Reference blueprints for AI in SAP analytics

The SAP AI stack operates across four integrated layers, each serving a distinct purpose in your analytics architecture:

Where the stack stands (November 2026): The generative AI hub in SAP AI Core offers foundation models from OpenAI, Anthropic, Google, Mistral AI, Amazon, Meta, Cohere and others, plus SAP's own models. SAP-RPT-1.6, SAP's relational pretrained transformer for predictions on tabular business data, is generally available in the AI Core model library, and SAP-RPT-1 retires by December 31, 2026; SAP's efficiency comparison with general-purpose LLMs (about 50,000× less energy and up to 3.5× better predictions) was published for SAP-RPT-1 and is SAP's own claim. Joule Studio's skill builder has been generally available since July 2025 and its agent builder since December 2025; the new Joule Studio announced at Sapphire 2026, with more than 70 AI models to choose from, has been rolling out to general availability in phases since October 2026, alongside Joule Work. Agents built in Joule Studio support the Agent2Agent (A2A) protocol, and Joule's bi-directional A2A capabilities were planned for general availability in Q4 2026. Model Context Protocol (MCP) servers give agents a standard way to reach tools—Joule Studio can generate one from an existing SAP OData or REST API—and SAP AI Agent Hub keeps agents and MCP servers registered and verified. For data and models that must stay in Europe, SAP unveiled EU AI Cloud in November 2025.

AI in analytics is not a single thing. It spans conversational interfaces, predictive planning, data-driven automation, and specialized model orchestration. Organizations often deploy AI capabilities in isolation—a chatbot here, a forecasting tool there—without a coherent architectural plan. This creates semantic drift, duplicated data pipelines, access control nightmares, and weak ROI.

The reference blueprints in this guide anchor AI placement to specific architectural layers within SAP's modern analytics stack. Each blueprint describes:

• When to choose this pattern (business context and maturity prerequisites)
• Which SAP and partner components participate (semantic layers, platforms, runtimes)
• What AI capabilities emerge (assisted analytics, forecasting, data orchestration, custom ML)
• How governance, access control, and data stewardship remain intact
• Prerequisite conditions and readiness indicators

By aligning your AI initiatives to one or more of these blueprints, you gain clarity on data flow, responsibility boundaries, cost allocation, and risk management. You also preserve the semantic trust and governed business definitions that make analytics defensible and auditable.

When to use this blueprint

Choose this blueprint when your primary AI objective is accelerating business analyst and decision-maker productivity. This includes:

• Interactive "just ask" queries against trusted semantic models
• AI explanations of answers and AI-assisted drafts of comment summaries
• Predictive baselines and AI-assisted planning logic
• Natural-language help with calculations and story scripts
• Self-service exploration tied to governed dimensions

This blueprint is ideal for organizations that have already invested in governed SAC or Datasphere models and are seeking to democratize analytics without exposing raw data or allowing uncontrolled prompting.

Architecture and components

The blueprint consists of four integrated layers:

Capability mapping

This blueprint unlocks the following AI-powered capabilities:

• Just ask. Users ask questions in natural language—for example, the top ten products by gross margin last quarter, or the change in revenue versus last year by region—and get a chart or table from an indexed SAC or Datasphere model. Supported question types include aggregations, filters (including fiscal and dynamic time), rankings, percentage of total, CAGR, variance, and hierarchy drill-down; calculations between measures, forecasts, and multi-step questions are out of scope. Generated AI explanations can add a short narrative to an answer.
• Smart Insights and AI explanations. Smart Insights runs statistical analysis on a selected value—change over time, top contributors, or how a calculation is built—on acquired, live SAP HANA, Datasphere, and seamless planning models. It does not write commentary; AI-assisted features can draft comment summaries for a person to review.
• Planning assistance. Smart Predict writes time-series forecasts into planning versions as a baseline, AI-assisted features generate advanced-formula scripts and documentation for data actions from natural language, and the AI prompt insights widget analyzes planning data. SAP Enterprise Planning, announced at Sapphire 2026, adds Joule agents to the planning cycle; its general availability was planned for Q3 2026, so confirm the status for your tenant.
• Calculations and scripts in plain language. AI-assisted features create calculations in data analyzer from a natural-language request and generate or explain story scripts, which cuts documentation lookups for builders. Review what they produce before you publish it.
• Agent-driven workflows. Custom agents built in Joule Studio can chain steps—gather variance data, call a tool, draft a summary for review—with human-in-the-loop steps where you need approval. You build and govern these agents yourself. At Sapphire 2026 SAP also announced AI story generation in SAC, which was in private preview at the time.

