- Home
- AI
- Machine learning
Machine learning in SAP
Where predictive models run in an SAP landscape, which scenarios SAP already ships, how to build your own, and how the new tabular foundation models change the work.
- HANA PAL · APL
- ISLM
- SAP AI Core
- SAP-RPT
- Smart Predict
Facts checked in October 2026.
At a glance
- In the database
- HANA PAL · APL
- In S/4HANA
- ISLM, since 2020
- On the platform
- SAP AI Core
- Without training
- SAP-RPT-1.6 · TabPFN
- For planners
- SAC Smart Predict
- For data scientists
- SAP Databricks in BDC
- Python client
- hana-ml 2.30 (Sep 2026)
Where machine learning runs in SAP
There is no single “SAP machine learning” product. Models run in four places, and the right one depends on where the data sits and who uses the result.
In the database
SAP HANA runs the algorithms next to the data with PAL and APL. Nothing is copied out.
Inside S/4HANA
ISLM manages the scenarios SAP delivers and the ones you create, and shows the results in Fiori apps.
On the platform
SAP AI Core trains and serves your own models in containers, plus SAP-RPT and the other models of the generative AI hub.
In analytics and data
SAC Smart Predict for business users, SAP Databricks in SAP Business Data Cloud for data scientists.
Machine learning or generative AI? Machine learning predicts a number, a class or a date from structured data: will this invoice be paid late, how much stock will we sell. Generative AI produces text. The two meet in SAP-RPT, a foundation model that predicts from tables and runs through the generative AI hub like a language model.
What came before, and where it went
SAP has renamed and replaced its ML tools several times. If a project document mentions one of these, this is where the work moves.
| Product | Status | Use instead |
|---|---|---|
| SAP Predictive Analytics2015–2024 | RetiredMainstream maintenance ended 31 Dec 2024; priority-one support only until 31 Dec 2026 | SAC Smart Predict, HANA APL and PAL |
| SAP Leonardo Machine Learning2017–2020 | RetiredFunctional service APIs switched off on 30 June 2020 | SAP AI Core and SAP AI services |
| SAP Data Intelligence2019– | LegacyListed as legacy; SAP moves its pipelines and ML scenarios to newer products | SAP Business Data Cloud and Datasphere for pipelines, SAP AI Core for ML scenarios |
| Data Attribute Recommendation–2026 | RetiredRemoved from the eligible cloud services on 31 Mar 2026; usable until the end of the subscription term | SAP-RPT, as SAP itself suggests |
| SAC Search to Insight–2025 | ReplacedDeprecated in the Q4 2024 release, removed in Q1 2025 | Just Ask (and Joule in SAP Analytics Cloud) |
ISLM: machine learning inside S/4HANA
Intelligent Scenario Lifecycle Management is the framework that trains, activates and monitors ML models in SAP S/4HANA, both SAP’s and your own. It replaced the Predictive Analytics integrator (PAi) with S/4HANA 2020.
Embedded
Same systemThe model is trained and runs in the HANA database of S/4HANA, with APL or PAL.
- Good for
- Regression, classification and forecasting on S/4HANA tables
- Lifecycle
- Train, then activate a model version: no deploy step
- Data
- Never leaves S/4HANA
- Licence
- Included in the S/4HANA licence
Side-by-side
On the platformThe model runs on SAP AI Core or an SAP AI service; S/4HANA sends the data and reads the result.
- Good for
- Documents, text, images, large volumes and generative AI
- Lifecycle
- Train, deploy, then activate
- Data
- Sent to the platform region of the service
- Licence
- The platform service is billed separately
The lifecycle in two Fiori apps
Create
Intelligent Scenarios app: pick the algorithm, the training and apply datasets (CDS views) and the target field.
Train
Intelligent Scenario Management: filters, a version description, how many records to use.
Deploy
Side-by-side scenarios only: the trained model is deployed on SAP AI Core.
Activate
Compare the model versions and their quality figures, then activate one.
Consume
Fiori apps and ABAP code read the predictions through the ISLM APIs.
Monitor
Watch the quality figures and retrain when the business changes.
