TL;DR
SAP AI Core is SAP’s managed runtime for your own AI models and LLM apps on the SAP Business AI Platform. It works in two ways, either training and serving custom models in containers or calling foundation models through the generative AI hub. SAP AI Launchpad is the console, and Joule is the assistant your users see on top. The generative AI hub needs the extended service plan, so the free tier and standard plan do not include it. Pricing is consumption-based, and SAP points buyers to its Discovery Center estimator instead of a single list price. Start with the 30-day trial and one workload with a clear baseline, then let the result decide the next step.
Key Takeaways SAP AI Core is a fully managed AI service that trains, hosts and serves models on the SAP Business AI Platform. It has two paths, your own model in containers or foundation models from several providers through the generative AI hub. AI Launchpad manages AI Core, while Joule and Joule Studio sit in separate layers above it. The generative AI hub is available only on the extended plan, and free or standard tenants can upgrade without losing work. Costs are consumption-based through Pay-As-You-Go or an enterprise agreement, with no single published list price. Data quality, grounding scope and a named owner decide whether a workload reaches production, far more than the model does. Watch on YouTube
What Does It Take to Deploy an Enterprise LLM?
Kanerika’s team walks through what it takes to move an enterprise LLM from a promising demo to a governed production service, the same gap SAP AI Core is built to close.
Gartner’s Warning: At Least 30% of GenAI Pilots Stall After Proof of Concept
In July 2024, Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. The reasons that Gartner listed were poor data quality, weak risk controls, rising costs and unclear business value. Not one of them is about the model.
A second Gartner forecast is sharper still, because it ties failure directly to data. Through 2026, it expects organizations to abandon 60% of AI projects that lack AI-ready data . For SAP teams, then, that pressure lands on a very practical question, which is where the model will actually run once the pilot ends.
SAP’s answer, in short, is SAP AI Core. Yet SAP now also sells Joule, Joule Studio, AI Launchpad and a generative AI hub. So teams keep asking which of them runs their model, what it costs and when it is worth building on.
What Is SAP AI Core?
SAP AI Core is, in SAP’s own words, “a fully managed AI service on the SAP Business AI Platform” (SAP product page ). Its stated job is to operate, govern and scale AI workloads in a standard way. In other words, it is where models run: you bring the code or pick a foundation model, and SAP handles hosting, scaling and plumbing.
A Managed Runtime on the SAP Business AI Platform
The platform name changed at SAP Sapphire on 12 May 2026, when SAP announced the SAP Business AI Platform. The new platform unifies SAP BTP, SAP Business Data Cloud and SAP Business AI in one environment. As a result, newer pages say Business AI Platform, while tutorials still provision the same service inside a BTP subaccount.
What It Does and What It Does Not Do
AI Core trains, hosts and serves models, and it routes calls to large language models through one governed interface. Its central model library holds SAP and third-party tabular, language and multimodal models next to the custom models you register.
The limits matter too. It is not a chat assistant or a data platform, and it does not read your S/4HANA tables on its own. Instead, grounding a model in SAP data is design work your team has to plan and fund.
The Benefits SAP Claims, and the Catch
SAP’s AI Core page lists three benefits. First, trusted models and data: SAP, third-party and custom models connect to enterprise data and business knowledge, so outputs carry business context. Second, governance, with consistent guardrails, observability and access policies across every AI interaction.
Third, scale, because a managed path turns one project into a reusable service across SAP processes and custom apps. In addition, all three sit on GPU-accelerated, auto-scaling infrastructure in public and private clouds. Still, these are SAP’s claims, not measured results, and every one of them assumes your data is ready.
Where AI Core Sits Next to Launchpad, the Hub and Joule
SAP’s developer hub calls AI Core the “runtime for training and serving ML + GenAI models”. AI Launchpad, by contrast, is the web interface for scenarios, models and deployments. The table below places the six names that teams confuse most often.