Prerequisites and readiness

This blueprint requires foundation work before Joule becomes effective:

• Governed semantic models. Datasphere models must be complete, well-documented, and actively maintained. Incomplete or inconsistent models lead to poor answers.
• Defined business entities. Master data—customers, products, organizational units—must be curated and linked within Datasphere so that Joule understands the business structure.
• Model selection and access rights. Add only the models that suit natural-language questions to just ask, define synonyms for business terms, run the model readiness check, and set model access rights in SAC and Datasphere before you switch it on.
• Quality baseline. Start with high-confidence models and datasets. Joule's value increases as underlying data quality improves.
• User acceptance readiness. Business users need onboarding on how to ask questions and when to trust AI-generated narratives. Skipping this leads to low adoption.

Architecture pattern

Data flows from operational or warehouse systems into Datasphere, where it is modeled as dimensions, facts, and calculated measures. SAP Analytics Cloud connects to these models, providing traditional analytics interfaces. Administrators add selected models to just ask, which indexes them; a natural-language question is matched to the model's measures, dimensions, and synonyms and answered with a chart or table, and analytical insights in Joule uses the same mechanism. Custom Joule Studio agents can chain several steps and call tools. Because answers come from governed models rather than raw tables, definitions stay consistent—provided model access rights are set correctly in SAC and Datasphere, which is why SAP's rollout guidance puts that step before switching just ask on. SAP also states that just ask does not store users' questions and that tenant data stays isolated (KBA 3574723).

When to use this blueprint

Choose this blueprint when your AI initiative centers on demand forecasting, supply planning, financial forecasting, or scenario modeling. This includes:

• Time-series forecasting of demand, sales, or inventory
• Driver-based planning that correlates business drivers to outcomes
• What-if simulation and sensitivity analysis
• Variance and driver analysis on plan and actuals
• Collaborative scenario governance and approval workflows

This blueprint is optimal for supply chain, finance, and revenue planning teams that need predictive power without building custom data science infrastructure.

Architecture and components

The blueprint integrates planning, analytics, and AI services:

Capability mapping

This blueprint enables:

• Demand forecasting. Smart Predict time-series models learn trend and seasonality from history and can use influencers such as promotions or prices—provided you also have their future values. Forecast accuracy is reported per entity, so you can see where the model is weak.
• Driver-based planning. Value driver trees model the driver relationships you define and let planners change a driver and see the effect. Smart Predict regression or time-series models with influencers show how much each candidate driver contributes, which helps you decide which drivers belong in the plan.
• What-if simulation. Value driver trees and data actions recalculate dependent figures when assumptions change. SAC Compass runs Monte Carlo simulations—statistical rather than AI—so planners see a range of outcomes and how likely they are; it needs an SAC for Planning standard or professional license.
• Variance analysis. Smart Insights shows the top contributors to a value and how it changed over time, which helps planners see where actuals moved away from plan.
• Forecast refresh and accuracy tracking. Schedule the multi action that retrains the model after each actuals load, and compare the plan with the updated forecast to monitor plan attainment.

Prerequisites and readiness

Predictive planning requires clean, consistent historical data and mature planning discipline:

• Historical data quality. Enough clean, reconciled history at the forecast grain (weekly, monthly): SAP's guidance is about five times the forecast horizon—60 periods of history for a 12-period forecast—and you need at least two full seasonal cycles if seasonality matters. Gaps, outliers, or corrections must be documented.
• Planning model maturity. The underlying planning model should be stable, with clear input assumptions, validated formulas, and documented business logic.
• Statistical baseline. Identify the baseline forecasting method currently in use (e.g., simple average, exponential smoothing, regression). AI models should be benchmarked against this baseline to prove lift.
• Business driver clarity. For driver-based planning, identify the key business drivers that influence the forecast and ensure they are captured in data and accessible to the model.
• Governance and approval workflow. Define who can build scenarios, who reviews forecasts, and who approves them for use in downstream planning or financial reporting.

