Articul8 CEO Arun Subramaniyan on Why AI Agents Must Use Domain-Specific Languages

Hello and welcome to the latest edition of the Techstrong AI Leadership Insight series. I'm your host, Mike Baar. Today we're with Dr.

Arun Ian, who's the CEO for Articulate, and we're having a little chat about how AI will go into various vertical industry segments as we kind of develop more domain specific languages. Doctor, and welcome to the show, Mike. Thank you so much for having me.

All right, so what will it take to accomplish this? 'cause I think if I look at most of the solutions that people are using today, their general purpose, they're horizontally oriented. And yet in each vertical industry segment there's different nomenclature, different terms, different workflows.

So how do we kind of flip this on its head as it were, and make this more vertical industry friendly? So actually, if you think about it, like the models have gotten significantly better, right? So the general purpose models, uh, today I would say have gone from being a middle school student to a very competent high school student, sometimes even, uh, a competent, uh, uh, college graduate.

However, the nuances of individual domains, as you said, are completely lost. And, um, like what I mean by that is you can take, um, um, say even a high school student, give them a textbook on something very complicated. Tell them that the answer to the question lies in the textbook and go find it.

As long as they know how to go look at the indices, they can go do the table of contents, figure out where mostly things are, they can go retrieve something for you, maybe even the right answer. But then when they do that, they don't know whether this is the right context. They suddenly don't know whether it's the right answer and the person consuming it has no idea whether the depth of the information has been considered before giving you an answer.

And that's what we mean by domain specificity versus generalist giving you an answer, right? Mm-hmm. And, um, the, the flippant example I can give you is if you have a question about, uh, say brain surgery, you go ask a brain surgeon.

The brain surgeon looks at a textbook, analyzes uh, the information you're giving them, and then gives you an answer. A high school student looks at the same textbook, gives you the same answer, which answer would you prefer? And hands down, it's a very obvious question, but when we go and ask a general purpose model that is designed to understand pretty much everything from a 5-year-old asking to give you a, a nursery rhyme to a, uh, financial analyst going and doing deep financial analysis of a particular investment thesis, the answers are more or less plausible, but they're anything but specific and most of the time anything but accurate.

So that delta is what we are trying to to navigate, right? I'll give you another example that may sound very und. So take an example of a case where you wanna understand, um, a table, but then not just understand a table, but you want as a financial analyst, or say you are a safety engineer in a plant somewhere, and you want to know that if this system looks at a table, it reproduces the table precisely every single time you look at the table.

Right Now, there's no computing system on the planet that will give you 100% reproducibility. 999% repeatable, you needed to have tested your algorithms, your systems at least 10 million times across a variety of different tables. That's what we mean by going from a general purpose system to a system that works in an enterprise in a responsible way.

Mm-hmm. How does that kind of get worked through? And I'm asking this question because a lot of the answers that you get from AI are probabilistic.

Yes. And you know, that generally means they're right. Some percentage of the time that's less than a hundred.

Yes. And a lot of the things in a vertical industry workflow or deterministic, right? Yes.

They're supposed to be done the same way every time. I'm exactly the same way. And precisely, um, how do I marry these two things together to come up with something that is, you know, where one plus one equals five or better.

Yes. So actually it is not really as much of, uh, a conflict as, uh, it seems on the outset, right? So for example, when you ask a language model to make a prediction, it's really only making a prediction of a probability.

There is no certainty there at all. And each time you ask it, depending on what, uh, is going on in the system, is going to give you a slightly different answer. However, you almost never put just a model to get you to an outcome.

In fact, it's always a system that gets you an outcome. Even today when somebody uses, uh, Chad GPT or uh, say cloud or Gemini three, they're not really just hitting a model unless you're hitting a model API, if you're using the application, you're hitting a system. So you ask a question, the system interprets your question before it hands it to a model.

And after the model gives it an output, there is a series of deterministic tools that interpret the outcome before you ever see an answer. So it is a combination of deterministic tools with probabilistic models that make the overall system determinist. And many times the final answer is something that is retried many times over.

Mm-hmm. And the user may never actually see the retries in between. Right.

And are we essentially also using multiple AI agents to check the work of each of the AI agents or the models or whatever it may be so that we're not overly dependent upon the single guess of one? Right. That's absolutely essential because, uh, like so we've, uh, developed a system called Model Mesh, which has now evolved into Asian mesh.

