What is business intelligence?

First the data has to come together

The basic problem of business intelligence has changed surprisingly little over the past decades.

The interesting information is almost never in one place.

The customer is in the CRM. Their invoices are in the ERP. Costs come from financial accounting. Production data from another system. On top of that there are Excel files, CSV exports, APIs and usually at least one data source that nobody remembers who built in the first place.

In a hospital it doesn’t get any clearer. Patients, cases, diagnoses, procedures, movements, costs, revenues, staff and quality data come from different systems and still have to do with each other.

Business intelligence brings this information together.

For that, data is loaded, checked, standardised and linked. In the past people almost always talked about ETL, that is, extract, transform, load. Today there is ELT, data lakes, lakehouses, streaming and a few more terms that you can know but don’t have to.

The task behind it has stayed the same: a reliable picture has to emerge from many separate data sets.

Because AI doesn’t turn bad data into good data. At best it turns it, faster, into a very convincing-sounding wrong analysis.

Then comes the context

For us, that is the really interesting part.

A company doesn’t consist of tables.

Customers buy products. Products are delivered. Employees look after customers. Invoices belong to orders. Orders arise at particular times and in particular places.

The data maps these relationships.

Classic BI has often built reports from them. Modern BI can go further and make these relationships themselves the subject of analysis.

When I select a customer, I don’t just want to see their master data. I want to see which products they bought, when they bought them, which salesperson looks after them, which complaints there were and what other information is connected.

If I also select a product, this data space changes. Certain values still belong. Others no longer do.

And what doesn’t belong can be extremely interesting, too.

That is the difference between a report and an analysis.

A report answers a question known beforehand. Analysis usually leads to the next one.

And now comes AI

Since generative AI is being built into practically every piece of software, you could get the idea that the whole topic of business intelligence will soon be settled.

You simply ask the AI: why did our revenue in the north fall in the second quarter? And the AI answers.

That is indeed what part of the future of BI will look like.

But behind this seemingly simple question there is a pretty big problem.

What is revenue?

Net or gross? Invoiced or booked? With or without cancellations? By invoice date, service date or payment receipt?

What is the north?

And which second quarter do we actually mean?

People in a company often know from context what is meant. A machine needs that context, too.

That is why a clean semantic foundation doesn’t become unnecessary with AI. It becomes more important.

A modern BI system doesn’t only have to know where a number is stored. It has to know what it means. Which dimensions belong to it? How is a key figure calculated? Which tables are connected? Which time logic applies? Which unit is used?

Only then can an AI work sensibly with company data.

Otherwise we have an extremely eloquent intern with access to the database.

The interface will change

Dashboards won’t disappear because of this.

Some things you simply want to see. Nobody should have to ask anew every morning about revenue trends, utilisation, costs, quality or deviations from plan.

But something new joins the dashboard.

We can talk to data.

Why is this number so high?

Which customer group is responsible?

Show me only cases where this effect also occurs.

What distinguishes these customers from the others?

An AI can translate such questions, run analyses, look for anomalies and offer possible explanations.

It gets interesting, though, when it doesn’t generate an isolated SQL query each time but moves within the same data model as the user.

The person selects something. The AI investigates further. The AI finds an anomaly. The person follows it up.

Business intelligence thus becomes less a pile of finished reports and more a shared workspace for people, data and AI.

This is exactly the direction in which we are developing our own technology.

Who is actually allowed to see what?

With AI, another, rather old BI topic suddenly becomes very topical again: permissions.

A dashboard is relatively easy to control. The user gets certain reports and, within those reports, certain data.

An AI is more curious.

If it is allowed to answer questions on its own, investigate connections and combine data from different areas, it is no longer enough to simply hide a menu item.

The permission has to be attached to the data.

A salesperson may be allowed to see their customers, the sales manager their region and the management the whole company. HR data may only be available to certain roles. In a hospital, particularly protected health data comes on top.

And these rules must also apply when nobody clicks on a report any more but simply asks: show me all employees with an unusual number of sick days.

Technically an AI may be able to answer that question. That is far from meaning it is allowed to.

That is why, for us, permissions belong in the data model itself. They must apply equally to dashboards, free analysis, APIs and AI.

The AI gets no special rights just because it is politely worded.

Business intelligence is actually getting more interesting again

We are therefore working on a BI architecture that brings together a few things that are often considered separately today.

Data is loaded from different sources and linked. Their relationships are preserved and can be explored associatively. Key figures and dimensions get a clear business meaning. Permissions become part of this model.

And on top of that, different kinds of use can emerge.

A classic dashboard.

A free analysis.

An application.

An API.

Or an AI that understands the model and works with it within the rules that apply to the respective user.

The AI is not supposed to simply invent answers that sound somehow plausible. It should be able to trace which data and connections led to a statement.

Because at the latest when someone spends money on the basis of an analysis, changes a process or makes a medical, economic or organisational decision, one question becomes surprisingly important again: how did we actually arrive at this?

That was already a pretty good question in Deming’s day. With AI it doesn’t get worse.