Product What it is Who uses it Relation to AI Core SAP AI Core Managed runtime for training, serving and calling models ML engineers, BTP developers The service itself SAP AI Launchpad Web console for scenarios, models, deployments and prompts Admins, data scientists The front end that manages AI Core Generative AI hub Model library and orchestration service for LLMs App developers Part of AI Core, extended plan only SAP Joule Assistant inside SAP cloud applications Business users A separate assistant layer for end users Joule Studio Low-code authoring for Joule skills, prompts and tool integrations Developers, consultants A separate authoring tool; classic edition GA since December 2025 SAP Business Data Cloud SAP’s data layer for business data Data teams Supplies the data that grounding and training depend on
So if your question is really about the assistant, our guide to SAP Joule and what it costs covers it in depth. If instead it is about where the data comes from, start with SAP Business Data Cloud . Otherwise, the rest of this post stays with the runtime.
How SAP AI Core Works: Two Paths
Every AI Core workload takes one of two paths. Either you train and serve a model you own, or you call a foundation model that someone else trained. The two share an account, a console and a security model, but almost everything else differs.
Kanerika Service
Generative AI Services
Kanerika designs grounded LLM applications, from data preparation and retrieval to evaluation and guardrails, so a pilot has a funded path to production.
Explore Generative AI Services Path 1: Train and Serve Your Own Model
The custom path runs on containers, so your team first packages training code as a Docker image. Then it describes the pipeline as an Argo WorkflowTemplate in a Git repository that AI Core syncs, as SAP’s training tutorial shows. Next, AI Launchpad lists each template as a scenario, and you run executions with a configuration inside a resource group.
Serving follows the same pattern, with a ServingTemplate that wraps a KServe InferenceService with Knative autoscaling and exposes an endpoint your apps call. If you have run models on Kubernetes before, the shapes will look familiar, because they are open-source parts with SAP’s management layer on top. For background, our explainer on AI inference vs training covers why the two jobs need different resources.
Path 2: Call Foundation Models Through the Generative AI Hub
The second path skips training entirely, since the generative AI hub gives access to OpenAI, Anthropic, Gemini, Mistral, NVIDIA, Amazon and SAP models. It then routes them through the AI Core orchestration service, according to LiteLLM’s SAP provider documentation . Versions change often, so check the current list on SAP Help Portal rather than trusting any blog’s version names, this one included.
Above all, orchestration is where the hub earns its place, because one API sits in front of every model. Then you add modules for prompt templating, content filtering, data masking, grounding, translation and model fallback, as the SAP Cloud SDK for AI documentation describes. So a team can switch providers or add a masking step without rewriting its application.
Shared Plumbing Both Paths Use
Both paths start from one BTP subaccount and one service key, which holds the OAuth credentials your code uses. Resource groups separate workloads and their secrets, while AI Launchpad sits on top as the console. Treat resource groups as an organizing tool, though, not as a full tenant-isolation design for separate business units.
Where SAP Data Comes In
Neither path, however, connects to SAP business data by default. For example, grounding an LLM in sales orders, or training a model on S/4HANA history, still needs extraction, cleansing and access rules. That is integration work, and our guide to SAP data integration walks through the main patterns.
In short, the table below sums up what each path asks of your team.
Component Path 1: your own model Path 2: generative AI hub What you own Model Your code, packaged as a Docker image Foundation model from the hub library Choice, testing and retirement Pipeline Argo WorkflowTemplate synced from Git Orchestration configuration Version control and reviews Serving ServingTemplate on KServe, inference endpoint Orchestration endpoint Latency targets and fallback Safety Checks you write into your code Filtering and masking modules Policy, thresholds and logs Data Training data you prepare Grounding sources you connect Quality, access and freshness Console AI Launchpad AI Launchpad Roles and approvals
What You Can Build With SAP AI Core
AI Core workloads fall into four groups, and each one maps to one of the two paths. The examples below are hypothetical patterns, not reported customer results, so treat them as starting points for your own baseline.
Predictions on SAP Tables Without Training
Tabular AI is the least discussed part of AI Core, yet it may be the most useful for SAP teams. For instance, SAP-RPT is SAP’s own foundation model for predictions on finance, sales and supply chain data. In the same library, PriorLabs TabPFN also gives zero-shot predictions on datasets of up to 1M rows without training, according to SAP’s AI Core page .
