on a service provider or an opaque tool.

ADA changes the equation on three points:

Autonomy. You no longer need to go through an agency for every analysis report. Once the templates are built, your team can rerun the analyses on its own.

Trust. Your decisions are grounded in analyses whose logic you can verify. It’s not “the AI said so” — it’s “here’s the calculation, here’s the result.”

Capitalization. Every analysis becomes a reusable asset. You progressively build a library of analytical processes tailored to your business, in line with our practical guide to automating your repetitive tasks with AI.

Can your current workflow produce an anomaly report on your data in under an hour, without outside technical help?

If the answer is no, that’s exactly where a tool like ADA creates measurable value.

Three takeaways before you take action

Before you close this tab, remember the essentials:

1. Algorithmic transparency is concrete. A tool that shows its calculations isn’t just more “ethical” — it’s more useful. You can debug, improve, audit. An opaque LLM doesn’t offer you that.

2. Open-source means long-term control. No dependency on a vendor, no surprise price hikes, no API that disappears. You control your analytics stack.

3. ROI comes from repeatability. A one-off analysis has little value. What changes everything is turning your recurring questions into automated processes. ADA is built for that.

---

Want to industrialize your data analysis?

At GDM-Pixel, we help SMEs move from “AI that answers stuff” to “AI that produces verifiable results.” Whether it’s connecting ADA to your business data, building your first analytical templates, or integrating this kind of tool into a broader workflow with n8n or Make — we’ve done the groundwork.

No promise of a revolution. Just an honest assessment of what can work in your context.

Contact us for a technical audit — we’ll tell you in 30 minutes whether it’s worth digging deeper.