What is data analysis and Business Intelligence? Dashboards, KPIs and fact-based decisions

1. What is data analysis and how does it differ from Business Intelligence?

⚡ In one sentence

Data analysis is the process of turning raw data into useful information, and Business Intelligence (BI) is the set of tools and practices that deliver that information to decision-makers in time.

💡 In plain terms

Imagine a trading company that sells products across several European markets. At the end of the month, the head of sales opens a spreadsheet with thousands of rows — orders, returns, revenue, shipping costs. The data is there. But to answer the question "which market brings us the highest profit once logistics costs are taken into account?", it has to be filtered, combined, margins calculated, periods compared. That takes hours — and it's never quite clear whether the result is correct.

Data analysis is exactly that process: from raw numbers to an answer to a business question. It starts with collecting data (from sales, warehouse and accounting systems), moves through cleaning it (removing duplicates, filling gaps) and ends with drawing conclusions — ideally ones you can act on.

Business Intelligence (BI) is a step further. BI is not a single tool — it's an approach in which data analysis is organized systematically: data from various sources is gathered in one place, processed and presented as dashboards, reports and alerts. Thanks to BI, the head of sales doesn't have to build a spreadsheet by hand every month. They open a dashboard and see the answer to their question — updated, broken down by market, product and period.

The difference? Data analysis is an activity — I can do it once, ad hoc, in a spreadsheet. BI is a system — it organizes the analysis so that answers are available continuously, without repeating the same work.

🔧 Deep dive

The term "Business Intelligence" first appeared back in 1865 — Richard Millar Devens used it to describe a banker who gathered market information faster than his competitors. The modern understanding of BI was shaped by Howard Dresner of the Gartner Group, who in 1989 defined BI as the concepts and methods that support fact-based business decision-making.

In practice, BI spans several layers:

• Data sources — a company's operational systems: ERP, CRM, warehouse systems, e-commerce, spreadsheets. Data can also come from the automatic processing of documents — invoices, contracts and correspondence turned into structured data. Without integrating these sources, BI has nothing to draw on.
• Storage layer — a data warehouse or data lake, where data from different systems is gathered in one place and standardized. The process of feeding the warehouse is called ETL (Extract, Transform, Load) — extracting data from the sources, transforming it (cleaning, unifying formats) and loading it into the warehouse.
• Analytical layer — the tools that process the data and create models, aggregations and calculations.
• Presentation layer — dashboards, reports and alerts that deliver information to end users.

BI doesn't replace thinking. It delivers information that lets you verify hypotheses faster, spot trends earlier and react to problems before they become a crisis.

2. What is a dashboard and how do you choose KPIs?

⚡ In one sentence

A dashboard is a panel that presents a company's key indicators (KPIs) in a single view — so that a decision-maker can see the state of things without opening five different systems.

💡 In plain terms

Back to the trading company with several markets. The head of sales needs answers to three questions: how much did we sell this month, which market is declining, and are we on plan? Without a dashboard they have to open the sales system, export the data, open a spreadsheet, calculate totals, compare against plan. With a dashboard — they open a single screen and see everything: current revenue vs plan, a breakdown by market, the monthly trend.

A dashboard is not decoration or a presentation for the board. It's a working tool — like the instrument panel in a car. The driver doesn't have to open the hood to check the engine temperature. A dashboard in a company works the same way: it shows the state of processes on an ongoing basis, without having to "look under the hood" of each system separately.

But a dashboard is only as good as the indicators placed on it. That's why the key lies in KPIs (Key Performance Indicators) — not twenty on a single screen, but three to five, chosen to match the questions a decision-maker actually needs answered. Choosing KPIs starts with the question "what decisions do we make and what information do we need for them?", not with whatever happens to be in the systems. If on-time delivery is critical — we measure OTD (On Time Delivery). If customer profitability matters — margin and CLV (Customer Lifetime Value). An indicator that doesn't lead to a specific action is decoration, not a KPI (see What is process optimization? — the section on KPIs).

