Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
SAP CEO Christian Klein’s argument is that enterprise AI becomes useful when it is built into modern business applications, grounded in governed data and allowed to act within real business processes. In a Computer Weekly opinion article published July 10, 2025, he describes a progression from legacy on-premises systems to cloud applications, better-organized data and AI agents that can help complete work. It is a coherent strategy, but also SAP’s commercial point of view—not independent proof that moving to SAP cloud products will produce a particular return.
The practical lesson is broader than any one vendor: AI needs reliable access to business context, clear permissions and a measurable job to do. Cloud can help provide those conditions; it does not create them by itself.
Klein’s argument: three foundations for enterprise AI
Klein’s thesis connects three layers that are often discussed separately:
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →- Modern applications: Replace or simplify fragmented, heavily customized systems that are difficult to maintain and hard to connect.
- Managed, contextual data: Make information current, searchable, deduplicated, governed and connected to the meaning it has in a business process.
- AI inside workflows: Use assistants and agents to interpret that context and help people—or, within defined limits, systems—complete work.
His example is an agent that identifies overdue invoices, investigates why they are late, helps address the exception and supports payment-target attainment. That is an illustration of the intended outcome, not independently verified evidence that an agent can reliably handle invoice resolution end to end in every production environment.
Klein also cites a McKinsey survey to say that more than 80% of organizations had not yet seen tangible profit impact from AI investment. That figure should be understood as a claim attributed to the survey he cites in his article, not as a universal current measurement of every organization or AI project.
Why AI pilots often stop short of business value
A model can answer questions or generate summaries without changing a business outcome. The gap between a successful demonstration and a profitable production process often comes down to the work around the model:
- No business baseline: A pilot may count prompts or time saved on a task without measuring cost, cycle time, error rate, working capital, revenue or risk.
- Weak or inaccessible data: Supplier, order, receipt and payment records may disagree, be stale or be unavailable to the system that needs them.
- Disconnected workflows: An AI tool may identify an issue but lack reliable interfaces or authority to move it to resolution.
- Review and rework: Human checks may remain essential, reducing the apparent time saving. Errors can add further work.
- Implementation costs: Integration, data conversion, security, testing, training and change management can absorb early gains.
- Low adoption: Employees may distrust the output, lack training or find the new process slower than their established workaround.
- Scale and control costs: Usage charges, model operations, monitoring and exception handling can grow as deployment expands.
A useful test is: Does this capability improve a defined process, against a measured baseline, using authoritative data, under accountable ownership and controlled permissions? If those conditions are missing, adding an AI layer is unlikely to fix the underlying process.
Recommended Free Tools
Cloud helps, but is not the prerequisite in itself
Klein presents a move from legacy on-premises software to cloud applications as the first step. That is a plausible route for organizations whose systems are fragmented, difficult to update or poorly exposed to other applications. But the underlying requirement is accessible, governed, well-contextualized systems—not one deployment model.
“Cloud” can mean materially different things:
- Public-cloud SaaS: A vendor-operated service, commonly built around standardized processes and frequent updates.
- Private-cloud ERP: A cloud-delivered environment that can allow broader functional scope and more flexibility for an existing estate.
- Hosted legacy software: A system running in a cloud data centre but retaining much of its old architecture and process complexity.
- Hybrid architecture: A combination of SaaS, private systems, data platforms and specialist applications.
These distinctions matter. Hosting a poorly designed process does not modernize it. Conversely, an organization may be able to use AI with on-premises systems if it has reliable APIs or event interfaces, adequate compute, clean data and effective governance. A data platform or carefully bounded process improvement may create value before a core ERP replacement is justified.
SAP’s own comparison describes Public Edition as more standardized, with SAP-managed operation and more frequent innovation releases, and Private Edition as offering broader scope and greater flexibility. The right choice depends on process fit, existing SAP investment, customizations, regulatory and residency needs, and the organization’s appetite for changing how it works. Product details and release schedules can change, so confirm the applicable edition, country, release and contract.
What “good data” means in practice
Klein’s “magic filing cabinet” analogy points toward data that can be found and understood rather than merely stored. For enterprise AI, that requires more than loading documents into a search index. The organization needs a data foundation that answers questions such as: Which supplier record is authoritative? When was this figure updated? Which business unit owns it? Who may see or change it? What does “overdue” mean in this process?
