SAS turned fifty this month, arguably earning the title for the oldest private AI company in continuous operations, and a private one at that. It may or may not go public in the near future to take advantage of the current market climate to grow faster and in alignment with its ethics, business, and people model. It’s not in a hurry because it has been profitable for all of those 50 years. I would go one step further to suggest this points to something seminal about what might be required to build long term foundation for trustworthy AI.
Of course, this year when people talk about AI, they mean the fancy LLM (Large Language Model) or agentic stuff. Sometimes the adjacent data harnesses get billing as supportive infrastructure to the rockstar, rather than the main event that makes or breaks this year’s darling memes. But most of the people talking in these loops seem to be in a hurry to go nowhere fast. Or at least their feet, exemplified by the direction of their dollars, seem to be moving in a different direction from the words.
I think there is a reasonable argument that SAS has been an AI company since July 1976 when it was founded. And granted, at least some large fraction of those years it might not have been helpful to the messaging as AI seasons turned from summer to winter at regular intervals. But across all of those seasons, their direction of travel has been towards helping evolve tools for making sense of the world in a more helpful way.
Perhaps that’s because SAS started as a group of researchers at Cox Hall at North Carolina State University, where a consortium of southern universities pooled effort on a statistical package for agricultural data to improve crop yields and field trials. They were wondering something rather unglamorous: how to feed a few more people with a little less effort.
It was officially incorporated in July 1976 with four employees. Its first product was Base SAS, consisting of around 300,000 lines of code, and it was leased rather than sold – predating the SaaS model by several decades. The company has been profitable since, has never borrowed, and is still owned by Jim Goodnight and John Sall. Last year, revenue passed $3 billion, and customers numbered more than 50,000 across 150-plus countries, according to the company.
Of course, none of this has ever been marketed as AI, because for forty of those fifty years the phrase would have been a liability. A bank’s model risk committee was more concerned with better modeling that was reliable and trustworthy and consistent rather than the meme of the year.
I put this to Bryan Harris, SAS’s EVP and CTO. He grounds the claim in neural network work that goes back half the company’s lifetime:
So, as to SAS being an AI company, I appreciate your position on that. We agree; we’ve been doing neural networks for almost 30 years, and so we were one of the first ones to be doing this stuff. We wrote some of the early guides on it and actually applied neural networks for customers doing predictive models, so we’re very comfortable with this space, and it also means that we also know what is real and what is not in this space.
Learning from experience
Real trust requires building a team that has learned the hard knocks from level set that occurs when one year’s bright ideas make hard contact with reality. The people now asked to govern AI in regulated industries did not train as AI people. They trained as statisticians, in a discipline with a hundred-year apparatus for saying how confident it is and why.
For example, I chatted with Dr. Iain Brown, who runs AI and data science for SAS across banking, insurance and capital markets globally and is an adjunct professor at Southampton. One of his first projects out of university in the early 2000s was doctoral research alongside a major UK bank, building machine learning models to judge credit risk. That bank did not survive the global impact of the 2008 financial crisis. He recalls:
So I was at university, and there was an opportunity actually to do doctoral research alongside one of the largest banks in the UK at the time, which was Halifax Bank of Scotland. They’re not around anymore. They got acquired as part of the 2008 financial crisis, so they became part of Lloyds Banking Group, which is Lloyd’s now as an organization.
Trustworthy was beginning to take on a new meaning across the industry when he joined Lloyds’ risk department in 2010 or 2011, then SAS around 2012:
That was after the financial crisis, and there was even more scrutiny then from a regulatory perspective on governance and how models were being generated and built. So I was part of that whole world of building more models, but safer models to protect the financial services sector as a whole, not just the banks. So more robust governance around these systems has been, from the outset, a bit of a focus for me because it’s in all of our interests to use these capabilities in the right way. So I’m an advocate of using AI in an ethical way, but also in an interpretable way that we can understand what’s going on behind the scenes, because I think that’s a bit that we’re now moving away from again as a society.
The trust dilemma
SAS, being an analytics company, has the data and supporting analytics models to back this up. SAS and IDC surveyed 2,375 respondents across five regions for the Data and AI Impact Report published last September. The central finding is that 78% of organizations say they fully trust AI, while only 40% have invested to make their AI systems demonstrably trustworthy through governance, explainability, and ethical safeguards.
One of the most interesting findings was the comparison between technologies. Among the organizations with the weakest foundations, generative AI was trusted roughly three times as trustworthy as traditional machine learning. This is backward according to what is practical today, since traditional machine learning is more explainable than newer LLM-based techniques that might excel at telling a more prosaic story. Brown offers:
There’s more trust in the generative systems than there is in the traditional machine learning methods of the past. It seems odd that that would be the case, because a lot of these systems are black box by nature; you don’t understand what’s going on behind the scenes, but there are also risks around the hallucinations as well, but because you can communicate with them, they feel like they understand what you’re asking. As humans, we tend to trust things that we can converse with and communicate with, and that seems to have been the route we’ve now taken culturally as humanity. We’re trusting these systems more and more without fully understanding them.
