AI in SAP analytics has moved from features to agents. In SAP Analytics Cloud, just ask and analytical insights in Joule answer natural-language questions from governed models, AI-assisted features draft scripts, formulas and comment summaries, and Smart Insights and Smart Predict explain and predict. Around SAC, the SAP Business AI Platform—launched at SAP Sapphire 2026—unifies SAP BTP, SAP Business Data Cloud and SAP Business AI around the SAP Knowledge Graph: Joule Studio to build agents with more than 70 AI models to choose from, Joule Work as the conversational front end (general availability rolling out in phases since October 2026), the SAP-RPT models for predictions on tabular business data, and SAP Reltio inside Business Data Cloud for master data. This is not one tool. It is a stack where each layer solves a specific class of problem. The question is no longer "can AI help?" It is "which AI pattern solves this problem at the right speed, with the right trust, in the right place?"
Executive Q&A with just ask and Joule, first-draft FP&A commentary, predictive forecasts as a planning baseline, spend and revenue analysis on SAP's intelligent content, custom agents for repeatable workflow steps, and risk simulation. These are the use cases most likely to show measurable value early, and they stay governed.
Joule is SAP's AI assistant across its applications, and in analytics it is the place users go when they have a question, need an explanation, or want to explore governed data. Four capabilities matter most today: just ask and analytical insights in Joule for natural-language questions, AI-assisted features for drafting and explaining content, AI in planning for forecasts and scenarios, and Joule deep research for multi-source analysis.
Ask business questions in natural language and get answers as charts and tables from acquired SAC models or SAP Datasphere models, with auto-complete for measure and dimension names and admin-defined synonyms. Results can go into a story, open in data analyzer, or export to CSV or Excel. Analytical insights in Joule runs on just ask once admins index the models; standard-release customers have had it since the Q2 2025 QRC.
What it solves: Routine questions that used to wait for an analyst: executive Q&A, variance checks, ad-hoc reporting.
SAC's AI-assisted features work where users already are: calculations in data analyzer, summaries of comment threads, advanced-formula scripts and documentation for data actions, story scripts generated or explained in natural language, an AI prompt insights widget, and chart summaries in the PowerPoint add-in. They draft; people still review.
What it solves: First drafts of commentary and documentation, and script and formula help for story and planning builders.
Smart Predict writes time-series forecasts into planning versions as a baseline that planners adjust, and the AI prompt insights widget analyzes planning data and can bring in a real-time web search through an LLM. SAP Enterprise Planning, announced at Sapphire 2026, adds Joule agents that sense signals, simulate, recommend and update plans; its general availability was planned for Q3 2026, so confirm the status for your tenant.
What it solves: A data-driven starting point for forecasts, faster scenario work, and earlier warning when actuals drift from plan.
Announced at SAP Connect in October 2025, deep research in Joule goes beyond quick answers: it combines internal SAP data with external intelligence, structured or unstructured, to produce strategic analysis and reports without leaving Joule. It entered beta in December 2025; check availability before you plan around it.
What it solves: Research-style questions that need context from several sources, such as a market or supplier briefing before a decision.
Joule Studio, part of the SAP Business AI Platform, is SAP's environment for building and running enterprise AI agents and workflows. Its skill builder became generally available in July 2025 and its agent builder in December 2025; at SAP TechEd 2026 SAP said general availability of the new Joule Studio is rolling out in phases. Business users get a guided low-code experience, pro-code developers work in their IDEs with Git and CI/CD pipelines, and builders can choose from more than 70 AI models. Deploying and running custom agents on the SAP-managed Joule Studio runtime is free through December 2026 for select customers, and agents built there reach users through conversations in Joule Work.
Best fit: Custom workflow acceleration, guided planning flows, data-driven process automation, role-based task assistance, end-to-end orchestrated workflows involving multiple agents and systems.
Critical rule: Agents thrive on repeatable decisions and governed data. Start with existing workflows that are already defined, not ones you are still trying to understand. Check what SAP already delivers before you build—the Autonomous Suite's assistants and agents cover many standard processes—and chain several agents only where each hand-off has a clear owner and checkpoint.
Operating principle: Agents follow the four-rule pattern: (1) intentional design—define what the agent does and for whom; (2) governed inputs—agents work with data that already has defined semantics, access rules, and quality standards; (3) human checkpoints—approve, validate, or override agent actions before they commit; (4) measured outcomes—track whether the agent actually saves time, improves decisions, or increases reuse.
SAP-RPT is SAP's relational pretrained transformer: a foundation model built for structured business data rather than text. Instead of training a model per use case, you send example rows with known outcomes together with the rows you want predicted, and the model predicts from that context—in-context learning. SAP-RPT-1 arrived at SAP TechEd 2025 in small and large versions in the generative AI hub, with an open-source version on Hugging Face and GitHub for research and non-commercial use. The current version, SAP-RPT-1.6, is generally available in the SAP AI Core model library as SAP-RPT-1.6 and SAP-RPT-1.6-large, and SAP plans to retire SAP-RPT-1 by December 31, 2026.
SAP-RPT handles classification and regression on tables—invoice classification, payment matching, cost-center assignment, spend prediction—and SAP positions the family for time-series analysis and forecasting as well. SAP's own claims are up to 2× the prediction quality of narrow AI models and 3.5× that of language models; for SAP-RPT-1 it also cited 50,000 times less energy, 100,000 fewer GPU FLOPs and 50 times more speed than state-of-the-art LLMs. Treat these as vendor figures and test on your own data. The generative AI hub puts SAP-RPT next to models from OpenAI, Anthropic, Google, Mistral AI, Amazon and others behind one API, so you can choose the right model for each task.