Useful disciplines include:
- Master-data management: Maintain consistent customer, supplier, product, employee and account identities across systems.
- Data quality and deduplication: Define validation rules, identify conflicting records and assign people responsible for correcting them.
- Catalogs, metadata and lineage: Document what data means, where it came from, how it changed and which reports or decisions depend on it.
- Business semantics: Establish common definitions and relationships, so an AI system does not confuse similarly named fields or processes.
- Identity and access controls: Apply permissions consistently to people, applications and agents, including sensitive or restricted data.
- Retention, privacy and regulation: Set rules for data use, location, retention and deletion that reflect applicable law and sector obligations.
- Transactional and analytical design: Decide which systems record transactions, which support analysis and how updates are reconciled.
For generative AI, retrieval-augmented generation can retrieve relevant enterprise material at answer time and ground a response in it. That does not guarantee correctness: the source may be stale, incomplete or contradictory. A dependable system should expose the records or citations behind its answer, show freshness where possible and say when it lacks enough information instead of silently filling gaps.
SAP positions SAP Business Data Cloud as part of this data layer. That is a vendor description of its offering, not neutral evidence that it will be the best fit for every SAP or non-SAP estate. A well-governed multi-vendor architecture can be integrated; a single-vendor architecture can still contain duplicated data and broken hand-offs.
Assistant, automation or agent? The distinction matters
“AI agent” is used loosely. A buyer should establish what the system can actually do:
- Copilot: Helps a person with a task, typically by drafting, summarizing or suggesting.
- Automation: Executes predefined rules or steps when stated conditions are met.
- AI assistant: Answers questions, summarizes records, recommends actions or helps navigate a system.
- AI agent: Can select and execute multiple steps using tools or enterprise systems, within its permissions and operating constraints.
- Autonomous process: A higher-risk arrangement in which a system can complete consequential work with limited human intervention.
For the overdue-invoice example, the difference between a useful assistant and a risky agent is concrete. Can it only flag an exception? Can it draft a message to a supplier? Can it change a payment date? Can it release a payment? Each additional action raises the stakes and demands stronger controls.
Before an agent acts, define which actions are allowed, which require approval, how permissions are inherited, how exceptions are detected and how a decision can be reconstructed later. Use least privilege, separation of duties, approval thresholds, audit logs and a way to stop or revoke access. Test incomplete or conflicting records, unusual invoices, fraud attempts and malicious instructions embedded in documents. Keep a human accountable for high-impact financial, employment, safety or regulatory decisions.
SAP describes Joule as enabling natural-language interaction with business information and selected navigation or transactional tasks. Its page lists a “price upon request” commercial approach and advertises a 90% faster execution claim for navigation and transactional tasks. Treat that performance figure as a SAP claim, not a guaranteed customer result; establish the relevant task, comparison baseline, edition, release, configuration and human-review time before using it in a business case.
What the SAP product story does—and does not—establish
Klein’s argument aligns with SAP’s current product positioning: SAP describes S/4HANA Cloud Public Edition as subscription-based cloud ERP with embedded AI and coverage spanning areas such as finance, supply chain, HR and sales. SAP also offers Private Edition, SAP Business Technology Platform for integration and extensions, Business Data Cloud and Joule.
That portfolio gives SAP customers a connected route to consider. It does not prove that all capabilities are included in every package, available in every region or release, or automatically integrated in a particular customer’s landscape. Nor does a product feature establish implementation success. Confirm licensing, model and data-processing arrangements, APIs, supported functions, release behavior and contractual terms with SAP and the implementation partner.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Rank #4
The choice between SAP-centered and best-of-breed approaches is also a trade-off, not a universal verdict. A SAP-centered stack may align ERP processes, data and AI more closely and reduce some integration work. It can also increase vendor dependence, switching costs and licensing complexity. Specialist products can provide stronger functionality in a particular domain or more flexibility across providers, but they increase the work of reconciling identities, data, permissions and process context. A single vendor does not guarantee integration; multiple vendors do not make integration impossible.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.A practical route from idea to production
- Choose one bounded process. Favor high-volume, repetitive work with a clear owner and a reasonably stable process, such as invoice-exception triage, case summarization or purchase-order status lookup.
