SAP AI Use cases and details of AI Launchpad and AI Core Services - Part 3

In this article we will take a closer look at both AI Launchpad and AI Core service – but before we get there, Abhijeet Kulkarni would like to share some of the possible use cases of AI services that he has come across.

Use case 1: Speeding up Integration.

One of the most common challenges faced by IT teams is identifying the most fit-for-purpose integration for a business requirement. While the architecture guidance and decision trees are helpful, those are never perfect. Often, teams need to fall back on expert help - i.e., a person - as the defined guidance – i.e., a document – is not comprehensive enough to appreciate the nuances unique to different business requirements. Note that each organization has unique set of principles and operating constraints, which makes generic guidance unsuitable. Essentially, this person-based approach slows down decision making and thus, delivery. Worse, it introduces biases and variations as decision makers change.

Introducing AI in this scenario promises tangible results. AI trained on variety of integration requirements of that organization, along with the pattern and technology chosen to meet it, far exceeds the capability of a single person – or a small group of decision makers – to arrive at the most ideal solution. Manual decision making requires factoring in numerous parameters, ranging from features to budgets to approved patterns to end-of-life considerations etc. A well-trained AI model can arrive at the decision in fraction of the time compared to a human. Further, any number of permutation and combinations of the parameters can be presented to this model to obtain multiple potential solutions, while keeping the time element flat. AI tools from reputed vendors are explainable, which adds to the credibility of the proposed solution and makes it more durable.

While the above use case is not readily available for exploiting, there are several that are. GitHub copilot is one such tool that is used by developers to aid and assist in building code. SAP released its own tooling under the branding of SAP Build Code that uses it Generative AI assistance Joule to provide working solutions – not just bootstrapped or boilerplate code – based on tailored prompts. The operative word here is tailored. At the time of writing, it is barely a quarter since Code was GA. It is a great start, and coming quarters should bring better outputs from the tool as its model is trained and tuned further.

Use case 2: Simplifying reporting.

There is another use case would, perhaps, resonate with everyone. The current practice of gleaning information from a system of records is by building reports. Hundreds of reports are built using variety of tools by extracting data at periodic intervals, or sometimes live connections. There are two major challenges in this approach. First, building reports take lot of time, and in many cases, still require interpretation or correlation with other reports.

Second, the way to query the information is not intuitive. The user is expected to tweak options to get a specific report from a specific system. This is very typical of a server client way of working. A very rigid pre-determined querying on a system to get pre-formatted response. That is not how boardroom discussions, or even small decisions take place. In discussions questions are ad-hoc and impromptu. The current way of reporting is not suitable to support such discussion and thus decision making is slow, and often inaccurate.

If, however, cleaned, and structured data is fed to a pre-trained Large Language Model and Retrieval-Augmented Generation (RAG) technique is applied on it, it opens the way to query the data in a more natural way. Instead of static reports, the information is scoured from vast collection of structured/unstructured data sources. Information from multiple sources can be poured into an evolving model.

Can it answer everything? Probably not, at least initially. But it is lot intuitive to use, will support ad-hoc querying and keep the complexity away from the end user.

With these use cases in mind, let’s get into how you can set them up with help of AI Core and AI Launchpad services.

AI Core is the service that is designed to handle execution and operation of your AI use case by connecting the foundational model to your SAP data and the front end of your choice (e.g. Fiori UI App). AI core can be deployed on Kubernetes or CF or Kyma environment and allow deployment trained ML model as a webservice which can be consumed by web-apps to make inferences. It allows model training, metric tracking, and model deployment like an API lifecycle process in a multitenant environment. AI API would be combination of AI Artifacts and workflows such as training scripts, org data, models, and model servers.

AI core is where the developer would register the Docker registry, synchronize AI content from git repository, and register object store for training data and trained models to be then served to consuming Apps. ?AI core is available in 3 plans – Free, Standard and Extended.

The key difference between standard and extended is the availability of SAP Generative AI Hub. For standard plan use of resources for custom workloads is charged using adaptable pricing according to the resource type (Storage and Compute) used, with the addition of a baseline charge for service use.

For the Extended plan, the costing is based on use of provided generative AI models and is charged using adaptable pricing according to model choice, and the number of tokens sent and received. AI Core service can be used in combination with Generative AI Hub, but it is not part of the normal usage metrics to measure costs.

SAP AI Launchpad is the next key service which allows management of different AI use case runtimes (e.g. AI Cores) and connects the Generative AI Hub into the solution. AI launchpad helps to connect one or more AI core runtimes, manage lifecycle of the prompts, generate endpoints for the AI services, track computer needs, and authentication of the AI APIs. Think of it as a management console to keep a track of all AI core runtimes.