Pretrained on tabular structures and relationships, it handles changing values and enterprise data formats without custom preprocessing or feature engineering. SAP-RPT-1.6 targets medium-complexity scenarios where latency and throughput matter; SAP-RPT-1.6-large aims at the highest prediction quality, and its deep context mode trades latency and cost for accuracy when you need more than 8,000 context rows. You can test it free in the RPT playground, whose API now allows 1,000 requests per hour.
Best for: Tables where a column has to be predicted from historical rows: invoice classification, payment matching, cost-center assignment, spend prediction and demand forecasting.
What this means for architects: For tabular predictions, try SAP-RPT before you train a narrow model or prompt a general LLM, and benchmark it against your current approach. For time series, compare Smart Predict in SAC with SAP-RPT-TS in SAP IBP (below). Plan the move off SAP-RPT-1 before it retires.
The generative AI hub in SAP AI Core offers models from OpenAI, Anthropic, Google, Mistral AI, Amazon, Meta, Cohere and others through one API. Its orchestration service adds templating, grounding, translation, content filtering, and data masking that anonymizes or pseudonymizes personal data before it reaches a model.
Best for: Narrative generation for executives, summarizing unstructured documents, open-ended analysis and conversational interfaces—tasks where language matters more than tabular prediction.
What this means for architects: Route tabular prediction to SAP-RPT and language tasks to an LLM, and choose each LLM on quality, cost and data-residency needs. Test on your own prompts; model rankings change faster than architectures.
SAP completed its acquisition of Reltio on May 7, 2026, and SAP Reltio is now generally available inside SAP Business Data Cloud, with unified data-product capabilities planned for Q1 2027. Reltio unifies, cleanses, harmonizes and governs core data such as customers, products and suppliers across SAP and non-SAP systems, using AI-based entity resolution to merge related records into a single golden record.
What this means for architects: Master data governance is a prerequisite for trustworthy AI, not an optional extra. Reltio automates matching and merging, but match rules, survivorship and stewardship still need owners—clean master data is a process, not a purchase.
Architecture pattern: Master data quality (SAP Reltio, SAP Master Data Governance) + governed semantics (Datasphere analytic models and BDC data products) + the right model per task (SAP-RPT for tabular prediction, LLMs for language) = AI you can trust. The SAP Knowledge Graph, at the core of the SAP Business AI Platform, gives agents a structured map of business entities, processes and relationships on top of that foundation.
The model line moves fast. SAP-RPT-1.5 added retrieval-augmented prediction—retrieving relevant records from your own data at inference time—and SAP-RPT-1.6 extends the family for large-dataset scenarios, with its biggest gains in high-cardinality classification. Around the models sits tabular orchestration, a managed workflow in the generative AI hub that SAP announced as available in September 2026: you register a data destination (a HANA Data Lake File System), the table schema and a scenario once, and each prediction becomes a single API call that retrieves and selects context rows, randomly or heuristically, before calling the model. SAP plans to extend it to other tabular foundation models, including TabPFN from Prior Labs, which SAP acquired in July 2026. At SAP TechEd 2026, SAP made SAP-RPT-TS available for predicting time-indexed business metrics such as demand in SAP Integrated Business Planning without custom machine learning, and announced SAP-RPT-2, which adds anomaly and outlier detection and embeddings.
SAP Business Data Cloud is the data foundation of the SAP Business AI Platform. It brings SAP Datasphere, SAP Analytics Cloud, SAP BW and SAP-managed data products together as a managed SaaS, and it shares data with partner platforms through BDC Connect without copies. For AI, its value is curated data products that keep business meaning intact—exactly what Joule, agents and models need.
SAP delivers and manages data products for its applications—finance, procurement, HR, supply chain and more—with business semantics intact. You install them into Datasphere spaces and build analytic models on top, or share them zero-copy to SAP Databricks, SAP Snowflake or partner platforms through BDC Connect.
What this means: Start from SAP's data products instead of rebuilding extraction and semantics for every use case, and spend the saved effort on governance and adoption.
SAP now calls its prebuilt line-of-business packages in BDC intelligent content: data products, semantic and analytical models, planning models, KPIs in My Metrics, dashboards, and AI and machine learning for use cases such as workforce planning or working capital. Packages include Finance Intelligence, Spend Intelligence, People Intelligence, Supply Chain Intelligence, Revenue Intelligence and Cloud ERP Intelligence Private, and partners such as Moody's and Thomson Reuters build their own on BDC.
What this means: AI and analytics content arrives prebuilt on governed data, so the first project is configuration and adoption rather than modeling from scratch.
SAP does not publish the internals of just ask, but it documents what the feature needs, which questions it supports and how it explains its answers. Design for that documented behavior rather than for assumptions about how queries are generated.
Admins choose which models just ask and analytical insights in Joule can use—acquired SAC models or SAP Datasphere models—and add synonyms for business terms. A readiness check on the model management page reports the model's overall readiness, describes each issue it finds, lists the affected measures, dimensions or members, and suggests fixes. Clear descriptions and consistent naming in the model matter more than clever prompts.
Supported question types: Aggregation by dimension, version or time; member, version, measure and time filters, including fiscal and dynamic periods; ranking and top-N; percentage of total, CAGR, variance and variance percentage; hierarchy drilling; currency conversion; and a requested chart type.
Just ask does not calculate between measures (a sum or ratio of two measures), produce predictive forecasts, run multi-step calculations or join across datasets, answer questions about model metadata, or filter below date level. When a question needs one of these, build the logic into the model or the story instead.