So for a credit team, that could mean scoring late-payment risk on open invoices from a table extract, then checking the scores against last quarter. Our piece on AI in demand forecasting also shows how to judge predictions like these before anyone acts on them.
Grounded Answers Over SAP and Document Data
The generative AI hub can also ground an LLM in documents, SAP data and business knowledge, with masking and filtering applied along the way. The trial also includes hands-on vector RAG search on the SAP HANA Cloud vector engine. If the pattern is new to you, our RAG pipeline guide explains each moving part.
Custom Models for Forecasting, Classification and Anomaly Scoring
When the logic is specific to your business, the custom path fits better. For example, a data science team might train a classifier for duplicate vendor records, then serve it as an endpoint that a BTP app calls. After launch, retraining and monitoring follow the same discipline described in our MLOps orchestration guide .
AI Inside Apps, Agents and ABAP Work
AI Core also runs the model behind custom apps and agents that call it as one step in a process. For the wider category, see our overview of agentic AI applications , and for SAP’s own agent line, our guide to SAP AI agents . For developers, the SAP-ABAP model in the library helps with code understanding and explanation, which pairs well with SAP ABAP development .
Each use case also needs its own data and its own measure of success, as the table shows.
Use case Path Data needed How you measure it Payment-risk scoring with SAP-RPT or TabPFN Model library, tabular models Clean table extract with history Hit rate against last quarter Grounded Q&A over policies and SAP records Generative AI hub with grounding Indexed documents, scoped SAP data Accuracy on a test set, with sources Duplicate vendor detection Custom model Vendor master history Precision and recall on reviewed samples Demand forecasting Custom or tabular model Sales and inventory history Forecast error against the current method ABAP code explanation Model library, SAP-ABAP Your custom code base Review time per change Agent step inside a workflow Either path Process data and tool access Completion rate and human overrides
SAP AI Core Plans and Pricing
SAP AI Core has three entry points, the free tier, the standard plan and the extended plan. One fact changes most decisions, because the generative AI hub is available only with the extended service plan .
Checklist
Generative AI Checklist: Secure Adoption and Governance
A step-by-step checklist for the security, governance and adoption work that a generative AI workload needs before it goes live.
Get the Checklist → Free Tier, Standard and Extended
Free and standard tenants can also move to extended later without losing data or work, according to SAP’s provisioning tutorial. That upgrade is not free, though, since moving from the free tier to extended starts to incur cost. So a team can explore on a small plan first and upgrade only when an LLM use case is ready.
Plan Generative AI hub Typical use Note Free tier No Learning and first experiments Upgrade later keeps data and work Standard No Paid workloads that do not need LLMs Confirm what is included in SAP Discovery Center Extended Yes LLM apps, orchestration and grounding Required for the generative AI hub
In practice, the plan question comes down to one test. If the first workload needs a foundation model, budget for extended from day one. If it does not, start smaller and upgrade when it does.
How You Pay
After the trial, SAP sells AI Core licenses through Pay-As-You-Go for SAP BTP or through an Enterprise Agreement, as the AI Core trial page explains. Usage is consumption-based, so the bill tracks what your workloads actually run.
SAP’s AI pricing page bills custom-built agents and apps running in production on compute, storage, agent memory, workflow execution and AI tokens. That page covers the Joule Studio runtime, not AI Core’s own price list, but it shows the kind of meters SAP uses.
In fact, SAP does not publish a single list price for AI Core on its product pages. So treat any firm dollar figure you see online as an estimate to check against SAP Discovery Center. If you are pricing Joule instead, its AI Units follow a different model, explained in the same SAP Joule guide .
Estimating the Full Cost
First, use the estimator in SAP Discovery Center for the service itself. Then add the costs that sit outside it, which are often larger. Data pipelines, SAP HANA Cloud vector storage, monitoring, security reviews and engineering time all belong in the business case as well.
A simple rule helps here, too. If the data work is not funded, the AI Core line item will hardly matter, because the pilot will stall long before it reaches production.
How to Get Started With SAP AI Core
The setup itself is short, since SAP’s booster provisions the service for you. Choosing the right first workload takes longer, because it decides whether the trial teaches you anything. SAP’s tutorials cover each click, so the steps below focus on the decisions around them.
