Why business intelligence produces figures that contradict each other
Business intelligence is one of the oldest data workstreams, and one that management readily complains about. The stock of reports grows without anyone knowing which ones are still consulted. Two dashboards presented to the same committee display two values for the same indicator, and the meeting is spent trying to work out which one is right.
The reporting tool is rarely to blame. The calculation rule is buried in each report: each author applies their own filters, period and exclusions. The “margin” or the “active customer” therefore exists in as many versions as there are reports. Self-service, opened up without a framework, has accelerated the phenomenon.
A second cause is less visible. Many dashboards were designed from the data available, and not from the decisions to be taken. They provide plenty of information, but do not help to make decisions.
The semantic layer: defining each indicator only once
The semantic layer sits between the platform’s tables (warehouse or lakehouse) and the reporting tools. It describes in business terms the indicators, the analysis dimensions, the relationships between tables and the access rules. An indicator is calculated there in a single place. The reports, connected spreadsheets and assistants that query it all obtain the same value.
Its design calls for three decisions.
Where to place it. In the reporting tool, it is simple to set up but only serves that tool. In an independent layer, it can be reused by several tools and by applications, at the cost of an additional component to operate. The choice depends on the number of tools in service and on the role AI will play in accessing data.
What model to base it on. The star schema, which separates the measured facts from the analysis dimensions, remains the most readable basis for the business and the most predictable in terms of performance. It is prepared in the data platform, not in each report.
Who validates. An indicator only enters the semantic layer with an agreed definition and a business owner. Without this rule, the semantic layer reproduces the disorder it was meant to resolve.
Self-service: three levels of autonomy
Self-service is not an on/off switch. Not all users have the same needs or the same skills, and the framework must reflect this.
| Level | What the user does | What they need to be provided with |
|---|---|---|
| Consultation | Reads, filters and explores certified dashboards | A small amount of documented content, with the source displayed |
| Analysis | Builds their own views from the indicators in the semantic layer | A readable model, training, a personal or team workspace |
| Production | Creates new datasets and new indicators | Prepared data, modelling rules, a review process |
Two mechanisms hold everything together. Certification visibly distinguishes what is authoritative from what is exploratory: a user can tell at a glance whether they can present a figure in committee. The life cycle retires what is no longer used, based on consultation statistics.
AI-augmented analytics: what it brings, what it requires
Three uses stand out: querying data in natural language, automatic explanation of a variation (which segments explain the fall in an indicator) and drafting summary commentary.
The risk lies in the nature of a language model. Placed in front of raw tables, it has to guess the joins, filters and business rules. It then produces a well-worded answer, sometimes wrong, with no error signal. For a figure intended for a decision, this behaviour is not acceptable.
The remedy is to have the model go through the semantic layer. It translates the question into a selection of existing indicators and dimensions; the calculation itself remains deterministic. Four conditions are added: carefully written business descriptions, including synonyms; access rights that remain those of the user; display of the indicators and filters used for each answer; a set of reference questions, with the expected answers, to evaluate the assistant before it is opened up and after each change. These principles are the same as those we apply to generative AI projects.
Our approach: from decisions to data
Our consultants work within your teams; decisions remain yours.
Measuring the existing estate. We inventory reports and their actual use. This survey highlights content to be retired and conflicting indicators.
Starting from the management bodies. For each committee, we identify the decisions to be taken, then the indicators that inform them. The indicator dictionary is agreed at this stage, with the business owners.
Building on a first domain. The semantic layer is implemented on a limited scope. The priority dashboards are rebuilt on it, and the old ones are decommissioned as you go.
Opening up in stages. Self-service is extended level by level, with the corresponding training and a network of lead analysts in the business areas. AI is introduced when the definitions are stable, not before.
The seemingly good ideas we advise against
- Changing tool to solve a definitions problem. The new tool will display the same discrepancies, with better presentation.
- Migrating all reports like for like. A change of tool does not justify carrying over reports that nobody opens any more.
- Opening self-service on raw tables. Each user then recreates their own calculation rule.
- Piling up indicators. A dashboard that displays dozens of them ranks nothing and triggers no decision.
- Treating a data defect in the reporting layer. When the discrepancy comes from the sources and not from the definitions, it falls within data quality and is corrected upstream.
How to know whether your BI serves decision-making
The indicators agreed at the scoping stage concern use and trust, not the volume produced.
- The share of reports actually consulted, and the number of items of content retired.
- The share of consultations that concern certified content.
- The number of management indicators with a single definition and an owner.
- The time taken to obtain a new analysis, and the share of requests handled independently by the business.
- For an analytics assistant: the rate of correct answers on the set of reference questions.
One qualitative sign is worth all the others. When committees stop discussing the validity of the figures and start discussing what to do with them, the BI is fulfilling its role.