Design implication: Model the metrics people actually ask for—ratios, margins, plan versus actual—as named measures, so just ask can find them by name.
With each answer, just ask can generate an explanation: a title, a summary, and a detailed account of how the question was interpreted—the model used, the measures and dimensions recognized, exact or similarity-based matches, and any filters, variables, hierarchies, variances and the chart type chosen. A Thinking section shows its reasoning while it works. Explanations are generated in English, French, German, Italian or Spanish and require the Execute permission on the Generative AI privilege.
Why it matters: The explanation is the trust layer. If just ask misread a term, users can see it and rephrase with recognized synonyms—so train users to read it, not skip it.
Users can add the chart to a story, open it in data analyzer, or export it to CSV or Excel. Analytical insights in Joule uses the same engine, so the same model preparation pays off twice. For a deeper look at a single data point, Smart Insights shows its top contributors and how it changed or is calculated.
Analytical insights in Joule: Generally available since 2025, it needs a Joule entitlement, SAP Build Work Zone and an SAC Enterprise license—it is not available with embedded-analytics LOB licenses—and standard-release customers received it with the Q2 2025 QRC.
Agents in Joule Studio are not lightweight scripts. They reason through tasks, call tools, take action in SAP and non-SAP systems, and can hand work to other agents. That makes design and governance decisions more important than the build itself. Treat the patterns below as design guidance, and check each capability against the current Joule Studio release.
Decide how each agent starts: on request (a user asks Joule to "run the spend analysis"), on a schedule (week-ahead reports every Monday at 8 a.m.), or on an event (a KPI breaches a threshold). Conditions can be simple (every day at 9 a.m.) or compound (revenue forecast below 90% of plan and inventory above 120% of target). Compound conditions avoid alert fatigue because the agent fires only when several signals align.
Best practice: Design trigger conditions conservatively. An agent that runs too often consumes cost and attention. An agent with too many false triggers gets ignored. Start with simple triggers (time-based), measure whether the agent output is actually used, then add event-based triggers only for high-signal scenarios.
An agent's data sources define what it can see and act on. A spend analyst agent might read from Datasphere (purchase orders, invoices, commitments), Salesforce (customer contracts), and SuccessFactors (headcount). It chains these sources: find invoices for a supplier, match to contracts, check for discrepancies, flag items for approval. Action chains define what the agent can do. The spend analyst can recommend 3-way match exceptions, flag pricing variance opportunities, and draft negotiation talking points. Creating a purchase order stays behind human approval—a boundary you set deliberately, not a default you inherit.
Common action chains: Data quality agents (suggest corrections, flag duplicates, recommend merged records), anomaly agents (detect and summarize), forecast agents (run predictions, compare to actuals, surface confidence), and workflow agents (route items to queues, set priorities based on business rules, track SLAs).
SAP delivers Joule Agents inside its applications, and the SAP Autonomous Suite announced at Sapphire 2026 groups them under more than 50 domain-specific Joule Assistants that orchestrate over 200 specialized agents across finance, supply chain, procurement and other areas. Examples from the past year include the Production Planning and Operations, Production Master Data and Material Reservation agents in manufacturing, and the Booking and Receipt Analysis agents for travel and expenses. Availability differs by agent and product, so check each one before you plan around it.
Customization path: Start with a delivered agent where one fits, run it in a controlled scope, collect feedback, then extend it in Joule Studio or build your own for the gaps. This beats building from scratch because you start from SAP's process logic.
Multi-agent workflows are where open protocols matter. In May 2026 SAP said agents built in Joule Studio will support the Agent2Agent (A2A) protocol natively and that Joule's bi-directional A2A capabilities would become generally available in Q4 2026. At TechEd 2026 SAP described Joule Work orchestrating assistants and agents through open protocols including A2A and the Model Context Protocol (MCP), with third-party agents invoking Joule Agents through a governed agent gateway. The SAP AI Agent Hub is the vendor-agnostic command center to discover, inventory and govern SAP and non-SAP agents and MCP servers.
What this means for architects: Design workflows as agent collaborations with explicit hand-offs: which agent owns which decision, what context passes along, and where a human approves. Interoperability makes chains easy to build; governance decides whether they are safe to run.
Use this pattern for: Month-end close (variance analysis agent → commentary agent → account reconciliation agent → executive summary agent → human review), planning cycles (demand forecast → supply plan → financial impact → resource planning → board presentation), and cross-functional risk (credit risk agent + liquidity risk agent + operational risk agent → aggregated risk dashboard).
Agent governance rule: Build agents for repeatable workflows that are already understood. Do not use agents to explore unknown processes or make decisions you do not yet know how to make. Agents amplify clarity, and they multiply confusion just as readily.
SAP's Production Planning and Operations Agent works with Joule in SAP S/4HANA Cloud Public Edition. Planners can release production orders in natural language while the agent automatically checks the prerequisites—material availability, capacity and scheduling constraints—flags shortages and suggests workarounds. SAP introduced it at SAP Connect in October 2025, planned general availability for Q2 2026 at Hannover Messe in April 2026, and lists it in its Q3 2026 release highlights among four production agents released in Q3 under the Manufacturing Assistant, whose general availability is underway in a phased approach; SAP's product page lists SAP AI Units as required.
What this means for architects: This agent does not turn a demand forecast into an MRP run. It shortens the last mile between planning and execution—the checks and release steps planners used to do across several apps. Measure it on order-release lead time and on how often planners override its recommendations.