Before you start
- SAP S/4HANA 2020 or later (cloud or on-premise); ISLM does not exist in ECC
- APL and PAL installed on the HANA database, with the script server enabled
- Business role SAP_BR_ANALYTICS_SPECIALIST for the two Fiori apps
- bgRFC set up for ISLM: a supervisor destination and the inbound destination ISLM_QBGRFC_INBOUND_DEST (queue prefix ISLM_)
Where ML meets generative AI. ISLM also carries the ABAP AI SDK (S/4HANA 2023 and later): the same framework that manages ML models now manages the calls from ABAP to large language models.
Scenarios SAP delivers
A selection of the ML scenarios shipped with S/4HANA. Check the Intelligent Scenarios app for the full list in your release.
| Scenario | Type | What it predicts or proposes |
|---|---|---|
| Supplier delivery predictionProcurement | Embedded | Predicts whether purchase order items will arrive early, on time or late (Monitor Purchase Order Items app). |
| Contract consumptionProcurement | Embedded | Predicts when quantity contracts will be used up. |
| Blocked invoices with quantity varianceAccounts payable | Embedded | Helps clear supplier invoices blocked for quantity differences. |
| Slow and non-moving stockInventory | Embedded | Detects early the materials that are about to stop moving. |
| Predicted delivery delaySales | Embedded | Flags sales order items likely to be delivered late, learned from about 130 standard fields. |
| Quotation conversion rateSales | Embedded | Probability that a quotation becomes a sales order. |
| Abnormal liquidity itemsCash management | Embedded | Flags unusual items in the liquidity forecast. |
| Project cost forecastProjects | Embedded | Forecasts the cost of projects in Enterprise Portfolio and Project Management. |
| Receivables line-item matchingAccounts receivable | Side-by-side | SAP Cash Application: matches incoming payments to open items from past clearing. |
| Payment advice extractionAccounts receivable | Side-by-side | Reads payment advice PDFs and emails for the matching. |
| GR/IR reconciliationGeneral ledger | Side-by-side | Proposes next step, priority and root cause for open GR/IR items (Intelligent Account Automation, S/4HANA Cloud; at least 1,000 open items). |
| Purchase requisition approvalProcurement | Side-by-side | Proposes the approval of purchase requisitions from past decisions. |
| Material group proposalProcurement | Side-by-side | Proposes the material group of free-text requisition items. |
SAP HANA: PAL and APL
The oldest and still the fastest way to run ML on SAP data: the algorithms run inside SAP HANA, next to the tables, so millions of rows are scored without leaving the database.
PAL
Predictive Analysis Library: more than 90 algorithms you choose and tune yourself. Part of the Application Function Library (AFL), included with SAP HANA.
APL
Automated Predictive Library: automated gradient boosting and time series, with the model built and checked for you. The engine behind most embedded S/4HANA scenarios.
hana-ml
Python client (version 2.30, September 2026): HANA DataFrames push the work into the database, plus visualisers, model storage and a scheduler. The R client is no longer actively released.
Algorithm families in PAL
Classification
Yes or no, or one of several classes.
- HybridGradientBoostingClassifier
- RDTClassifier
- LogisticRegression
- SVC
- MLPClassifier
- NaiveBayes
- KNNClassifier
Regression
A number, such as an amount or a number of days.
- LinearRegression
- PolynomialRegression
- GLM
- SVR
- HybridGradientBoostingRegressor
Time series
Forecasts, including intermittent demand.
- AutoARIMA
- AutoExponentialSmoothing
- Croston
- AdditiveModelForecast
- LSTM
- Hierarchical_Forecast
Clustering
Groups of similar customers, materials or items.
- KMeans
- DBSCAN
- HDBSCAN
- GaussianMixture
- SOM
Anomaly detection
Unusual records and changes in a series.
- IsolationForest
- OutlierDetectionTS
- CPD
- BCPD
AutoML
Searches the algorithm and its settings for you.
- AutomaticClassification
- AutomaticRegression
- AutomaticTimeSeries
- Pipeline
Example: late payments with hana-ml
# Which open items will be paid late? Trained and scored inside SAP HANA.
from hana_ml.dataframe import ConnectionContext
from hana_ml.algorithms.pal.partition import train_test_val_split
from hana_ml.algorithms.pal.unified_classification import UnifiedClassification
conn = ConnectionContext(address='<hana-host>', port=443, user='<user>',
password='<password>', encrypt=True)
items = conn.table('AR_ITEMS_HISTORY', schema='ML') # a HANA DataFrame: no data is copied
train, test, valid = train_test_val_split(data=items, id_column='ITEM_ID',
training_percentage=0.8,
testing_percentage=0.1,
validation_percentage=0.1)
model = UnifiedClassification(func='HybridGradientBoostingTree', n_estimators=200)
model.fit(data=train, key='ITEM_ID', label='PAID_LATE')
result = model.predict(data=test.deselect('PAID_LATE'), key='ITEM_ID')
print(result.head(5).collect()) # only these five rows come back to PythonPython with hana-ml: the HANA DataFrame only builds SQL, so training and scoring run in the database.