Step 1: Pick One Workload and a Baseline
First of all, choose one use case with an owner, a data source you trust and a number you can measure today. Without a baseline, after all, a pilot can only report that the demo worked. Our AI proof of concept guide shows how to scope this step.
Step 2: Start With the 30-Day Trial or a Free Tier
SAP’s 30-day basic trial includes AI Launchpad for designing and managing LLM prompts, hands-on vector RAG search and a playground. Alternatively, a free-tier instance is the other route, and it can be upgraded later without losing work.
Step 3: Add the Entitlement and Run the Booster
In a BTP subaccount, add the SAP AI Core entitlement, and choose the extended plan if you need the generative AI hub. Then run the AI Core booster, which sets up the service instance for you.
Step 4: Create a Service Key and Connect a Client
Next, create a service key and store it like any other production secret. You can then connect through AI Launchpad, Postman or Python. Also, the SAP Cloud SDK for AI is available for Java, JavaScript and Python.
The key itself is a JSON object, and four of its fields do the work. So when you set up a BTP destination, SAP’s connection guide maps them like this:
Name: my-aicore
Type: HTTP
URL: [serviceurls.AI_API_URL]
Authentication: OAuth2ClientCredentials
Client ID: [clientid]
Client Secret: [clientsecret]
Token Service URL: [url]/oauth/tokenThe same guide also names the error you get when the key is missing: with no service binding and no destination, the SDK stops with “Could not find service credentials for AI Core. Please check the service binding.” For local work, then, paste the whole key into an AICORE_SERVICE_KEY environment variable. Also, escape any $ in the client secret as \$ inside a .env file.
Step 5: Run a First Execution or Orchestration Call
On the custom path, first sync a workflow template and run one execution. On the hub path, instead, send a first orchestration call with a prompt template and a content filter. Then test the output against the baseline from step 1, not against your hopes.
Step 6: Plan Production Before You Celebrate
Finally, decide who owns the endpoint, how you will watch quality and cost, and how you will roll back a bad model. For that, our guides to LLM observability and to moving an AI pilot to production cover those checks.
Watch on YouTube
Can LLM Gateways Make Enterprise AI Safer?
A Kanerika discussion of how an LLM gateway sets one place for model access, cost limits and logging, the pattern that lets coding tools reach SAP AI Core safely.
Using Coding Tools and Gateways With SAP AI Core
Developers increasingly want their everyday tools to reach company-approved models, and AI Core supports that through gateways. For example, LiteLLM supports the hub’s orchestration service with a sap/ route prefix for chat completions and embeddings. It authenticates with an AI Core service key and reads the resource group from AICORE_RESOURCE_GROUP, which defaults to “default” (LiteLLM docs ).
Cline, meanwhile, goes a step further and lists SAP AI Core as a provider, set up with service key credentials (Cline provider docs ). So a developer can point the coding assistant at company-approved models instead of a personal account.
Coding Agents Behind a Gateway
Other coding agents, Claude Code among them, still have no native AI Core connector that we could confirm. Instead, the workable pattern is a gateway such as LiteLLM in front, subject to company policy. Our explainer on the LLM gateway pattern shows how one layer can cap spend, log prompts and enforce model choices.
One caution, however, applies to every option. A service key is a secret with real spending power, so keep it out of repositories and rotate it on a schedule.
Security, Data Privacy, and Governance in SAP AI Core
SAP states that AI Core uses end-to-end encryption, identity and access management through SAP Cloud Identity Services, and audit logging (SAP AI Core FAQ ). It also cites GDPR and ISO 27001 compliance. On data use, SAP says customer data is never used to train SAP models or shared with third-party vendors for training LLMs.
What Stays Your Job
Those controls cover the platform, but your side still owns a lot. In particular, least-privilege roles, masking rules for prompts, the scope of each grounding source and logs that someone actually reads all sit with you. Tools for data masking help, and an AI governance framework then turns these duties into named owners and review cycles.