Smart Predict in SAP Analytics Cloud builds classification, regression and time-series forecasting models from your data in predictive scenarios—no coding, with a guided workflow. SAP-RPT needs no training at all: it predicts from example rows sent with each request. In 2026 they complement each other. Use Smart Predict when an analyst owns the model inside SAC—time-series forecasts for planning, or classification and regression with influencer reports people can read. Use SAP-RPT when a developer or an agent needs predictions on tables through an API, especially where training and maintaining a model per use case is not worth it.
You define the target and Smart Predict does the modeling: it encodes the data, trains candidate models on a training subset and uses a hold-out validation subset to pick the best one. For classification and regression the split is random, 75% training and 25% validation; for time series it is sequential, so the first 75% of the history trains the candidates and the remaining 25% selects the best. You choose the target and candidate influencers and exclude variables that would mislead the model—above all leakers, which are results of the target rather than causes.
What you get: Classification and regression models come with influencer contributions—the relative importance of each variable, summing to 100%—and category influence, and prediction explanations can add per-row reasons to the output. Time-series models show forecast accuracy by segment and a breakdown of the signal.
Typical use: Sales forecasting, revenue projection, demand planning, cash flow estimation, inventory planning.
Smart Predict treats the history as a signal and decomposes it into a trend (where the series is heading), periodics (seasonal and recurring patterns), fluctuations (dependence on recent past values) and residuals, the random noise that is left. The model is additive: the forecast is the sum of trend, cycles and fluctuations. You can add influencer variables such as a promotion flag, but you must then supply their future values, and you can segment the forecast by entity, such as product or country.
Rule of thumb: SAP suggests a history-to-horizon ratio of about 5:1—to forecast six months, provide about 30 months of history—and recommends discarding history that is too old to reflect today's business.
Planners can use a planning model as the input to a time-series scenario, generate segmented forecasts, and write them into a planning version—typically a private copy of actuals—as the baseline they then adjust with business judgment. A multi action's predictive step retrains the model on current data and writes a new forecast to a public or private version, and multi actions can be scheduled in the calendar, so the baseline refreshes after each actuals load.
Why it matters: Planners monitor plan attainment against continuously updated predictions instead of a forecast that was right only on the day it was made.
For supply chain planning, SAP-RPT-TS—made available at SAP TechEd 2026—applies the RPT approach to time-indexed business metrics, predicting values such as demand in SAP Integrated Business Planning without custom machine learning. It complements rather than replaces Smart Predict: Smart Predict serves forecasts in SAC planning models, while SAP-RPT-TS targets forecasting in IBP.
Typical use: Demand and supply forecasts in IBP where building and maintaining a statistical model per product line is costly. Compare it with your current IBP forecast models on the same history before you switch.
Retraining is something you schedule, not something Smart Predict decides. Accuracy is reported when you train—for example forecast accuracy by segment for time series—and you choose when to retrain: manually, or through a scheduled multi action whose predictive step retrains on the latest data. Track the error of past forecasts against actuals yourself, and investigate before you trust a sudden change: did the business change, did the data change, or did the model?
Typical use: Monthly rolling forecasts refreshed after each close, seasonal planning where the model must adapt before peak periods, and any forecast that feeds a decision someone signs off.
For regulated industries and EU customers, SAP unveiled EU AI Cloud on November 27, 2025: a sovereign AI and cloud offering for Europe that supports EU data residency and full sovereignty. Customers choose the level of control—SAP Sovereign Cloud on SAP Cloud Infrastructure (EU), SAP's own infrastructure in its European data centers, where all data stays in the EU; SAP Sovereign Cloud On-Site, SAP-operated infrastructure in a data center the customer owns or selects; selected hyperscalers per market; or Delos Cloud for the German public sector. Models from Cohere, Mistral AI, OpenAI and others are integrated into SAP BTP and run on SAP Cloud Infrastructure and SAP BTP in European data centers. At SAP TechEd 2026, SAP added SAP AI Core, Joule and SAP Datasphere to SAP Sovereign Cloud on AWS, with general availability planned for Q4 2026 in the UK, Australia and Canada.
What this means for architects: Sovereign hosting answers where data and models run; it does not make an AI system compliant on its own. EU AI Act duties such as risk classification, documentation and human oversight stay with you as provider or deployer. SAP's AI Governance Assistant in the SAP AI Agent Hub, planned for general availability in Q4 2026, is designed to help with EU AI Act risk classification. Check which models are offered in which deployment option before you design around one.
BDC Connect shares SAP data products with partner platforms zero-copy and in both directions, so data science and analytics teams can use SAP data where they already work and share results back. As of November 2026, BDC Connect is generally available for Databricks (since October 2025), Snowflake (since May 4, 2026) and Google BigQuery (since July 2026); Microsoft Fabric is in restricted public preview with general availability planned for H2 2026; and Amazon Athena is planned for H1 2027.
BDC Connect for Databricks shares SAP data products into Databricks Unity Catalog together with semantic metadata—comments, keys and governance tags—so teams can build ML and analytics on SAP data without extraction. SAP Databricks is the option when you want Databricks running inside BDC itself.
Typical fit: Data lake analytics, machine learning feature stores, modern data platforms.
The Snowflake connector (generally available since May 4, 2026) and the Google BigQuery connector (generally available since July 2026) extend zero-copy sharing into those clouds, and SAP Snowflake runs Snowflake inside BDC. Build the semantics once in SAP and share governed data products instead of maintaining extraction pipelines.
Typical fit: Multi-cloud strategies, cloud-native analytics, data that already lives in these warehouses.
BDC Connect for Microsoft Fabric shares SAP data products into OneLake without copies. SAP lists it as a restricted public preview with general availability planned for H2 2026, so treat it as a pilot path rather than a production dependency today.
