Watch this Ask the Expert session to explore how the integration of AI into SAS Viya enables users to build and improve modeling workflows.
Watch the Webinar
You will learn more about:
-
How SAS Viya Copilot makes modeling, coding and building reports faster and easier.
-
What you can do with SAS Viya Copilot in SAS Model Studio versus code.
-
How you can get access to SAS Viya Copilot in SAS Model Studio.
The questions from the Q&A segment held at the end of the webinar are listed below and the slides from the webinar are attached.
Q&A
Does that mean SAS uses Microsoft Copilot LLM embedded in the Viya environment?
Not exactly. SAS Viya Copilot is SAS's own AI-powered assistant. It's a separate product from Microsoft Copilot. However, it does use large language models provided by Microsoft, specifically OpenAI's GPT-4.1 models hosted on Azure AI. SAS has built its Copilot experience on top of those models, tailoring it specifically for SAS Viya workflows like building machine learning pipelines in Model Studio. So, the underlying LLM provider is Microsoft/Azure, but the assistant itself is SAS-built.
I arrived a little late so this may have already been answered--did SAS work with Copilot to train the AI model used in Viya?
Yes!
With all the volume of whitepapers across user groups from global to the local level, that produces a sizable source of training data.
Sure. A large portion of our training data comes from SAS documentation, but we have also had development teams within SAS contribute dedicated training data from procedure experts. Since SAS PROCs can become complicated due to the number of options they support, the developers of those procedures have supplemented our training data to ensure we have the most comprehensive dataset. Additionally, as new options are added, we keep SAS Copilot updated on what the new PROCs can do. This will be a key differentiator between SAS Copilot and commercial offerings, as we always remain current with the latest options.
Can Copilot be used to iteratively develop code in a more interactive format? I often use AI to iteratively build longer pieces of code. I don't need it to figure out how to analyze data, which the examples seemed focused on so far.
Currently, as I demonstrated, we have the ability to generate code from comments. As a sneak peek, we are introducing enhanced chat capabilities. This will allow you to directly query the LLM about the generated code, make suggestions, and have it modify the code for you—features that are coming soon. I would also add that this iterative development capability exists within Model Studio. You can use Copilot to help develop potential model pipelines, interpret results, and update configurations as needed.
Do we have Viya Copilot for SAS MRM?
We are offering a variety of SAS Viya Copilots for our solutions, and I can confirm that model and risk management Copilot capabilities are in our pipeline. While I cannot confirm any timing at this point, I will say that this is a priority for us as we expand Copilot’s capabilities. These features will not be limited to the platform—they will be available in the majority of our solutions, including retail. We have just celebrated the announcement of our clinical data and asset liability management Copilot tools. You can expect to see more tools like these emerging throughout the year as we announce them. This is something we are excited about, as SAS Viya Copilot will not be limited to your SAS Viya experience, but will also be integrated into our associated solutions.
One you register a model say gradient boosting, will the same model be useful for different loan applications? Each case might be different with some of them having issues with data quality or different parameters.
I think the best practice would be to train a new model for each new type of loan application. I am assuming, in this instance, that you are referring to two different loan applications processes/types. In that case, I recommend training a new model with data specific to that application. However, if you are using a trained model to predict on a similar loan application, then in theory, yes, you could use the same model to make a prediction on that.
Can organizations tune Copilot to align with internal modeling standards or approved pipeline templates? What about in-house developed SAS Macro Library or PROC FCMP Functions/Call Routines?
Right now, it is somewhat generic, as it tends to follow procedure best practices. However, in the future, we are planning to enable you to specify internal RD libraries and procedures. Since this is still in development, unfortunately, I cannot share much more information at this time. But keep an eye on SAS innovation—it's very exciting, and by the end of this, I'm sure our viewers will be eager to hear about what's coming and to see the latest advancements at SAS.
How you handle Model Distillation in SAS Viya Copilot?
Model distillation is a topic that we are gradually expanding upon. At present, I don't believe there is an easy way to perform model distillation with code assistance, as we do not currently provide the ability to custom train a model. However, this is something we are actively working on right now. The same applies to Model Studio.
It seems to me that the showed examples are extremely simple. Could SAS Viya Copilot be used in a similar way to other LLM tools (e.g. to make SAS Copilot build more pipeline and make them compete them and creating a full report with a single prompt, instead of guiding them through simpler prompts referring to each single step of this process)?
I think this is referring to multi-step processes and building them either with our code system or through model pipeline development. For our code assistant, the current publicly released version of Copilot focuses on individual steps, which roughly correspond to a single procedure call or two. However, the next agentic version will support multi-step, single-prompt execution, enabling you to generate multiple procedures and data steps, and then review them before execution.
Any copilot features integrated in JMP Pro?
Unfortunately, JMP Pro is not something I have a lot of exposure to in my position, so I don't think I'm the best person to speak on that. I believe I can speak for the team here when I say that our main focus is generally on SAS Viya. Our visibility into the JMP organization is somewhat limited, as we primarily concentrate on SAS products. While our organizations are connected, our interactions with JMP tend to be less direct and not at the ground level.
How does the LLM keep data secure?
We are very privacy conscious here at SAS. When you submit your code to the LLM, it is transmitted using secure data channels with encryption. The code is only parsed by the LLM; there is no logging involved, and no history is kept. Once your code has been processed by the LLM and the results have been returned, the data is deleted. Your information remains secure at all times, from both SAS employees and external parties. Yes, it’s reassuring for our users to know their security is a priority.
I was just going to add that we also implement tenant isolation and access controls. Even within the same tenant, your prompts are specific to you.
Will this be rolled out to 9.4 (or maybe 9.5 or a non Viya release? Or JMP?
All I can say right now is our focus is on Viya, and our strategy moving forward with Copilot for 9.4 is unfortunately not something we have announced publicly, at least as far as I am aware.
Peyton, feel free to add if you have more information. No plans have been announced publicly, but there are possibilities for the future. However, we cannot confirm or deny anything in this webinar at this point.
Is there a study guide for Proc CAS?
You can contact Stacey Syphus for more information about the programming courses.
Does "Copilot" = "Microsoft Copilot" -- and can we choose another AI tool?
When we say SAS Copilot, we do not mean Microsoft Copilot, nor do we mean to imply that SAS Copilot uses Microsoft Copilot. SAS Copilot is the name that we give to the AI-assisted features across the SAS Viya platform. Depending
On the usecase, SAS Copilot may use a generic LLM or a finetuned SLM behind the scenes.
Want more tips? Be sure to subscribe to the Ask the Expert board to receive follow up Q&A, slides and recordings from other SAS Ask the Expert webinars.