Watch on YouTube
AI Security for Enterprises: Kanerika’s CISO on Prompt Injection, ISO 27001 & SOC 2
Kanerika’s CISO explains prompt injection, access control and audit evidence for enterprise AI, the controls that sit on your side of a managed runtime.
Even so, the failure modes below show up most often once a workload goes live.
Risk Warning sign Control Owner Sensitive data in prompts Personal data in prompt logs Masking module and field-level rules Data protection lead Over-broad grounding Answers cite records the user cannot open Scoped sources and role checks App owner Leaked service key Calls from unknown clients, cost spikes Vault storage, rotation and alerts Platform team Model drift Accuracy falls against the baseline Scheduled evaluation and rollback ML owner Unchecked spend Token use rising without new users Budgets and gateway caps Product owner
What the 2024 SAPwned Research Teaches
In July 2024, Wiz Research disclosed SAPwned , a chain of vulnerabilities in SAP AI Core. Together, the flaws could have let attackers take over the service and reach customer data. Afterward, SAP fixed every issue, and Wiz reported that no customer data was compromised.
For customers, then, the lesson is about isolation. A shared runtime is only as safe as the boundaries around each workload, so review your roles, secrets and resource groups with equal care. Our LLM security guide also lists the checks worth running before go-live.
When Should You Use SAP AI Core?
SAP AI Core is the right choice when a workload needs SAP context, SAP’s own models or central governance on BTP. On the other hand, it is the wrong first step when the capability already ships inside your SAP applications. In between, then, sits a judgement call about where your data and ML skills already live.
Strong Fit
In general, choose AI Core when the use case is SAP-centred, such as predictions on finance or supply chain tables with SAP-RPT. It also fits grounded LLM apps that must respect SAP roles, and teams that want one governed place for models across BTP apps.
Use What Is Embedded First
Many needs, in fact, are already met by embedded SAP AI and Joule, which arrive with SAP cloud applications. Check those first, because a built-in feature needs no model of your own to build, test and maintain.
When an Existing Platform Wins
Suppose your data and ML pipelines already run on Microsoft Fabric, Databricks or Azure Machine Learning. Moving them to AI Core rarely pays off by itself, so keep the models where they are and send results to SAP through integration. Our notes on MLOps in Microsoft Fabric and Databricks MLOps cover that route.
Situation Choose Why A Joule or embedded SAP AI feature already meets the need Embedded SAP AI No model to build or run Predictions on SAP tables with little ML staff AI Core, tabular models SAP-RPT and TabPFN predict without training LLM app grounded in SAP data and roles AI Core, extended plan Hub orchestration with masking and grounding Custom model that SAP apps must call AI Core, custom path Endpoint close to SAP, governed in BTP Data and ML already on Fabric, Databricks or Azure ML Existing platform, integrated with SAP Avoids a second ML stack Unclear value and no baseline Neither yet Run a scoped proof of concept first
The verdict, in AI Core terms, is therefore short. Use embedded SAP AI first, then build on AI Core when you need your own model or an SAP-grounded LLM app.
How Kanerika Helps Teams Put SAP AI Core to Work
Kanerika works on the parts that decide whether an AI Core workload reaches production, which are the use case, the data and the operating model. Its generative AI strategy work follows the IMPACT methodology, which ties each initiative to a measurable business outcome before any code is written. That outcome test is also where a team learns whether AI Core, embedded SAP AI or an existing data platform should carry the workload.
SAP-Side Engineering With Review Gates
On the SAP side, the technology stack of Kanerika’s SAP ABAP practice includes the ABAP AI SDK and the generative AI hub. It also lists SAP Joule for Developers and the SAP ADT MCP server for governed, agent-based development. Every AI-drafted change still passes ATC checks, ABAP Unit tests and developer review before it reaches a transport.
That practice also runs AI adoption in three steps, which are Foundation, AI-Assisted Development and Governed Automation. The same order suits an AI Core rollout, since standards and baselines come before any automation.
Next comes the data underneath the models. Kanerika’s data integration services handle extraction and quality rules, while its generative AI services design the grounding and build against an agreed baseline. Governance then runs through every stage, from masking rules to model owners, through AI governance services .