Typical fit: Microsoft-native shops, Power BI integration, unified Microsoft analytics stack.
Sharing is zero-copy and governed by the data product, but once data is shared, each partner platform applies its own access controls and compute. The BDC Connect instance must sit in the same region and on the same hyperscaler as the data product (the partner system can sit elsewhere), and inter-region transfer and API requests are metered in capacity units. Design governance on both sides, not just in SAP.
Typical fit: Enterprises with heterogeneous data platforms who want unified analytics governance and AI.
Smart Discovery, which built a model to surface the key influencers of a measure, was built on the classic experience of SAP Analytics Cloud and has been phased out with it; SAP deprecated the classic experience at the end of Q2 2026. The questions it answered—what drives this outcome, which segments behave differently—still matter, and SAC now answers them with Smart Predict, Smart Grouping and Smart Insights.
To ask what drives an outcome, build a classification or regression scenario in Smart Predict. Its influencer contributions report ranks the variables that explain the target and shows which categories within each influencer push results up or down, and prediction explanations show the main reasons behind each individual prediction. Contributions show association, not causation—validate a driver before you act on it.
Use this for: Exploratory analysis when you have a metric but don't know what drives it. Root cause investigation. Hypothesis validation before deep-dive analysis. Market segmentation and customer profiling.
Smart Grouping clusters the points of a bubble or scatter chart into groups so natural segments become visible; the story designer sets the number of groups. For a single unusual value, Smart Insights shows how it changed over time and which dimension members contribute most. For outliers across whole tables, SAP has announced SAP-RPT-2 with anomaly and outlier detection, planned for general availability in Q1 2027.
Use this for: Customer segmentation, product portfolio analysis, supplier concentration checks, and spotting unusual values worth a closer look.
Smart Insights answers the first question after a surprise in a chart: what is behind this number? Select a data point or a variance in a story, and the Smart Insights panel combines a short text with a visualization. SAP describes it as statistical analysis that provides insights automatically, and it works on acquired data, live SAP HANA connections, and SAP Datasphere and seamless planning models.
Smart Insights answers three kinds of question about a selected value: how it changed over time, which dimension members contribute most to it, and how it is calculated, including the formula and aggregation. It works on most chart types, table cells and all variance types in stories, and you can add Smart Insights charts to a story. It does not build a price-volume-mix bridge: effects you have not modeled still need explicit calculations, for example in a waterfall chart.
Use this for: Month-end and quarter-end variance reviews, board pack preparation, and quick checks on which regions, products or customers moved a number.
Smart Insights is not available for structures, calculated columns created in tables, blended models, embedded mode, or when several data points are selected, and it does not support models with a fiscal-year setting, user-managed time dimensions or week-based date patterns. For hierarchies with more than 200,000 nodes, top contributors fall back to the leaf level, and they are not returned when the backend result exceeds one million rows; on Datasphere and seamless planning models, top contributors are unavailable when there are more than 20 candidates (KBA 3705341).
Enterprise benefit: Faster first answers in variance reviews. The commentary itself still needs a person—or a draft from SAC's AI-assisted features that a person checks.
Planning is where predictive and generative AI meet governed numbers. Smart Predict supplies baselines, AI-assisted features help builders write and document planning logic, Compass simulates uncertainty, and SAP Enterprise Planning brings Joule agents into the cycle.
Instead of anchoring on last year plus a growth rate, start from a time-series forecast built on actuals. SAC predictive planning writes Smart Predict forecasts into a planning version at the level of detail you need—top-down or bottom-up—and planners then supplement them with business judgment. Value driver trees let you test what-if changes to drivers on top of that baseline.
Use this for: Annual budgets and rolling forecasts that start from a data-driven baseline, expense and sales forecasting, workforce planning, and scenario work on top of a forecast.
Allocation and planning logic in SAC stays rule-based: you define it in data actions. What has changed is how you build it. AI-assisted features generate advanced-formula scripts and comments for data actions from natural language and generate data action documentation, while the AI prompt insights widget analyzes planning data and can bring in a real-time web search through an LLM. Treat generated logic like any code: review it, test it on a copy, and document why it exists.
Use this for: Faster build and documentation of allocation and planning logic, onboarding new planning modelers, and keeping complex data actions understandable.
Instead of three static scenarios, SAC Compass runs Monte Carlo simulations—statistical simulation rather than AI—so planners see a range of outcomes and how likely they are; since the Q3 2026 release, a simulation-results widget with smart boundaries can show those results in stories. Compass needs an SAC for Planning standard or professional license. SAP Enterprise Planning, announced at Sapphire 2026 on BDC data products and SAC, embeds Joule agents that sense signals, simulate scenarios, recommend actions and update plans; general availability was planned for Q3 2026, so confirm the status before you plan around it.
Use this for: Strategic what-if analysis. Risk and sensitivity planning. Contingency planning and stress testing. Opportunity assessment for M&A and new initiatives.
Data quality is foundational: AI models, just ask and Smart Predict all inherit whatever errors the data carries. SAP's tooling helps in specific places—master data, validation, lineage and metadata—but no single AI cleanses everything automatically.
SAP Datasphere offers data validation for views with key and hierarchy validation rules, but it has no general-purpose data quality rule engine comparable to SAP Information Steward (KBA 3493549). For profiling, rules and monitoring at scale, customers add partner tools such as Collibra Data Quality & Observability, which pushes checks down to SAP HANA Cloud and Datasphere.
Use this for: Catching broken keys and hierarchies before they reach stories and models, and deciding early which data quality tool owns rules and monitoring.
