BI & decision intelligence modernisation

Context

Leadership reporting was produced monthly, by hand, by a small number of people who had become the only ones who knew how. The pack was good. It was also five days old the moment it landed, and it existed nowhere except in the workbooks that produced it.

The business problem

Two problems, and only one of them was visible. The visible one was speed: a five-day cycle meant the leadership meeting discussed a month that had already ended.

The invisible one was worse. Different reports defined the same metric differently — not maliciously, just accumulated over years of one-off requests. Meetings opened by establishing whose number was right. Once that happens regularly, the reporting has stopped being an input to decisions and started being a negotiation.

Constraints

• Definitions genuinely differed between business units in some cases for legitimate operational reasons — not everything could or should be forced into one shape.
• The existing pack had real authority; replacing it with something leadership trusted less would have been a net loss.
• No additional analyst headcount was available, so the new model had to reduce work rather than add a maintenance burden.
• Source data quality varied by unit, and the modernisation could not wait for all of it to be fixed.

Approach

The work started with the KPI dictionary, not the dashboard. Every metric in the pack was inventoried and taken through a definition workshop with the people who argued about it: one owner, one calculation, one stated grain, and an explicit note where a business unit legitimately needed a local variant. Writing down the variant is what stops it from quietly becoming a contradiction.

Only then was the semantic model built, so every report reads the same measures instead of each one re-implementing them. Transformations were version-controlled and tested, so a definition change is a reviewable commit rather than an edit somebody made in a workbook on a Friday.

Dashboards were then built against named decisions — what someone is deciding, how often, and what would change their mind — rather than against the set of fields that happened to be available. The last step was the one that is easiest to skip and most important: actually retiring the spreadsheet. A dashboard that runs in parallel with the old pack has not replaced anything.

What I did

• Ran the metric inventory and the definition workshops, and authored the KPI dictionary.
• Designed and built the semantic and data models underneath the reporting layer.
• Built the executive and operational dashboards, each tied to a named decision and cadence.
• Set up hourly and near-real-time refresh where the decision cadence justified it, and left the rest daily — refresh frequency is a cost, not a virtue.
• Ran the adoption phase through to the point where the manual pack was switched off.

Systems and technologies

Power BI, Tableau, Looker Studio and SAP BusinessObjects on the presentation side; dbt, SQL Server and SAP HANA for modelling and transformation; Git for version control of the models and definitions.