Enterprise AI in the SAP ecosystem: which use cases are really worth the effort?

There has never been so much artificial intelligence in presentations and so little in transactions. The gap is not a matter of models: the models are good, and they improve faster than companies’ ability to absorb them. The gap lies in what surrounds the model. Without a connection to the management system, without reliable data, without a governance framework, the prototype remains a prototype. The AI that matters reads your real documents, understands your SAP master data and acts within your transactions, under supervision proportionate to the risk. That work is as much engineering as data science, and that is where the gap between a demonstration and a service is played out.

The right question is not “can we?” but “should we?”.

Almost anything can be given an agent. Almost nothing should be given one without prior examination. We use a three-box reading grid, which we apply to every use case before writing a line of code.

Where an agent is relevant

When the situation varies with every case, when the decision requires cross-checking several sources, and when the action remains reversible or supervisable. Exception handling in a supplier flow, incident resolution from a knowledge base, purchasing or replenishment recommendations from heterogeneous signals: these are areas where variability justifies reasoning, and where a human can take back control without harm.

Where automation is enough

When the rule is stable and the volume is high. A workflow, a business rule, an integration on the SAP BTP platform do the job at lower cost, more predictably and in a way that is easier to audit. AI adds nothing there that good configuration does not already do, and it adds a share of uncertainty that nobody asked for. Putting an agent where a rule is enough means paying more to be less sure.

Where nothing should be automated

When the error costs more than the time saved, when accountability must remain with a named person, when the data is not ready. A contractual commitment decision, a customer credit decision, a decision that affects an individual: assistance is possible, delegation is not. And when the data is not ready, we say so, and we suggest starting with it. That is the subject of our point of view on the data foundation.

An AI use case is chosen like an investment: by the process it improves, the measure that proves it, and the risk one agrees to carry.

Documents first, in French and in Arabic.

If we had to start with only one family of use cases, it would be this one. Companies in the region still largely live to the rhythm of documents: supplier invoices, delivery notes, contracts, customs files, administrative correspondence. Many arrive in French, some in Arabic, some in both languages on the same page, often scanned, sometimes photographed. Processing them keeps entire teams busy re-keying, reconciling, checking.

Document AI, integrated into the S/4HANA core, changes the nature of that work. Automatic extraction and reconciliation of invoices with orders and receipts, checking of mandatory information, classification of incoming documents: these are cases where the value is immediate, the measure obvious, and the risk manageable since every discrepancy goes back to a human. The difficulty is not in the model; it is in the real diversity of your documents, in the quality of your supplier master data, and in bilingualism. That is precisely why the Casablanca Centre of Excellence works on documents in French and Arabic, with teams that read both.

Joule, agents and assistants: an order of priority.

The SAP ecosystem now offers its assistant, Joule, pre-built agents by domain, and the building blocks to build others on BTP. The temptation is to activate everything that is available. We recommend the opposite: start from the processes that hurt, select two or three cases per domain, and check for each that the data is accessible, that human control is defined and that the gain is measured.

• In finance: closing assistance, detection of unusual entries, assisted matching of cash receipts. Repetitive tasks with high variability, where the assistant saves time without making the decision.
• In procurement and supply chain: compliance checks of requisitions against the contract, stock-out or supplier-delay alerts with recommended action. The agent is relevant there because every situation differs and the action remains supervisable.
• In customer service: qualification and routing of requests, in French and Arabic, and assisted replies based on history. A case where language is the first difficulty, and where the region has a head start if it gives itself the means.
• In the SAP run itself: incident analysis, assistance with development and code review, monitoring of flows. Internal cases, less visible, but which often fund the next ones.

Prototype, measure, govern, industrialise. In that order.

The method matters as much as the use case. We have stabilised it in four stages, which we refuse to reverse.

The prototype is built on your real documents and data, in a sandbox connected to SAP, on a narrow scope, in a few weeks. A result that can be shown, not an intention. The measure is defined before the prototype and compared with the existing situation: extraction rate, automatic reconciliation rate, time saved, errors avoided. Without a prior measure, every prototype succeeds, and nothing gets decided. Governance sets human supervision proportionate to the risk, traceability of decisions, personal data protection under the law applicable to each entity, ownership of models and prompts. Industrialisation, finally, integrates the case into the S/4HANA core, monitors it, improves it continuously. At that stage, the use case becomes a service, handled by the service centre, and no longer a project.

That is the role of the group’s AI Centre of Excellence in Casablanca: to design, measure, govern and industrialise these use cases for the group’s clients and for companies in the region, with teams that know SAP from the inside and speak the languages of your documents. Those who scope the use case are those who put it into production. We know of no other way to be accountable for it.

What to remember

Before building, examine: agent, classic automation or nothing. Variability, reversibility and data maturity decide.

Where to start

Documents. Invoices, delivery notes, contracts, customs files, in French and Arabic, automatically reconciled in the S/4HANA core with a human on every discrepancy.

The method

Prototype on real data, indicators defined beforehand, governance proportionate to the risk, then industrialisation handled by the service centre.

The reflex

When the data is not ready, say so. And start with the data foundation rather than with a doomed prototype.

To go further: the Casablanca AI Centre of Excellence, our Data & Analytics offering, and the group of which the Centre is one of the three pillars.