Duplicates in customers, suppliers and products are a master data problem. SAP Reltio, generally available inside BDC, uses AI-based entity resolution to match and merge related records from SAP and non-SAP sources into a golden record, and SAP Master Data Governance on SAP S/4HANA is now a core BDC component for governed master data processes.
Enterprise benefit: A single view of each customer or supplier, so revenue and spend are no longer split across duplicates—provided someone owns the match rules and the stewardship queue.
The Datasphere catalog's impact and lineage analysis shows where an asset's data comes from—sources, transformations, associations and data access controls—and what is affected when it changes, and task-chain and replication-flow monitoring shows which loads failed. Statistical checks on the content of each load, such as sudden spikes or unusual values, are something you build into your pipelines or add with a data quality tool.
Use this for: Impact assessment before changes, faster root-cause analysis when a number looks wrong, and confidence in analytics and AI outputs.
Datasphere's content generation for catalog uses AI to help data stewards generate business descriptions, business terms and KPIs, and the catalog links glossary terms, KPIs and hierarchical tags to assets. BDC added data product governance—terms and conditions, steward approval and access agreements—in September 2026, and BDC Ontology Manager, which adds AI-assisted domain views of the SAP Knowledge Graph, was planned to enter beta in November 2026. Stewards still review and approve what the AI proposes.
Use this for: Reducing time spent on governance documentation, keeping glossary and KPI definitions consistent across assets, and making data products easier to find and trust.
SAP's Global AI Ethics Policy—version 3.0, effective June 15, 2026—is grounded in UNESCO's Recommendation on the Ethics of Artificial Intelligence and sets 10 guiding principles, among them fairness and non-discrimination, transparency and explainability, human oversight, privacy and data protection, and responsibility and accountability. Every SAP AI use case goes through an AI ethics impact assessment before deployment. Your own models and agents need the same discipline.
Smart Predict does not certify fairness for you. Review the influencer contributions: if a sensitive attribute, or an obvious proxy for one, drives a prediction about people—credit, hiring, retention offers—exclude it as a candidate influencer, retrain and compare. Test outcomes across groups on a hold-out set and document what you checked; under the EU AI Act, uses such as creditworthiness assessment or recruitment decisions can be high-risk and carry documentation and oversight duties.
Use this for: Ensuring models are fair and defensible. Meeting regulatory requirements. Maintaining brand trust—biased models damage reputation.
Smart Predict can add prediction explanations to classification and regression output: for each prediction, the top influencers, their values and a normalized strength of their impact, plus each influencer's contribution for regression models, with a baseline so the result can be shown as a waterfall. Instead of a bare score, a churn prediction arrives with the reasons behind it, which managers can check against what they know.
Use this for: Regulated decision-making (credit, insurance, hiring). Building analyst confidence in model recommendations. Identifying which interventions matter most (if NPS is the biggest risk factor, improve NPS).
SAP states that it does not share customer data with anyone to train or improve large language models. In the generative AI hub, the orchestration service's data masking can anonymize or pseudonymize personal or confidential data before a prompt reaches a model, and content filters screen inputs and outputs. For data that must stay in Europe, the EU AI Cloud options keep data and models in European data centers.
Use this for: Using LLMs on business data without exposing personal data, meeting privacy regulations, and building customer trust in AI systems.
SAP data is valuable. Your models are valuable. SAP does not need to own both. Share SAP data products with the platform your data scientists already use—Databricks, Snowflake, BigQuery—or extract what other ML services such as SageMaker, Vertex AI or Azure ML need, build models there, and bring the results back into SAP for governance and decisions.
BDC Connect is generally available for Databricks (since October 2025), Snowflake (since May 4, 2026) and Google BigQuery (since July 2026); Microsoft Fabric is in restricted public preview with general availability planned for H2 2026, and Amazon Athena is planned for H1 2027. Data products are shared rather than copied, so data scientists train on governed SAP data in place and can share results back into BDC. SAP Databricks runs Databricks inside BDC itself.
Use this for: Advanced ML use cases (time-series deep learning, LLM-powered anomaly detection, computer vision for supply chain) where Databricks, SageMaker, or Vertex AI are stronger than Smart Predict.
SAP AI Core runs training workloads and serves models: you can deploy your own trained models for inference and call them through an API, and Joule Studio agents and SAP BTP applications can call such endpoints, or other APIs, as tools. This brings specialized models—computer vision on supply chain images, NLP for contract analysis, a forecasting model trained on years of competitor pricing—into SAP processes without rebuilding them in Smart Predict.
Use this for: Specialized use cases where external ML platforms excel. Leverage data science team investments in other cloud platforms. Build proprietary models that stay proprietary (deployed in external platforms, but called from SAP).
SAP is not the only vendor with AI analytics. Honest positioning matters. SAP's advantages are real but different from competitors. Choosing SAP is not choosing "the best AI." It is choosing "the right AI for ERP-native, governed, integrated analytics."
Microsoft builds Copilot into Power BI and Fabric, and Copilot replaces the classic Power BI Q&A, which retires in December 2026. Fabric data agents publish into Microsoft 365 Copilot, and Planning in Fabric IQ, generally available since July 2026, gives Microsoft native planning with write-back. Copilot needs a paid Fabric F2 or higher or Premium P1 or higher capacity. SAP's advantages: Joule works on SAP business semantics and processes, and SAC combines BI and planning on governed SAP data. Microsoft's advantages: Office and Teams integration, a broad user base and enterprise agreements. If you are Microsoft-native, Copilot is easier; if your numbers live in SAP, Joule and SAC stay closer to the source.