Grounded retrieval under strict access rules is also a pattern Kanerika has delivered, though not on AI Core. For a leading global investment bank, it built chat-based access to enterprise data, reporting 43% faster information retrieval and 100% role-based compliance. What carries over is the discipline of scoped data and role checks that any grounded LLM app needs.
Kanerika also holds 250+ technology certifications across Microsoft, Databricks and Snowflake, and it is a Databricks Consulting Partner and an OpenAI Select Partner. So, to scope a first AI Core workload, book a conversation with our team .
Case Study
43% Faster Data Retrieval for a Global Investment Bank
An AI solution that gave employees chat-based access to enterprise data, with 43% faster information retrieval and 100% role-based compliance.
Read the Case Study → Where SAP AI Core Fits in Your Plan
SAP AI Core is the managed runtime for your own models and LLM apps, with AI Launchpad as its console and Joule on top. Also, the generative AI hub needs the extended plan. Costs are consumption-based, so estimate them in SAP Discovery Center and add the data work around them.
Whichever path you take, the first step is the same. Start a trial with one workload and a baseline, and let the result decide what comes next.
Frequently Asked Questions
What is SAP AI Core used for?
SAP AI Core is used to train, host and serve AI models and to run LLM applications on the SAP Business AI Platform. Teams use it for custom machine learning endpoints, predictions on SAP tables with tabular models, and grounded question answering through the generative AI hub, all managed from one governed service.
What is the difference between SAP AI Core and SAP AI Launchpad?
SAP AI Core is the runtime where models are trained, served and called. SAP AI Launchpad is the web console that manages it, showing scenarios, executions, deployments and prompts in one interface. You can also drive AI Core through Postman, Python or the SAP Cloud SDK for AI without opening Launchpad at all.
Is SAP AI Core the same as the generative AI hub?
No. The generative AI hub is part of SAP AI Core, not a separate product. It adds a model library and an orchestration service for large language models, with templating, filtering, masking and grounding. You only get the hub on the extended service plan, while the rest of AI Core also runs custom models.
Which LLMs can I use in SAP AI Core?
Through the generative AI hub, SAP AI Core gives access to models from providers such as OpenAI, Anthropic, Google, Mistral, NVIDIA, Amazon and SAP itself. Specific model versions change often, so check the current catalog on SAP Help Portal before you design around one model, and plan a fallback model from the start.
Does SAP AI Core have a free tier or trial?
Yes. SAP has a free tier and a 30-day basic trial that includes AI Launchpad for prompt work, hands-on vector RAG search and a playground. Free-tier tenants can later move to the standard or extended plan without losing data or work, although the upgrade from free tier to extended does incur cost.
How much does SAP AI Core cost?
SAP AI Core pricing is consumption-based, and licenses come through Pay-As-You-Go for SAP BTP or an Enterprise Agreement. SAP does not publish one list price on its product pages, so use the SAP Discovery Center estimator for the service, then add data pipelines, vector storage, monitoring and engineering time to your budget.
Do I need SAP AI Core to use SAP Joule?
Joule is SAP’s assistant inside SAP cloud applications, and it is a separate product from AI Core. AI Core matters when you want to build your own models or LLM applications on SAP’s platform. Licensing details for Joule and related AI features vary by contract, so confirm them with your SAP account team.
Can SAP AI Core run on premise?
SAP AI Core is a managed cloud service. SAP describes its model hosting as running in public and private clouds, and it does not describe an on-premise runtime you install yourself. Teams with on-premise SAP systems usually connect their data to the cloud service through integration rather than hosting AI Core locally.
What does the orchestration service in SAP AI Core do?
The orchestration service puts one API in front of many generative AI models. It adds modules for prompt templating, content filtering, data masking, grounding, translation and model fallback. So an application can change models or add a safety step through configuration, without rewriting code each time the model choice or the policy changes.
Can I connect coding tools like Cline or LiteLLM to SAP AI Core?
Yes. Cline lists SAP AI Core as a provider set up with service key credentials, and LiteLLM supports the generative AI hub’s orchestration service for chat completions and embeddings. Other coding agents can reach approved models through a gateway such as LiteLLM, subject to your company’s security policy and spending limits.