Architecture pattern

Historical operational and planning data lands in Datasphere, where it is enriched with business context and driver definitions; with seamless planning, the planning model's own data can live there too. SAC planning models consume these sources and establish baseline planning structures. Smart Predict trains time-series models on the history—optionally segmented by entity—and predictive planning writes the forecasts into a planning version as the baseline that planners adjust with business judgment. Value driver trees, data actions, and Compass handle what-if work, and a scheduled multi action refreshes the forecast after each actuals load. Data audit logs changes to public planning data (time, user, old and new values), data locking controls who can still change a version, and calendar processes route plans to reviewers who approve or reject them.

When to use this blueprint

Choose this blueprint when your organization needs:

• Cross-domain data unification (SAP and non-SAP sources)
• Curated, governed data products for analytics and AI consumption
• Managed semantic enrichment without building custom pipelines
• Secure, cataloged data sharing across teams and business units
• A managed platform that scales from analytics to AI applications

This blueprint is ideal for large, multi-domain organizations (e.g., global enterprises with finance, supply chain, sales, and HR) that need a unified, governed data foundation for both analytics and AI applications.

Architecture and components

Capability mapping

This blueprint unlocks:

• Data Product Creation. Teams curate and publish data products—e.g., "Customer 360," "Product Catalog," "Supply Chain Network." Each product is versioned, documented, and backed by defined ownership.
• Semantic Enrichment. Raw data products are enriched with business logic, derived measures, and contextual relationships within Datasphere, making them immediately useful for analytics and AI.
• Governed Sharing. Data products are shared across teams with fine-grained access control. Consumers can trust the quality and governance of shared data without replicating or transforming it locally.
• Lineage. Datasphere's impact and lineage analysis and the catalog show how objects depend on each other, which supports audit, impact analysis, and data quality investigations.
• AI Model Training Foundation. Clean, semantically enriched data products reduce the data engineering overhead for ML teams. Models can be trained on trusted, documented data with known quality characteristics.

Prerequisites and readiness

This blueprint requires organizational commitment to data stewardship:

• Data product ownership. Each major data product must have a clear owner who is accountable for quality, documentation, and refresh schedules.
• Catalog maturity. Invest in tagging, documentation, and discovery. A data catalog without governance is a false start.
• Cross-team data governance. Define policies for data access, PII handling, and sharing approval. Governance rules must be enforceable in Datasphere and BDC.
• Integration readiness. Plan for connectors from all major source systems (ERP, CRM, data warehouse, third-party datasets). Each integration must have a defined owner and refresh SLA.
• Skills readiness. Teams need to understand semantic modeling, data product design, and governed sharing practices. Training and change management are essential.

Architecture pattern

SAP-managed data products flow from SAP applications into Business Data Cloud, and other sources—cloud apps, data warehouses, third-party providers—arrive through Datasphere replication flows, federation, or BDC Connect. Within Datasphere, business analysts and data stewards enrich products with semantics, calculations, and business context. These enriched models are then available to SAP Analytics Cloud, embedded AI services (Joule, Smart Predict), and external consumers via APIs. Data access controls defined in Datasphere apply to consumers that query through Datasphere, such as SAC; when you share a data product to Databricks, Snowflake, or BigQuery, access on that platform has to be governed there too. For organizations on Databricks, the same products can be used in notebooks and ML frameworks through SAP Databricks or BDC Connect, without copying the data.

When to use this blueprint

Choose this blueprint when you need:

• Custom AI model development and training
• Large language model (LLM) orchestration and grounding via the generative AI hub
• Retrieval-augmented generation (RAG) for business-specific knowledge
• Agent-based AI patterns with multi-step reasoning and external tool access (MCP)
• MLOps and model lifecycle management
• Non-standard inference infrastructure or frameworks

This blueprint is for organizations with ML engineering capabilities and the need to go beyond built-in analytics AI. It is often used in parallel with other blueprints—for example, to build specialized recommendation engines, anomaly detection systems, or natural language understanding layers that feed into SAP Analytics Cloud.