Tableau Pulse, generally available since February 2024, delivers metric digests with AI-generated insights. At Tableau Conference 2026, Tableau repositioned as an agentic analytics platform: Tableau MCP servers are generally available, and Tableau Agent conversational analytics is generally available in Tableau Next and in beta for Cloud and Server dashboards. SAP's Smart Insights covers part of the same ground for individual data points. Tableau's advantage: visualization-first and simple to set up. SAP's advantage: analytics, planning and forecasting on the same governed SAP data. Tableau works better for pure analytics; SAP works better for connected analytics, planning and decision support.
ThoughtSpot is search-first analytics. Its first generative AI layer, Sage, never reached general availability and was removed in June 2026, when remaining Sage customers moved to Spotter, ThoughtSpot's AI analyst agent; Spotter 3 reasons through questions and validates its own work. SAP's Joule is similar in intent. ThoughtSpot's advantage: laser focus on search and discovery and fast time to value. SAP's advantage: deeper integration with planning and SAP processes, and SAP-delivered agents across business domains. For pure discovery, ThoughtSpot is simpler; for integrated decision-making, SAP is more complete.
Qlik's agentic experience in Qlik Answers and its Qlik MCP server became generally available in February 2026, AutoML is now Qlik Predict, and prediction and automation agents followed in June 2026. With Smart Discovery retired, SAP's counterpart for driver analysis is Smart Predict. Qlik's advantage: its associative engine and strong data integration through Talend. SAP's advantage: analytics, planning and prediction in one ecosystem on SAP semantics. Qlik excels in standalone analytics; SAP excels in enterprise analytics plus planning plus decision-making.
At SAP Sapphire 2026, SAP framed its strategy as the Autonomous Enterprise: Joule Work, the SAP Autonomous Suite and the SAP Business AI Platform. SAP Connect and SAP TechEd 2026 followed with general availability rolling out in phases from October 2026. These are architectural directions, not single features, and they shape what you should build now.
The SAP Autonomous Suite equips SAP's existing business applications with AI agents that can run processes from start to finish: more than 50 domain-specific Joule Assistants orchestrating over 200 specialized agents across finance, supply chain, procurement and other areas. At TechEd 2026 SAP described a suite of assistants and agents across five horizontal domains, with industry specialization. The shift: from adding AI features to existing apps to apps that carry out work with AI.
Agents no longer work in isolation. Joule Work plans and orchestrates the steps that assistants and agents need, and open protocols—A2A and MCP—connect SAP agents with third-party agents through a governed agent gateway. The Production Planning and Operations Agent shows the pattern on a small scale: an agent that checks prerequisites and helps release production orders while the planner stays in control. Expect the same pattern across domains, with human checkpoints where decisions carry risk.
Joule Work is designed to run on desktop, mobile and voice across SAP and non-SAP systems; its mobile app was generally available by May 2026, and general availability of Joule Work overall has been rolling out in phases since October 2026. In analytics, SAP announced at Sapphire 2026 Joule agents in Business Data Cloud that generate SAC stories from data models and business context—as a private preview, so check availability. The direction: describe the outcome, and the system builds the first draft.
MCP and A2A are becoming the plumbing of enterprise AI. The SAP AI Agent Hub discovers, inventories and governs SAP and non-SAP agents and MCP servers, and the SAP Enterprise AI Gateway, planned for general availability in Q1 2027, is designed to consolidate authentication, policy enforcement, routing, monitoring, usage tracking and auditing for AI traffic. Your architecture becomes more plug-and-play, but only if every new endpoint is registered and governed.
SAP-ABAP-1 is SAP's first foundation model built specifically for ABAP development. It is available through the generative AI hub and powers ABAP assistance in Joule for Developers and Joule for Consultants, from explaining existing code to helping developers build and migrate applications. This extends the agent paradigm into custom development: your developers get an AI collaborator inside their tools.
SAP's completed Reltio acquisition (May 2026), now generally available inside Business Data Cloud, signals its view that AI quality depends on master data quality. Together with SAP Master Data Governance and the SAP Knowledge Graph, it gives agents and models consistent entities and relationships to reason over. This completes the stack: clean data → clear semantics → capable models.
What this means for architects: 2026 is not about adding more features. It is about shifting the architecture: from "AI features in applications" to "applications that carry out work with AI." From "integrate systems manually" to "discover and govern agents and tools through open protocols." From "hope your data is clean" to "govern master data with SAP Reltio and MDG." From "complex modeling" to "natural-language intent." The bar has been raised. Your AI investments today should build toward this architecture, not compete with it.
AI investments fail when ROI is not measured. Joule is delightful, but if nobody uses it, it is not ROI. Agents are powerful, but if they do not save real time, they are not ROI. SAP provides tools to measure. You need a measurement framework that tracks three dimensions: analyst productivity, decision quality, and business outcome.
Track time before AI (how long did month-end close take?) and after (actual hours post-deployment), in timesheets or task tracking—not estimates. A worked example: if 50 analysts each save 10 hours a month at a loaded cost of $150 an hour, that is 50 × 10 × $150 = $75,000 a month. Count only hours that are actually redeployed to higher-value work.
Time saved is visible. Quality improvement is harder to measure, but try. Set a baseline for how long it takes from data request to decision, then measure month by month after deployment. Also measure decision volatility: do plans change less because better information arrives earlier? Do executives override forecasts less often once they can see how a forecast was built? These are subtler but real returns.
Forecast accuracy is measurable—for example, the mean absolute percentage error of your forecasts before and after. Tie it to a business outcome. A worked example: if better forecasts let you hold 5% less safety stock on $50M of inventory at a 20% annual carrying cost, the saving is $50M × 5% × 20% = $500,000 a year. Measure the downstream financial impact, not just the forecast metric.
