At SAS Innovate 2025 in Orlando, SAS unveiled its roadmap for agentic AI, making the case for its position as an organization that has been quietly engaged on clever resolution automation lengthy earlier than AI brokers grew to become a trending matter. The most recent enhancements to its SAS Viya platform goal to assist enterprises design, deploy, and govern AI brokers that mix automation with moral oversight.
Whereas many tech distributors are racing to point out off what number of AI brokers they will spin up without delay, SAS CTO Bryan Harris dismisses such counts as an arrogance metric. What actually counts, he stated, isn’t the amount of brokers however the high quality of their output.
“The metric that issues,” Harris informed AIwire, “is what sort of choices you're working within the enterprise, and what's the worth of these choices to the enterprise?”
How SAS Defines Agentic AI
Agentic AI, as outlined by SAS, isn’t merely about automating duties, however is about constructing techniques that make choices with a mix of reasoning, analytics, and embedded governance. The SAS Viya platform helps this imaginative and prescient by integrating deterministic fashions, machine studying algorithms, and huge language fashions right into a unified orchestration layer. The purpose is to allow enterprises to deploy clever brokers which can be able to performing autonomously when applicable but in addition present transparency and human oversight when the stakes are excessive.
SAS Innovate 2025. (Supply: The Writer)
Udo Sglavo, VP of utilized AI and modeling R&D, described SAS’s agentic push as a pure evolution from the corporate's consulting-driven previous. “We’ve been doing this sort of modeling train for a very long time, however usually it was a one-to-one relationship. You got here to me with an issue, I’d ship in consultants, they’d clear up it, off we go,” Sglavo informed AIwire. “Now the thought is, if you happen to’ve executed this ten, 100 instances for a similar sort of problem, why not take all this IP and put it right into a software program product?”
This shift from providers to scalable options, in response to Sglavo, has been accelerated by rising consolation with LLMs. "There’s been a mindset change. Clients at the moment are extra prepared to undertake fashions they didn’t construct themselves," he stated. That shift has cleared the best way for wider adoption of prepackaged fashions and agent-based techniques.
The Limits of Giant Language Fashions
Each Harris and Sglavo emphasised that LLMs, regardless of their widespread enchantment, are just one piece of a a lot bigger enterprise AI image. At SAS, LLMs are seen as beneficial however restricted elements that should be paired with different types of intelligence to drive dependable, repeatable choices.
The SAS executives defined that in contrast to deterministic fashions, which return constant outputs for a similar inputs each time, LLMs could be unpredictable. “If I run a deterministic mannequin with the identical circumstances a thousand instances, I’ll get the identical reply a thousand instances,” Harris stated. “That’s not the case for giant language fashions.” This variability makes them ill-suited for high-stakes functions the place auditability and management are important. As an alternative, SAS makes use of LLMs the place they excel: rushing up repetitive duties and producing prototype options that people or extra deterministic techniques can later refine.
One instance of repetitive job speedup is in schema mapping, a job that usually requires area information and painstaking handbook evaluate. With metadata as enter, LLMs can quickly recommend column matches and generate code, decreasing a multi-week effort to minutes. Nonetheless, as a result of accuracy can differ, SAS integrates confidence scoring and at all times features a human-in-the-loop to validate outcomes.
In additional superior use instances, SAS has additionally carried out strategies that permit LLMs to iterate on their very own outputs by revisiting earlier steps, rethinking mappings, and difficult preliminary assumptions. This iterative self-checking conduct is a key design precept in SAS's agentic AI framework, the place brokers don’t simply settle for the primary reply however purpose by way of issues dynamically.
Giving Brokers a Objective
The important thing distinction SAS attracts between conventional automation and agentic AI lies in purpose orientation. Quite than merely executing a set of predefined directions, brokers are designed to pursue an outlined purpose and modify their conduct dynamically till that purpose is met. This functionality displays a shift in how organizations are serious about AI, pushed partially by the disillusionment that adopted early enthusiasm round LLMs.
Udo Sglavo, SAS VP of Utilized AI and Modeling R&D
Sglavo defined in an interview what number of enterprise leaders initially hoped that generative fashions would supply a sort of common intelligence the place you possibly can drop in a enterprise downside and get out an answer. As an alternative, LLMs proved greatest fitted to slim duties like textual content evaluation. The emergence of agentic AI, he stated, represents an effort to mix the statistical, machine studying, and optimization strategies developed over many years with the newer capabilities of LLMs and retrieval-augmented information techniques.
On this framework, brokers develop into orchestrators of these instruments. Quite than being explicitly programmed for every step, they’re handed an goal, comparable to rising occasion registration numbers, and are then tasked with deciding easy methods to obtain it. For instance, an agent might generate emails, establish potential recipients utilizing a statistical mannequin, and proceed refining its marketing campaign till an outlined goal is reached.
This type of agent, Sglavo famous, is well-suited for low-risk situations like advertising and marketing campaigns. However when the stakes are greater, comparable to choices about credit score approvals or healthcare outcomes, the method should shift. Human-in-the-loop oversight turns into important, and clear governance frameworks should outline the place autonomy ends and accountability begins.
Governance and Belief on the Core
The SAS executives pressured that agentic AI can’t be responsibly deployed with out built-in governance. SAS Viya contains mechanisms to detect bias, consider equity, and supply full transparency into how choices are made. "We give our prospects perception into when a mannequin is poor," stated Harris. " After which they will make the selection to enhance the info or enhance the mannequin.”
(Supply: Suri_Studio/Shutterstock)
Governance additionally contains controls over how a lot autonomy brokers are granted. That is particularly important in high-risk domains like finance, healthcare, and public providers. SAS contains guardrails that guarantee transparency and lets prospects fine-tune how a lot autonomy brokers are allowed.
SAS additionally emphasizes the significance of localized information sources. Quite than counting on internet-sourced info, brokers could be configured to attract solely from enterprise-specific information repositories. Retrieval-augmented era (RAG) setups allow brokers to entry inside information bases to make contextual choices with out compromising safety or accuracy.
A Market of Brokers Is Coming
Trying forward, Sglavo expects agentic AI to evolve into an open market, the place enterprises can combine and match specialised brokers from completely different distributors. In that future, decision-making can be distributed throughout interconnected agent networks that talk and collaborate utilizing shared protocols like MCP or Google's open supply A2A. This imaginative and prescient additionally redefines how enterprises take into consideration deployment. Quite than transport large monolithic AI techniques, firms will deploy nimble brokers, every with a slim focus however deep specialization.
“It will develop into {the marketplace} of brokers,” Sglavo stated. “As a result of whereas we might say we’ve got the most effective provide chain optimization agent, one other vendor might declare the identical factor. After which it turns into a query of belief, pricing, observe file. Have they executed this earlier than? Are they only a startup that’s good at tech however hasn’t labored with precise prospects?”
Sglavo added that enterprises will need the pliability to pick and mix brokers based mostly on their wants. “You’ll say, I need to use this agent, this one, and this one—and simply deliver all of them collectively.”
A Future Constructed on Accountable AI
Bryan Harris, CTO at SAS
As generative AI continues to seize headlines, SAS is inserting its wager on decision-first AI. For firms in regulated sectors the place the price of a nasty resolution could be measured in lives or billions, the corporate argues, transparency and belief should come earlier than experimentation or scale.
Because the enterprise AI dialog shifts from experimental prototypes to extra sensible, accountable techniques, SAS is staking out an area the place belief, interoperability, and resolution high quality come first.
"You’ll be able to't forestall irresponsibility," stated Harris. "However we are able to provide the instruments that permit you to make the best resolution."