Enterprise AI: the use cases that really create value.

Rarely has a technology been adopted so quickly in speeches and so slowly in processes. Every leadership team has seen impressive demonstrations. Many have launched pilots. Very few can point to a complete process whose performance has durably changed. This gap owes little to the maturity of the models; it comes from the way companies choose their use cases. Or, more precisely, from the way they fail to choose them.

Why “AI everywhere” fails

Sprinkling AI over every subject at once invariably produces the same result: a collection of orphan pilots. The mechanisms are always the same.

• The demonstration replaces the process. An assistant that shines in a steering committee and a treatment that keeps its promise over thousands of real operations, with their exceptions, their disputes and their badly scanned documents, are two different objects. The first is obtained in a few weeks. The second requires reworking the process around it.
• Nobody owns the result. When the use case is born in a lab rather than with a process owner, nobody answers for the gap between promise and production. The pilot succeeds, the rollout is left orphaned, the subject fades away.
• The data is discovered too late. Enterprise AI does not reason over the internet, but over your orders, your invoices, your contracts, your contact histories. If they are incomplete or contradictory, the model industrialises confusion. Many AI projects die of a data problem that an honest scoping would have revealed on day one.

What sets a good use case apart

Our reading grid is deliberately austere. Five conditions, all necessary.

• A frequent process that is expensive in human attention. The value of AI is counted in occurrences; be wary of prestigious, low-volume subjects.
• A result measurable in business terms: lead time, error rate, workload avoided, response time to the customer. Defined before the project. Indicators reconstructed after the fact prove nothing.
• Accessible and sufficiently reliable data, or a data-quality workstream explicitly included in the scope.
• A business owner who wants the result enough to accept changing their own process to get it.
• A path to production: integration into the system, exception handling, supervision. Without an industrialisation plan, a use case remains a prototype, whatever its budget.

Replace “where can we put AI?” with “which process do we want to see working differently in a year?”: half the sorting is already done.

On SAP and Salesforce, what makes it to production

The use cases that reach production look alike. We see them cluster around five areas, two of which live mostly in the core business system and two mostly in the customer relationship.

Document extraction

Supplier invoices, purchase orders, bills of lading, customs declarations, identity documents when opening an account: a considerable share of administrative work consists of reading documents to re-key their content into SAP or the CRM. Assisted extraction, matching and posting form the safest ground for enterprise AI. Volumes are high, the result can be verified, and the gains are visible to the teams themselves, which changes everything for adoption. In an economy where many documents still circulate on paper or as phone photos, the quality of capture conditions everything: it is part of the project.

Augmented customer service

On Salesforce Service Cloud, AI summarises a customer’s history before the call, suggests a reply to the adviser, classifies and routes incoming requests, whether they arrive by e-mail, WhatsApp or through the branch. The value is real when the adviser remains in charge of the answer and the system learns from their corrections. It disappears when you try to replace the adviser on requests that commit the customer, a dispute or a refund.

Forecasting anchored in operations

Demand forecasting, anticipating supply delays, cash projections: forecasting creates value when its output feeds a recurring, tooled decision, a supply plan or a cash review, rather than one more dashboard. The decisive criterion has nothing to do with the sophistication of the model. A business action has to change because of it.

Scoring

Lead prioritisation, a subscriber’s churn risk, propensity to take up an offer: scoring in Salesforce Sales Cloud or Marketing Cloud is useful if it is explained, contestable and reviewed. A score nobody can justify ends up ignored by the sales team, or worse, applied indiscriminately to customers it classifies badly. And as soon as it touches individuals, it falls under Law No. 2008-12: declared purpose, information of the persons concerned, right to contest.

Agents

Assistants that answer with the company’s knowledge, agents that prepare or execute transactions under human control, on SAP as on Salesforce: the most promising family is also the most demanding. It requires a strict framework. A delimited scope of action, traceability of every action, human validation of anything that commits the company. An agent without governance is nothing more than an operational risk with a good interface.

Three conditions that enthusiasm cannot replace

None of these families holds without three foundations. Clean data, first: master data shared between the ERP and the CRM, common definitions, governed access. It is the thankless workstream that conditions all the others, and the reason our Data & Analytics offering treats data and its uses together. Human supervision, next: someone looks at what the model produces, corrects it, and their corrections go back to the model. Governance, finally: who decides what AI is allowed to do, who answers for its errors, how the persons concerned are informed. We refuse to see these as brakes on innovation. They are what separates an enterprise capability from a permanent experiment.

What the field has taught us

A few lessons, drawn from our projects without naming any. The confidence threshold of an extraction is tuned with the teams who verify, not in an architect’s office: too low, and they stop trusting; too high, and they re-key everything. A model drifts: what worked on January’s documents degrades on June’s when a supplier changes format, and nobody sees it if nobody measures. Teams adopt what removes a chore and reject what adds one, however small. Finally, the best use case is almost always less spectacular than the one presented to the committee, and it is the one to fund.

Because these foundations live in the core of the system and in the CRM, the AI trajectory cannot be thought out independently of the ERP trajectory and the customer relationship trajectory. A clean SAP S/4HANA core, with harmonised processes and reliable data, and a Salesforce where the customer view is single, remain in our view the best AI investments most companies can make. We explain this in our reading of the ECC 2027 deadline and in our viewpoint on the customer relationship in banking and telecoms. Our AI & automation offering starts from there.