Determinism versus probabilistic
One of the hardest parts is figuring out where the LLMs sit against everything else. They are prone to changing their output for the same question for reasons that are hard to distill. But the probabilistic models that have powered statistics are a bit different, as the same model yields the same result each time. Harris unpacks this a bit:
I would say there’s just two things here: deterministic and non-determinism, right? Because probabilistic can still be deterministic. Machine learning, a lot of machine learning models, are definitely probability machine-like. They predict out of probability, but they’re deterministic. If you give them the same inputs, they produce the same outputs. If the problem with large language models is about non-determinism, it’s that there are probabilities, but the space in which it samples is non-uniform at times, or not consistent, and so you can get significant differences based on subtleties of the prompt itself. One word change, or a couple phrases of a prompt going in, can drastically change where it lands, has significant impacts to the changes of where it lands in response, and that’s one of the biggest issues that we’re all trying to deal with in the market for AI and agentic AI.
The non-determinism that makes these models good at mimicking creativity is the same property that can cause a compliance failure. Harris says:
It’s a function of the complexity of the models themselves. I mean, they’re just complicated things, that’s what makes them put it this way, what makes them creative. The reason you can get creative writing out of them is because of this non-deterministic nature.
Brown, coming at it from the statistics side, describes what gets lost in the translation:
From a statistics perspective, we’re very confident. We like to know the facts, and we like to have confidence, and we know that something’s within a 95% confidence… The problem with these large language models, potentially, is it’s all presented as this is accurate, this is the answer, with no degree of confidence, or no degree of, well, this might not be the case, or this is a best guesstimate on the information you provided… and particularly when that comes down to consumer, and we have very heavy regulations in that, in the country we’re in, around consumer duty, being able to present exactly why a decision was made to a customer, and being able to defend that.
Harris brings a willingness to sit with probabilities when describing the current state of tools. This level of candor feels like a precondition to building trust, which is strangely lacking from most vendors in the industry. Here’s one example from our conversation:
So, at best, a simple agent today, at best, state of the art with five different tools to orchestrate for an agent. So, you give it a prompt, it tries to select the right tool and execute that tool as part of an inferencing workflow. It is 80% to 90% accurate, something in that range. So, by default, you have a 10% error rate in a simple agent today. If you go to multiple multi-agent orchestration with five to ten tools that are registered to each agent, that accuracy rate falls somewhere between 50 and 85% depending on the workflow, so there’s just inherent error in these things.
He compares the current state of how LLM tools might affect things we actually care about to how you might prioritize the value of AI in a new car. It would be like telling you that your brakes work 100% of the time. But after AI, it would be like telling you they work 90% of the time:
I mean, no one’s excited about getting the AI-enabled car with brakes that fail 10% of the time. That’s where the agentic market is right now. It is exciting, but it needs to be treated properly to make sure we address it, especially in our world, where our biggest industries are banking, life sciences, insurance, and government, public sector. Every one of those industries has high-risk decision-making that cannot just sit there and be right sometimes.
One fix is to remove the discovery problem rather than improve the discovery:
We can reduce this this error rate to zero by allowing customers to design more directed agent workflows that allow them to kind of specify the path, because they already know what the business process is, they don’t need the agent to discover the business process, most people know how to automate something, and so there’s no reason for an agent to be discovering how to make a call to an API.
The ownership question
At the firm’s recent SAS Innovate event at Anfield soccer stadium in Liverpool, copies of a May 2026 Forbes article were on display with a picture of Jim Goodnight on the cover under the headline ‘Private Matters: James Goodnight’s SAS is profitable, private, and 50. Can it get to an IPO?’ as if that should somehow be the main aim of all companies. And this is not the first time the idea has been publicly teased. In 1995, Goodnight told The News & Observer that running a public company would “ruin my life.” In 2000, SAS said it would list within 12 to 18 months and engaged Goldman Sachs and then the dot-com collapse intervened. In July 2021, days after breaking off reported talks to sell to Broadcom for between $15 billion and $20 billion, it announced an intention to be IPO-ready by 2024.
This year it might, but it does not seem to be in a hurry. When I asked Harris about the Forbes article and his take on how an IPO might shift company culture for better or worse, he responds:
While certainly we’ll get an injection of capital by going towards an IPO, we want to be able to choose who we do that with, with the idea that we’re going to make sure that any investor understands our value system as a company. Dr. Goodnight would be very key, obviously, in the decision-making of ensuring that he trusts that if we were to go towards that and go and take investment, and that know that this is who we are, this is how we’re going to operate, and this is how we’re going to prioritize our execution as a business.
Most of our competitors that are going IPO have to make short-term decisions because they’re going to run out of money. We don’t have that problem. When you’ve been profitable for 50 years, you are in an incredibly financially strong position, so you get to actually define the terms a bit for yourself, because we don’t need anyone else’s money… It’s been controlled growth for us. It means that we don’t overhire and we don’t overfire, because we think about controlled growth and we think of that as providing a much more stable culture for people, which they love working here.
My take
Some of the largest players pioneering AI these days talk a lot about what they are doing as safe AGI that will be good for humanity. But then every week there is some new version of their actions, inactions, or negligence not quite aligning with that grand vision in some way or another. I don’t think the problem is the people per se; it’s just that they are too much in a hurry to pay attention to what they are really calling into being in the world.
There can be real value in taking the slow, steady route across all of the ups and downs that life brings. In one sense, SAS sells models and tools for making sense of them. But in another, it also sells the institutional harnesses around them, like confidence intervals, lineage, and the ability to sit in front of a regulator, customer, or others that might be impacted and say why.
That’s a rare talent for a company of any shape or size this year. Here’s to hoping for another fifty years.