🔧 Deep dive

The BI literature distinguishes several types of dashboard:

• Operational — monitors processes in real time or near real time. Example: a warehouse dashboard showing current stock levels, orders in progress, delays. User: the operations manager.
• Tactical — analyzes trends over weeks or months. Example: a sales dashboard with a comparison against plan, conversion, revenue structure. User: a department head or director.
• Strategic — aggregated indicators for the board. Example: product-line profitability, market share, progress against annual targets.

Designing a dashboard is a discipline of its own. The key principles:

• One dashboard, one question (or a few closely related ones). A dashboard that tries to answer everything answers nothing.
• Visual hierarchy — the most important indicator is visible immediately. Details are available on interaction (drill-down).
• Context — a number on its own, without comparison, is worthless. Is revenue of €500,000 a lot or a little? It depends on the plan, on the previous month, on the same period a year earlier. A dashboard always shows an indicator in context: vs plan, vs previous period, vs target.
• Refresh frequency — matched to the type of decision. An operational dashboard refreshed once a day loses its point, because operational decisions are made on the fly. A strategic dashboard refreshed every minute just generates noise.

KPIs are not a one-off choice. As the company changes, so do the questions it needs answered — and the indicators along with them. A regular review of dashboards (every quarter, every six months) lets you remove indicators no one looks at and add the ones that have become relevant.

3. The three levels of analytics — descriptive, diagnostic, predictive

⚡ In one sentence

Descriptive analytics tells you what happened, diagnostic analytics explains why, and predictive analytics forecasts what is likely to happen.

💡 In plain terms

The trading company from our example notices on its dashboard that sales in one of its markets dropped by 18% over the quarter. That's descriptive analytics — it answers the question "what happened?".

The next step: why did it drop? Maybe a competitor cut prices. Maybe the exchange rate moved and the products became more expensive. Maybe a supplier delayed deliveries and customers switched to the competition. To establish this, you have to combine sales data with data on competitor prices, exchange rates and delivery timeliness. That's diagnostic analytics — it looks for causes.

The third level: what will happen in the future? If the trend holds at the current exchange rate and with the offer unchanged — what will sales be in three months? Is it worth increasing the stock of product X because the data points to rising demand? That's predictive analytics — it uses Machine Learning and statistical methods to forecast based on patterns in historical data.

Most companies operate at the descriptive level: reports, summaries, breakdowns. That is already value — because it turns data into information. But every further level opens up new possibilities: not just knowing what happened, but understanding the causes and reacting before the problem deepens.

🔧 Deep dive

In business practice, predictive analytics most often means:

• Demand forecasting — how many units of product X will the company sell next month, quarter or season? ML models analyze historical sales data, seasonality and market trends and generate a forecast. The result: better inventory planning, fewer surpluses and shortages.
• Customer scoring — which customer is most likely to leave (churn prediction)? Which has the greatest purchasing potential? Models analyze purchase history, contact frequency, time since the last transaction.
• Price optimization — what price will maximize revenue given the elasticity of demand?

Predictive analytics doesn't give certainty — it gives probability. A demand-forecasting model will say: "with 80% probability, sales of product X in June will be between 1,200 and 1,500 units." That isn't a "yes/no" answer — it's a range that lets you make a better decision than intuition based on last year's experience.

The condition: data. Predictive analytics requires historical data — organized, consistent and in sufficient quantity. If a company doesn't measure delivery timeliness, it won't build a model that forecasts delays. That's why descriptive analytics isn't a "lesser" level — it's the foundation. Without reliable descriptive data there's no material for diagnostics and forecasts. This applies to companies in every market — in Poland just as in Germany or Scandinavia — regardless of industry and scale.

There's also a fourth level — prescriptive analytics — which not only forecasts but recommends actions. Example: "given the current demand trend and logistics costs, the optimal stock level of product X at the central warehouse is 800 units, and the allocation to market A should be 55%." In practice it requires advanced optimization models and is used by companies with mature data infrastructure.

4. Case study: optimized goods distribution at a B2C company

⚡ In one sentence

A multi-market trading company replaced manual inventory planning with an ML model that forecasts demand and tests different scenarios before a decision is made.

💡 In plain terms

A company selling products across many European markets had a problem familiar to anyone who manages inventory at scale: how much to order, where to send it and when?