- Record a baseline. Measure current cycle time, labor, error and rework rates, exception volume, service level and relevant financial outcomes. Include review time, not just the AI’s response time.
- Map the process and systems. Inventory the applications, interfaces, custom code, reports, data owners and dependencies involved. Identify where authoritative records live.
- Classify what should change. For each process or application, decide whether to retain, redesign, retire, replace or migrate it. Do not assume every legacy system must be replaced before any AI work can begin.
- Set data and clean-core rules. Profile and cleanse the relevant master and transactional data. Decide which customizations are essential and how extensions will avoid making the ERP core harder to upgrade.
- Design integration and identity. Verify API and event quality, common identifiers, role mapping, access boundaries and audit requirements. Specify what the agent can read, draft, change or commit.
- Pilot with human review. Test normal cases, edge cases, conflicting data, missing data, permission failures and recovery from a bad action. Keep high-impact steps under human approval.
- Test the whole service. Validate data conversion where relevant, security, segregation of duties, performance, integrations, upgrade behavior and business-continuity or rollback plans.
- Train users and assign ownership. Make clear who monitors outcomes, handles exceptions, maintains data quality and can pause the system.
- Measure before expanding. Compare results with the baseline, including adoption, total cost, exception and rework rates, and financial impact. Expand only when the process and controls work reliably.
For an ERP migration, the same sequence needs additional decisions: greenfield redesign, brownfield conversion or selective transformation; data mapping and reconciliation; cutover timing; dual-running costs; and rollback or continuity plans. Migration is a business transformation, not simply a hosting change.
A SAP S/4HANA Cloud Public Edition trial is described as a limited 30-day experience using sample data, with restricted customization, master-data management and SAP BTP integration. It can help someone explore the product, but it is not a production implementation environment or a reliable test of a company’s migration complexity.
Choosing the first use case—and avoiding the wrong one
Good first candidates have repetitive work, a clear baseline, accessible source data, stable ownership and a consequence of error that can be managed with human review. Examples include invoice-exception triage, customer-service case summaries, cash-collection prioritization, delivery-status lookup, supply-chain alerts, finance close-task assistance and comparison of procurement documents.
Do not start by granting an agent unsupervised authority to release high-value payments, make hiring or termination decisions, commit major procurement spend, take safety-critical manufacturing action or submit regulated reporting without sign-off. Those may eventually benefit from automation, but they need a much higher assurance bar than a read-only lookup or draft recommendation.
Questions to ask SAP and an implementation partner
- Which exact product, edition, release, region and features are in scope?
- Which AI capabilities are included, and which require separate licensing or consumption charges?
- Where is data processed and stored? Which models and subprocessors are involved? Is customer data used to train models, and under what terms?
- What APIs, events and business objects can the proposed agent use? How are they secured and versioned?
- Can the system show its source records, data freshness and actions in an audit trail?
- Which actions are read-only, draft-only, approval-gated or fully executable? Who can revoke access or stop the agent?
- How are privacy, residency, retention, segregation of duties and regulatory requirements addressed?
- How will cloud releases affect custom extensions, integrations and AI behavior?
- What is the full five-year cost, including migration, partner services, integrations, testing, training, internal staff, dual running, support and AI usage?
- What measurable baseline and acceptance criteria will determine whether the pilot succeeds?
- What are the cutover, rollback and business-continuity plans if migration or automation fails?
For enterprise ERP, subscription price is only one part of total cost. SAP’s public pricing material describes packages, while final cost depends on package, users, country, contract and scope. Do not generalize a regional package price to another market or treat a software quote as the cost of a completed transformation.
Verdict: the strongest part of Klein’s thesis
Klein is right about the direction of travel: AI is more valuable when it can use reliable business context and participate in an owned process, rather than sitting apart as a generic chatbot. His cloud-data-AI sequence is a useful way to think about the dependencies. It is incomplete if read as “move to cloud, then AI will deliver.” The real prerequisite is governed access to trustworthy data and systems, combined with process redesign, accountable execution, security and a way to prove business value. Cloud ERP may be the right path for some organizations; others can first modernize interfaces and data, choose a hybrid model or select a different platform.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