All of these pieces work because they are architected to fit together. Here is the complete picture—five layers from data to automation.
AI only works at scale if governance, ownership, and monitoring stay with it. This means semantic consistency, model performance tracking, agent approval workflows, and clear liability for data and decisions. The operating model connects Datasphere governance to SAP's intelligent content, Joule, Smart Predict and SAP AI Core. It is not optional.
Define KPIs, metrics, dimensions and business rules once—in Datasphere analytic models and the catalog's glossary—and reuse them in SAC stories, just ask and planning. Change a definition once, and every consumer of that model sees the change. This eliminates the "shadow analytics" problem where every tool has its own version of truth.
Decide where agents may act alone and where a human must approve. Joule Studio provides identity and access controls, pre-launch testing and visibility into agent behavior; you define which actions need approval, who approves, and what happens after.
Track the accuracy of Smart Predict forecasts against actuals, retrain on a schedule you control, and evaluate agents continuously—Joule Studio adds evaluation and observability for agent behavior and failures. Set thresholds for when a human reviews a model's or agent's output before it is used.
The Datasphere catalog's impact and lineage analysis traces data from source to analytic model and story, and data product governance in BDC adds terms, steward approval and access agreements. Data quality rules need an owner and a tool; when quality drops, you need to know where and why.
Design so that every agent decision, every forecast and every Joule answer can be traced back to the data, rules and model behind it: just ask explanations, Smart Predict influencer reports and prediction explanations, and agent observability all help. This is critical for regulated industries and high-stakes decisions.
Track which agents are used, how often Joule answers are acted upon, whether Smart Predict forecasts improve decision quality, and whether analyst time is actually saved. AI investment should show measurable return, not just capability.
Not every organization approaches AI the same way. Here are three proven paths, each starting from existing strength and scaling from there.
If: You have good semantic models in Datasphere or solid dashboards in SAC. Your users ask the same questions over and over.
Path: Enable just ask, and analytical insights in Joule, on your existing models; run the readiness check and add synonyms. Let users ask questions in natural language. Measure question-to-answer time saved and user adoption.
Timeline driver: Model readiness—descriptions, synonyms and the readiness check—more than the feature itself; disruption is minimal.
If: You need better forecasts. You have historical data and clear KPIs to predict.
Path: Use Smart Predict predictive scenarios to build forecasts on top of your governed data, and write them into planning versions or show them in dashboards.
Timeline driver: History and data preparation—SAP suggests about five periods of history for every period you forecast—plus agreement on how forecast accuracy will be measured.
If: You have repeatable workflows that are slow or manual. You know exactly what you want an agent to do.
Path: Use an SAP-delivered Joule Agent where one fits, or build a custom agent in Joule Studio. Start with guided workflows, add approvals, measure time saved.
Timeline driver: Clarity of the workflow, approvals and data access rather than the build itself. Measure time saved on the targeted workflow against a baseline.
The same AI discussion means different things depending on who is reading it. Good architecture work makes that explicit.
Start with Joule and one piece of SAP's intelligent content. Learn that AI in SAP is now modular, not monolithic. Pick one pattern, prove value on a small scope, then decide what comes next.
Map your use case to one of the five layers. Check whether you are solving a Joule question, a Smart Predict forecast, a Joule Studio agent workflow, or a custom AI Core pattern. Each has different governance, timeline, and ROI characteristics.
Design the semantic layer first. Everything else—Joule, Smart Predict, agents, intelligent content—depends on governed Datasphere models. Get semantics right, and AI adoption accelerates. Get semantics wrong, and you will be rebuilding agents and retraining models forever.
Invest in AI where it reduces decision cycles on repeatable choices or measurably frees analyst time for real work. Set clear ROI targets: time saved, forecast accuracy improvement, or decision velocity increase. Do not fund AI pilots that don't track these metrics.
SAP publishes learning paths, help docs, and reference architectures across the AI stack. Here are the key entry points for each layer.
AI in analytics delivers value when it shortens decision cycles, improves analyst productivity, or surfaces governed intelligence at scale. These illustrative scenarios show patterns that work.
Illustrative pattern: month-end variance analysis and commentary take days of manual work, and analysts spend more time writing explanations than analyzing root causes.
Illustrative pattern: forecasts rely on simple averages and manual adjustments, confidence is low, and inventory is over-purchased against weak signals.
Illustrative pattern: sales forecasting runs on spreadsheets and manual CRM updates, reconciliation takes days, and reps have no quick view of opportunity and pipeline health.
AI in analytics is powerful, but it is only valuable when it connects to semantics, governance, and measured outcomes. These neighboring topics show what AI requires to work at scale.
Sapphire and TechEd 2026. At SAP Sapphire 2026 in May, SAP launched the SAP Business AI Platform, which unifies SAP BTP, SAP Business Data Cloud and SAP Business AI around the SAP Knowledge Graph, with Joule Studio as the build environment, the SAP AI Agent Hub to discover, inventory and govern SAP and non-SAP agents and MCP servers, and Joule Work as the conversational front end. Anthropic’s Claude is among the foundation models the platform uses, Mistral AI and Cohere provide sovereign model options on SAP’s cloud infrastructure, and NVIDIA OpenShell is the secure runtime for Joule Studio. At SAP TechEd 2026 (Berlin, October 27–29), SAP said general availability of Joule Work and Joule Studio is rolling out in phases, opened Joule Studio to more than 70 AI models, made SAP-RPT-TS available for time-series forecasting, and announced SAP-RPT-2 with anomaly and outlier detection, planned for Q1 2027. Read the full Sapphire 2026 analysis →