Requirements
- On-premise: an AFL version that matches the HANA revision, and the script server enabled
- SAP HANA Cloud: the script server enabled, which needs at least 3 vCPUs
- Scoring in the database: SAP cites under 10 ms per call and hundreds of thousands of time series in parallel
Not for classic ML: the vector engine of SAP HANA Cloud (REAL_VECTOR type, COSINE_SIMILARITY and L2DISTANCE functions) serves retrieval-augmented generation with LLMs.
Your own models on SAP AI Core
When SAP’s scenarios and the HANA libraries are not enough, SAP AI Core runs any containerised model on Kubernetes: training on Argo Workflows, serving on KServe, managed in SAP AI Launchpad.
Git repository
Training and serving templates (YAML) are synced from Git into SAP AI Core.
Docker image
Your code in any framework (scikit-learn, PyTorch, TensorFlow), pushed to a registered registry.
Object store
An object store secret points to the training data and receives the model artifacts.
Train
Executions run on Argo Workflows and log metrics; the model is saved as an artifact.
Serve
Deployments run on KServe and give a URL for inference, with autoscaling.
Operate
SAP AI Launchpad shows executions, logs, metrics and deployments.
Serving template
apiVersion: ai.sap.com/v1alpha1
kind: ServingTemplate
metadata:
name: late-payment-server
annotations:
scenarios.ai.sap.com/name: "late-payment"
executables.ai.sap.com/name: "late-payment-server"
labels:
scenarios.ai.sap.com/id: "late-payment"
ai.sap.com/version: "1.0"
spec:
inputs:
artifacts:
- name: model # from the training run
template:
apiVersion: "serving.kserve.io/v1beta1"
metadata:
labels: |
ai.sap.com/resourcePlan: starter
spec: |
predictor:
minReplicas: 1
maxReplicas: 2
containers:
- name: kserve-container
image: docker.io/<account>/late-payment:1.0
env:
- name: STORAGE_URI # mounted at /mnt/models
value: "{{inputs.artifacts.model}}"Stored in Git and synced into SAP AI Core. The resource plan label sets the size of the model server.
Concepts
| Scenario | The use case, grouping its executables and versions |
|---|---|
| Executable | A training or serving template from Git |
| Configuration | Parameters and input artifacts bound to an executable |
| Execution / deployment | A training run, or a running model server |
| Resource group | Isolates teams or tenants (a Kubernetes namespace) |
| Plan | Memory | CPU | GPU |
|---|---|---|---|
Starter | 2.5 GB | 1 | – |
Basic | 10 GB | 3 | – |
Basic.8x | 115 GB | 31 | – |
Infer-S | 10 GB | 3 | Standard |
Infer-M | 25 GB | 7 | Standard |
Infer-L | 55 GB | 15 | Standard |
Train-L | 47 GB | 5 | Advanced |
Service plans: free, standard and extended. The generative AI hub, and with it SAP-RPT, needs the extended plan. Custom models are billed on compute, storage and an hourly baseline.
ML for planners and data scientists
Two audiences that rarely touch S/4HANA directly: business users in SAP Analytics Cloud, and data teams in SAP Business Data Cloud.
Smart Predict
SAP Analytics Cloud
Classification, regression and time-series scenarios without code. Time series are split into trend, cycles and fluctuations; the model trains on 75% of the data and is tested on 25%.
Predictive planning
SAP Analytics Cloud
Forecasts written into planning models, but only into a private version that the planner then publishes.
Smart Insights and Just Ask
SAP Analytics Cloud
Smart Insights explains a value (drivers, changes over time); Just Ask answers questions in natural language and replaced Search to Insight in 2025.
SAP Databricks
SAP Business Data Cloud
Notebooks, Unity Catalog and MLOps on SAP data products shared through Delta Sharing, without copies. Generally available since 30 April 2025; results can be shared back to BDC.
