No reliable AI without a data foundation: where to start, from headquarters to subsidiaries.

The scenario is familiar. A leadership team launches an AI project on a promising use case. The prototype works on a carefully prepared sample. Then comes scaling up, and with it the real data: customer master records duplicated between head office and subsidiaries, articles coded differently from one country to another, analytical hierarchies that have changed three times without anyone documenting why. The model is not mistaken; it faithfully learns the disorder it is given. And the project stops where trust in the figures had already stopped.

The data foundation: what the term covers.

The term has become so common that it ends up meaning nothing. For us, a data foundation consists of three inseparable layers.

The data model. A shared definition of the objects that matter: the customer, the product, the supplier, the site, the cost centre. Not their technical structure, but their business meaning: what is an active customer, from what point is an order a sale, which product hierarchy is authoritative for group reporting. As long as these questions have several answers depending on who asks them, no tool will reconcile them.

Governance. Named owners for each master data domain, rules for creation and modification, quality controls that escalate to the people accountable and not to an anonymous mailbox. Governance is not one more committee; it is the answer to a simple question: when a piece of data is wrong, who corrects it, and who decides what is right?

The platform. Only then comes the tooling: where the data is consolidated, historised, exposed to analysis and to models. It is the most visible layer and, paradoxically, the least decisive. A good platform on a bad model produces beautiful dashboards that are wrong.

Business Data Cloud, Datasphere, SAC: an architecture, not a catalogue.

The SAP ecosystem now offers a coherent chain for this foundation. SAP Datasphere plays the role of semantic and integration layer: that is where the data model takes shape, where SAP and non-SAP sources are reconciled, and where business definitions become reusable objects. SAP Analytics Cloud carries analysis, planning and reporting. Business Data Cloud, more recent, aims to bring these building blocks together around prepared data products and to open the way to AI use cases on governed data.

Our position is nuanced. These tools are good, and their integration with the S/4HANA core is a real advantage: the semantics of management data come with them, instead of being rebuilt by hand. But no tool replaces the work of definition. We have seen platforms deployed in a few months that remained devoid of meaning because nobody had settled what a customer was. And we have seen groups where most of the value came from work on the model and governance carried out even before choosing the platform. The right order is this one: define, govern, then tool up. The reverse is expensive and disappointing.

A data platform is chosen last. What fills it — the model and the governance — is decided first, and decided at leadership level.

The dashboard is the last step.

This is the most counter-intuitive conviction, and the one we defend most firmly. Data projects almost always start with reporting: a department wants “its” dashboard, and the project is built backwards to fill it. The result is a dashboard that displays figures nobody can explain, disputed from the first meeting where two departments bring two versions of the same indicator.

A dashboard is not a project; it is the visible result of an invisible foundation. When the model is shared, when governance is in place, when the platform exposes reconciled data, the dashboard becomes almost trivial to produce, and, above all, it is believed. The same rule applies to AI: an agent that recommends replenishment based on a stock level that three systems assess differently recommends nothing. It propagates an inconsistency with confidence.

This does not mean waiting years before showing anything. It means that the first visible deliverable of a data programme must be a narrow, entirely reliable scope, rather than a broad and approximate one. One correct indicator, explained, that management actually uses, is worth more than a wall of screens.

From headquarters to subsidiaries: the real testing ground for governance.

For Moroccan and African groups, the question of the data foundation takes a particular form. Head office wants a consolidated view; each subsidiary has its systems, its coding schemes, its local obligations, sometimes its currency and its working language. The temptation is to standardise everything by decree, or to give up and consolidate by hand in spreadsheets. Both fail.

• A common model, with explicit local extensions. The group defines the core, the objects and the hierarchies that must be comparable everywhere. Each subsidiary keeps what is specific to it, in attributes identified as such, and not in misused fields.
• Federated governance. Data owners at head office for the core, in the subsidiaries for local data, and an arbitration body that settles conflicts of definition. Without an arbiter, definitions diverge again within a few months.
• Deployment in waves, subsidiary after subsidiary, starting with the one where the data is most mature and management most committed. The first success serves as a model and as proof for the next ones.
• Compliance with local personal data law, country by country, built into the model from the outset: which data can be consolidated, in what form, and which must stay in its country.

This is work that is carried out from the region, with teams who know the subsidiaries’ systems and the constraints of each country. It is also what makes possible everything that comes next: reliable consolidation, planning, and the AI use cases we describe in our point of view on enterprise AI in the SAP ecosystem. The Casablanca AI Centre of Excellence almost always starts there, in fact: when the data is not ready, we say so, and we start with it.

What to remember

AI fed with inconsistent data produces results that are convincing and wrong. The data foundation comes before any use case.

The order

Define the data model, put governance in place, then tool up with Datasphere, SAP Analytics Cloud and Business Data Cloud. Never the reverse.

The dashboard

Last step, not starting point. A narrow and entirely reliable scope is worth more than a wall of approximate screens.

The subsidiaries

Common core and explicit local extensions, federated governance with an arbiter, deployment in waves, local law respected country by country.

To go further: our Data & Analytics offering, the AI Centre of Excellence that builds on this foundation, and our approach, from scoping to delivery.