Architecture and components

Capability mapping

This blueprint enables:

• Multi-model LLM orchestration. The generative AI hub gives one point of access to models from several providers. Choose the model per task by testing candidates on your own prompts and data; the harmonized API keeps client code unchanged when you switch.
• Custom model training. Data scientists package training code in Docker images and define workflow templates in Git; SAP AI Core runs them as executions on its managed Kubernetes cluster and stores the resulting models in your registered object store.
• Retrieval-augmented generation (RAG). Ground LLM responses in business-specific knowledge: retrieve relevant passages from policies, procedures, or past analysis—for example from a vector store in SAP HANA Cloud—and inject them into prompts through the orchestration service's grounding module, so answers draw on your own content.
• Agent reasoning with MCP. Agents built in Joule Studio reason across steps and reach SAP and external systems through MCP servers and APIs, with human-in-the-loop steps where an action needs approval. SAP AI Agent Hub keeps those agents and servers registered and verified.
• Model monitoring and retraining. SAP AI Core tracks metrics as part of the AI scenario lifecycle. Drift checks and the retraining pipelines they trigger are yours to build on top—decide the thresholds before go-live.

Prerequisites and readiness

This blueprint requires substantial technical and organizational maturity:

• ML engineering team. You need data scientists and ML engineers who understand model development, training, validation, and deployment. This is not a business analyst function.
• Model lifecycle discipline. Establish practices for versioning, testing, deployment, and rollback. Include code review, unit tests, integration tests, and shadow deployments.
• Inference infrastructure readiness. Plan for compute capacity, latency requirements, and scaling. Some models need GPU, others benefit from CPU-optimized runtimes. Cost modeling is critical.
• Data quality and feature engineering. Custom models are only as good as their input features. Invest in feature engineering, validation, and monitoring.
• Integration with source systems via MCP. Define how agents will discover and invoke tools, register MCP servers and agents in SAP AI Agent Hub, and agree API contracts with backend system owners.
• Governance and explainability. For high-stakes decisions (credit scoring, pricing, hiring), models must be auditable—and credit scoring and hiring are high-risk uses under the EU AI Act, with obligations applying from December 2, 2027. Plan for SHAP values, LIME, or other explainability techniques.

Architecture pattern

ML engineers keep training and serving templates in a Git repository that SAP AI Core syncs, and build Docker images into a registry that AI Core can pull from. Training runs as an AI Core execution on Argo Workflows; the trained model lands in the registered object store and is served as a KServe deployment behind an API. SAP AI Launchpad gives teams one place to manage scenarios, executions, deployments, and metrics; CI/CD pipelines, approval gates between environments, and drift-triggered retraining are yours to build around it. For LLM-based solutions, the generative AI hub's orchestration service routes prompts to the selected model with templating, grounding, filtering, and masking. Agents built in Joule Studio can call AI Core deployments or MCP servers as tools, and SAP AI Agent Hub keeps those agents and servers registered and verified. Log model versions, decisions, and performance for audit and compliance.

As you evaluate and implement these blueprints, adhere to the following design rules to avoid common pitfalls:

Use this matrix to identify which blueprint or combination of blueprints fits your organization's current state and next-step priorities:

Interpreting the table: Identify your top priority in the left column. Cross-reference your organization's maturity level. The recommended blueprint(s) are the most direct starting point for that priority. Time-to-value ranges are planning assumptions, not SAP benchmarks: they assume standard implementation discipline and no major data quality blockers.

Blueprint 1 + 2: A common pairing is conversational analytics (Blueprint 1) with predictive planning (Blueprint 2). Users explore historical patterns with "just ask," then apply predictive models to build scenarios. Both blueprints share the Datasphere semantic layer. Custom Joule Studio agents can automate parts of that loop, and SAP Enterprise Planning brings SAP-delivered planning agents (general availability was planned for Q3 2026).

Blueprint 2 + 4: For advanced forecasting, combine Smart Predict in SAC (Blueprint 2) with custom models on SAP AI Core (Blueprint 4). SAC handles interactive planning; AI Core hosts specialized demand sensing or price elasticity models that feed into plans. The generative AI hub can draft narrative explanations of model outputs for a person to review.

Blueprint 3 + 1: Establish a managed data foundation with BDC and Datasphere (Blueprint 3), then enable conversational analytics on top (Blueprint 1). This way, Joule answers from high-quality, governed data products. Access to BDC data products from inside Joule Studio is planned for general availability in Q1 2027.

Blueprint 3 + 4: Use Business Data Cloud to curate AI-ready data products (Blueprint 3), then train custom models on those products using SAP AI Core (Blueprint 4). This reduces data engineering friction for ML teams. The generative AI hub provides unified LLM access for model interpretation and documentation.