Manual planning — based on spreadsheets, the experience of planners and order history — worked while the scale was small. As markets multiplied, as seasonality looked different in each of them, as warehousing and logistics costs began to rise — manual planning could no longer keep up. Too many variables, too little time, too much risk of error.

The solution combined three elements:

• A predictive ML model — forecast demand for specific products and markets, taking into account seasonality, trends and historical data.
• Stochastic optimization — an algorithm that doesn't look for a single "ideal" plan but accounts for uncertainty: what if demand is higher than the forecast? What if a supplier delays a delivery? The result is a plan that minimizes risk at an acceptable cost.
• Scenario simulation — before a decision is made, the system tests variants: what happens if we increase stock by 20%? What if we reduce the number of warehouses? This makes it possible to assess the consequences of a decision before it's made.

The effect: inventory planning moved from manual work based on intuition to a systematic, data-driven approach — with a measurable improvement in forecast accuracy and a reduction in the costs tied to surplus or shortage of goods.

"Before we start building a model, I always ask the same question: what data do you have and what state is it in? Because even the best algorithm won't extract value from data that is incomplete, inconsistent or scattered across five spreadsheets. In this project, the first few weeks went not into modeling but into organizing the data — unifying formats, filling gaps, connecting sources. Only then did the model start producing forecasts you could trust."

🔧 Deep dive

The project brought together three technical areas:

• Data analysis and BI (this article) — sales data, logistics costs and stock levels as the source for forecasting.
• Machine Learning (see What is AI?) — a predictive model that identifies patterns in historical data and generates forecasts.
• Process optimization (see What is process optimization?) — the result of the analysis translates into an operational change (better inventory planning).

Stochastic optimization differs from ordinary optimization in that it accounts for the randomness and uncertainty of the input data. Classical optimization assumes the data is certain (demand = 1,000 units). In stochastic optimization, demand is a random variable with a probability distribution (demand = 900–1,100 with a median of 1,000). The result is not a single point but a strategy that's robust to fluctuations.

Scenario analysis is a technique in which the model is run repeatedly under different assumptions to assess the range of possible outcomes. In a distribution context: what happens to costs if demand in market A rises by 15% while demand in market B falls by 10%? The result is not a single answer but a map of risks and opportunities.

5. CRISP-DM — how to run an analytics project step by step

⚡ In one sentence

CRISP-DM is a six-phase methodology for analytics and data-mining projects that starts with understanding the business problem — not with the data or the tools.

💡 In plain terms

Companies that decide on an analytics project — whether building a dashboard or a demand-forecasting model — often start with the question "which tool should we buy?". CRISP-DM (Cross-Industry Standard Process for Data Mining) reverses that order: start with the problem, then check the data, then build the solution.

The six phases:

• Business understanding — what problem are we trying to solve? What decision should the analysis support? Without this phase it's easy to build a dashboard that looks good but doesn't answer any question that actually comes up when decisions are made.
• Data understanding — what data do we have? Where is it stored? What state is it in (completeness, quality, format)?
• Data preparation — cleaning, combining sources, filling gaps, standardization. In practice this phase consumes the clear majority of the time in an analytics project.
• Modeling — building a model (statistical, ML or a simple aggregation) that answers the business question.
• Evaluation — does the model answer the question we set at the start? Is the result credible? Does the decision-maker understand and trust the results?
• Deployment — going live in the production environment: a dashboard, an automated report, a model fed with current data.

This approach isn't linear — after the modeling phase you can go back to the data (because it turned out a key variable was missing). After evaluation — back to business understanding (because the question was poorly framed). CRISP-DM is cyclical: a deployed solution generates new questions that kick off the next cycle.

🔧 Deep dive

CRISP-DM was developed in 1996 and published in 1999 as an open standard, the result of collaboration by a consortium of companies (including NCR/Teradata, Daimler and OHRA) within the European ESPRIT program. Although it has not officially been updated beyond version 1.0, it remains the most widely used methodology for data-mining and analytics projects.

