Data Analysis with AI

Recorded in Polish. English subtitles are available on the recording via YouTube's auto-translation.

The third webinar in the Witbee series — and this time the lead role goes to AI rather than analytics. We look at the future of data analysis, a future already available today, and at how the individual models cope with real business data.

Report-based versus conversational analytics

Report-based analytics is the kind we all know: I open a report over my morning coffee and see the key numbers. Repeatable structure, a shared language across the organisation, access from a phone, trend visualisation. Its limitation: it does not answer why. I see the numbers and the relationships, but the interpretation is on me — and not everyone can do it.

Conversational analytics answers "why" and "what should we do about it". It lets you ask ad-hoc questions without building a report, draws on many data sources at once, and is not confined to a predefined visualisation. This is genuine democratisation of analytics: I no longer need an analytics degree or ten years of report-reading experience to get an answer in plain language.

Which is better? It depends — the most effective companies use both, because they complement each other. Report-based for daily monitoring, conversational for ad-hoc questions, interpretation, and digging deeper.

The context gap — what AI does not know about you

AI knows the entire internet, but almost nothing about your company. The missing pieces: consistent data from every system, order within that data, access to historical data, table documentation, business context, knowledge of what you are actually doing, and examples of the analyses you expect. That gap is exactly what separates "pasting a table into a chat" from real conversational analytics.

Four steps to AI-driven analytics

1. Data. A data warehouse — BigQuery in my case. But what matters is not merely collecting the data, it is preparing it. Pointing AI at raw, heavy tables (a single Shoper instance yields a dozen or more tables; large implementations produce daily tables measured in terabytes) ends in bills of 7,000–10,000 USD for querying alone — and AI queries repeatedly before it answers. Instead: a handful of light, joined, optimised tables, cheap enough that many companies stay within the free tier.

2. Context. Documentation of tables and metrics — is the "cost" column in zloty or dollars, net or gross, with or without shipping? That distinction changes results enormously. Then: business definitions (a good ROAS is 200 for one company and borderline at 500 for another), instructions and example analyses, industry and product context, goals and KPIs. Without these, recommendations come out generic — fitting everyone, therefore no one.

3. Logging your activity. AI does not sit beside you, cannot see your screen, and is not in your stand-ups. Without a record of what you launched, it will recommend exactly the same thing again two weeks later. And faced with a margin drop it will conclude demand fell — rather than pointing to the pricing change it knows nothing about.

4. MCP. Model Context Protocol connects your analytics infrastructure to AI. Thanks to all the preparation above, AI queries the light data marts according to your instructions, keeps token usage down, and enables repeatable, deterministic analyses tailored to your business.

None of this is rocket science — four steps, and you can talk to your own data.

The duel: Gemini vs Claude vs ChatGPT

First, the market context. Looker Studio Pro introduced the Conversational Analytics idea, but its agent connects to only one table, builds no full context, produces no visualisations of its own, and requires a licence — which is why it sits out the duel. ChatGPT has a code interpreter but no native warehouse integration, leaving the API or MCP. Claude interprets data well and is pushing MCP hardest.

Then the same questions and the same data go to each model — including a request to summarise marketing performance for a given period, deliberately phrased as a very simple prompt from a layperson's perspective. Claude returned an interactive summary: revenue, ad costs, orders, blended ROAS, daily revenue versus cost, average revenue by day of week, top campaigns, and a separate customer tab — and it spotted the wrong year in the question on its own. ChatGPT did not build an interactive page but delivered a dense interpretation with strong insights.

The verdict belongs to the audience in the chat. My own take: I personally like Claude most and consider it the strongest model for reasoning over data, but ChatGPT sometimes produces better substantive recommendations despite the weaker visual packaging. Models shift with every update, so the newest is usually the best.

Repeatability and security

The same prompt can yield different results — yet you want an identical report month over month. The answer is Skills: save the analysis you have refined and replay it on new data.

And one critical point before you put company data into any model: make sure you have turned off model training on your data. In many tools — including paid ChatGPT plans and non-enterprise Gemini — the setting is on by default and must be disabled manually.

Who it is for

Businesses operating online, e-commerce above all. The session assumes your data warehouse is already in place — if it is not, start with the first episode in the series. As I put it at the close: if you do not do this now, your competition will, any day.