Datasphere is for business data users, SAP Databricks for data engineers and ML developers. Companies that already run their own Databricks connect it with BDC Connect (since October 2025). At TechEd 2026 SAP announced more Databricks features natively in BDC (Lakeflow, Agent Bricks, Lakebase).
SAP-RPT: predictions without training
SAP-RPT (relational pretrained transformer) is a foundation model for tables. You send labelled example rows and the rows to predict in one request, and it answers from the context, the way a language model answers from a prompt. SAP now recommends it as the starting point for classification and regression.
| Classic ML (PAL, APL, AI Core) | SAP-RPT | |
|---|---|---|
| Training | Train a model per use case, then retrain | None: examples go in the request |
| Time to a first result | Days to weeks | Minutes |
| Data per prediction | Unlimited history for training | Up to 2,048 rows of context, or 65,536 with the large model (1.6) |
| Latency | Milliseconds in the database | Seconds; not for sub-200 ms calls |
| Explanations | APL contributions, SHAP | Feature importance and most influential rows (1.6) |
| Where it runs | HANA, S/4HANA, SAP AI Core | Generative AI hub in SAP AI Core, HANA Cloud |
Versions
- Dec 2025Live
SAP-RPT-1
Classification and regression from examples; small and large models, plus an open-source variant (sap-rpt-1-oss) for non-commercial use.
- Jul 2026Live
SAP-RPT-1.5
Shown at Sapphire 2026: first explanations of the predictions.
- Sep 2026Live
SAP-RPT-1.6
Column importance, most influential rows, Deep Context Mode on the large model and the Tabular Orchestration workflow.
- Oct 2026Live
SAP-RPT-TS
Zero-shot time-series forecasting, in SAP HANA Cloud, SAP AI Core and SAP IBP.
- Q1 2027Next
SAP-RPT-2
Planned: anomaly detection, embeddings of structured data and several tables in one request.
An independent view. A February 2026 study of the open-source variant on synthetic SAP-like data found SAP-RPT-1 reaching about 91–96% of the accuracy of a tuned gradient boosting model, at about 28 seconds for 2,048 rows. Its advice matches SAP’s own: use RPT to prototype fast, and keep a trained model where accuracy and latency matter most.
TabPFN. In 2026 SAP also bought Prior Labs, the company behind the TabPFN tabular model. TabPFN-3.5 Plus has been available in SAP AI Core since September 2026. You can try SAP-RPT for free in the playground at rpt.cloud.sap.
Choosing the right tool
Based on SAP’s own architecture guidance (October 2026): start with what SAP ships, then the simplest tool that fits.
A classification or regression on SAP tables, quickly, with a few thousand rows of history
SAP-RPT
Forecasts for many series, anomaly detection or clustering, with data that must stay in the database
HANA PAL or APL
A prediction shown inside an S/4HANA Fiori app
An ISLM scenario (check SAP’s first)
Payment matching or GR/IR clearing
SAP Cash Application or Intelligent Account Automation
Deep learning, your own framework, A/B tests and CI/CD for models
SAP AI Core
Forecasts made by planners, without code
SAC Smart Predict
Data science across SAP and non-SAP data products
SAP Databricks in Business Data Cloud
Data trapped in PDFs and emails: invoices, payment advices, orders
SAP Document AI
Practical tips
What decides whether an ML scenario is used after go-live.
History beats algorithms
The model learns your past decisions. Inconsistent clearing, missing reason codes or a recent process change matter more than the choice of algorithm.
Enough volume
SAP asks for at least 1,000 open GR/IR items for that scenario; the delivery delay model reads about 130 fields. Small company codes may not reach a useful model.
Confidence thresholds
Decide what is posted automatically. A common Cash Application setup clears automatically at 90% confidence or more and sends 70–90% to a clerk for review.
Retrain on purpose
There is no fixed retraining cycle. Retrain after reorganisations, new payment behaviour or a new chart of accounts, and compare the new version before activating it.
Data residency
Embedded scenarios keep the data in S/4HANA. Side-by-side services send it to the platform region: check it with your data protection officer, especially in the EU and the UK.
Licences
Embedded ISLM scenarios are included in S/4HANA. Cash Application, Intelligent Account Automation, SAP AI Core and Document AI are paid services; agents use AI Units.