In practice CRISP-DM guards against two situations that regularly end in wasted time and money:

• A company buys a tool and then goes looking for something to display on it. The result: an expensive BI platform with charts glowing away that no one uses, because they don't answer the questions that actually come up in management. CRISP-DM forces the reverse order: the business question first, then the data, then the tool.
• An analysis gets produced but never makes it into daily work. An analyst prepares a model or a report, presents the results — and that's where it ends. No one acts on the conclusions, because no one planned how and by whom they'd be used. CRISP-DM requires planning deployment from the very start of the project — so that the result of the analysis reaches the place where decisions are made.

For a company just starting out with analytics, CRISP-DM can be boiled down to three questions: what decision do we want to support with data? Does that data exist in the company? Who will use the result, and how?

6. Where to start with data analysis in your company?

⚡ In one sentence

Start with one business decision that today rests on intuition — and check whether the data you already have can support it.

💡 In plain terms

Data analysis doesn't require a BI platform, a data warehouse or a team of analysts to get started. It requires one decision, one question and data the company most likely already has.

Step 1 — Choose one decision. Which decision costs you the most time, stress or money? How much stock to order? Which customers to prioritize? Whether a marketing campaign pays off? One decision, not ten.

Step 2 — Check whether you have the data. To make that decision you need information. Does that information exist somewhere in the company — in the sales system, in spreadsheets, in email? If so, you have a starting point. If not — start by measuring: pick one indicator and begin recording it.

Step 3 — Build the first view. Not a dashboard right away. Maybe a single chart in a spreadsheet that puts sales against costs and shows the trend is enough. If that one view changes how you make a decision — you have proof that data analysis works in your company. That's the foundation for building something bigger.

Step 4 — Integrate the sources. Once the first view works, it gains value if you feed it data from other systems. Sales + logistics costs + delivery timeliness → a fuller picture. This is where systems integration and BI tools come in to automate feeding the dashboard. Data can come from various sources — including Computer Vision systems (defect frequency, defect types) or automated processes (lead times, the number of cases handled).

Step 5 — Consider forecasting. Once historical data is organized and the dashboards answer descriptive and diagnostic questions — the company is ready for predictive analytics. An ML model forecasting demand, customer scoring, price optimization — each of these requires data the company already has by this stage.

"Working with companies in Poland and Germany, we see that the data usually already exists in their systems — what's missing is the connection between it and the everyday choices. Our role is to design that connection: choose the indicator, build the view and make sure the decision-maker actually reaches for it — every day, not once a quarter."

🔧 Deep dive

At larger scale it's worth considering BI platforms that simplify creating dashboards and automate data feeds. The market offers both cloud and open-source solutions. The choice of tool should depend on the scale of the data, the existing IT infrastructure and the team's skills — not on the tool's popularity.

For companies in the European Union — especially in Poland and Germany — where analytical data is stored matters. Cloud platforms offer flexibility and scalability, but they require verifying GDPR compliance: server location, DPAs with providers, access control. For sensitive data (customers' personal data, financial data), on-premise solutions or a private cloud can be an alternative (see What is RAG and an AI agent? — the section on security).

Data analysis is a capability, not a one-off project. A company that builds a dashboard once and never updates it will quickly fall back on intuition-based decisions. That's why it's worth planning from the start: who will be responsible for maintaining and developing the dashboards? Who verifies that the indicators are current and relevant? Who decides on new analyses? In practice, it works best to designate one person responsible for the "data culture" in the company — even if it isn't their only job.

Frequently asked questions (FAQ)

Want to know which data in your company could work harder? Let's talk — we'll help you identify the question, choose the indicators and design your first dashboard.

Related articles in the cm-opti Knowledge base

• What is Artificial Intelligence?
• What is process optimization?
• What is automation?
• What is OCR, NLP and how does AI read documents?
• What is RAG and an AI agent?
• What is Computer Vision?
• What is systems integration?

Concepts explained in this article → Glossary

data analysis, Business Intelligence (BI), dashboard, KPI, ROI, data warehouse, ETL, descriptive analytics, diagnostic analytics, predictive analytics, CRISP-DM, demand forecasting, churn prediction, stochastic optimization, Machine Learning

Sources and references

• The term "Business Intelligence" — Richard Millar Devens, 1865. Modern definition: Howard Dresner, Gartner Group, 1989.
• CRISP-DM — Cross-Industry Standard Process for Data Mining, published 1999, consortium of NCR/Teradata, Daimler, OHRA.