And this was developed more than three years ago from a concept standpoint, at a time when it was not at all obvious to people that you needed more than one model to get you to an outcome. In fact, the the overall narrative was it is a model war, one model or two models are going to win. And we went into this knowing that you need a combination of different models to work together at runtime to get you to an outcome.

And second, even at runtime, your judge cannot be the, the same family as your player. That's what is the main differentiator. Because even if you say multiple agents, one is generating an answer, another agent is actually critiquing it, the agents cannot come from the same lineage of models.

If they do, you more or less going to go to the average. You at least need the judge to have an independence beyond just saying an instruction set. Right.

And that one, we've seen it multiple times over. And in fact there is a, a recent paper from Stanford that quantified it. They called it the semantic collapse, which is once you get to about 10,000 documents, if all you're doing is vector search, almost all documents look similar to all of the documents.

Ir respect to what question you're asking. Hmm. That's why most of these kinds of question answer systems work really well.

When you have a few hundred documents, you go to a few thousand, they start getting fuzzy. And then beyond a few thousand documents, they're completely fuzzy. And that's the reason why, Um, as you kind of think about this for a minute, we've had domain specific languages for a long time, but they were, shall we say, arcane and only a few people could master them.

And now it seems like we're gonna put a natural language interface in front of them, which then makes them more accessible. Yeah. So does that mean we'll see a, an explosion in r and d around domain specific languages?

'cause I think a lot of people didn't go build those because they were like, well, only a handful of people could use it, but now maybe everybody can use it. So actually it is the, the other way around, right? So where before you needed a domain specific language to describe a set of actions and activities related to the domain and only the advanced practitioner.

So that domain would even use that. Today, pretty much the interface is natural language and the translation layer is what is missing. Because say you ask an aerospace engineer about, um, what is lift compared to a manufacturing engineer?

What is lift? It's the same English word means two completely different things. And the context might also be relevant in terms of what you're talking to them.

But knowing that context and knowing the domain makes a difference between giving you the right answer versus giving you the absolute wrong answer. Right. And to your question of do I need to learn, uh, the domain at all?

Do I need to know the nuances of the domain? That's where the unlock is because today the finance person who doesn't understand the industry walks into a new company, they will have to learn the lingo of the company and the industry before they can actually like function properly. Tomorrow.

When you have these domain specific models properly deployed, they can ask the same question, the answer will be contextualized to wherever they are. Mm-hmm. That's really what is important.

Mm-hmm. So what is your best advice to the leaders of various vertical industry segments, whether it's manufacturing or finance, about how to infuse AI into their workflows? I think everybody's kind of having the same issues.

They're like, we love the idea and the concept and we see the potential, but when it comes to the, uh, the actual execution, everybody seems to struggle. Yes. So I'm, I would say more than advice, I would say I would have three recommendations, right?

So first and foremost is we are well fast. The question of is AI useful? If you are still doubting, if AI is useful, you'll really have to look at people who have actually made it useful for themselves.

Now you can ask the question, I can see them finding use, I can't see it myself. Right? That's where the gap is and that's the gap that you are, um, that highlighting.

The second point to that one is if you accept the fact that we do live in a different world and you don't have to take it from uh, a, it's a self-serving answer for me because I'm running an AI company, you can take it from the fact that we live that daily. We can't operate if we don't find use from ai. That's number one.

If second one is the general purpose, AI systems that are out there are necessary but not sufficient to get you to an outcome when you are running a deep industry. What I mean by that is take a time when email didn't exist. Take a time when intranet in a company didn't exist the first time a few companies started implementing those, those were actually significant differentiators in terms of how they operated.

They got more productivity than their competitors. But very quickly, everybody caught up Today, nobody would say having email is a differentiator, but everybody would say, not having email is a blocker. That's where we are going with general purpose ai.

So not having AI in your company in a safe way that everybody can use would significantly be a blocker. Mm-hmm. Okay.

Excuse me. But the third one is you just having AI is not gonna make a differentiation. What is going to be different is what do you do with your own know-how, with your own data to improve your own operations that only you can do?

And that's where domain specificity comes in. We call that hyper-personalization. How do we take AI systems that are either domain specific or general purpose, make it specific to you and take advantage of that.

Now, without that, everybody kinda sort of looks the same. Like today, if somebody sends you a one page document that looks well written, not that much of a differentiator, unfortunately, unless you can very clearly see the writing is original. And unfortunately we are very quickly going to a point where original writing is very rare.

Hmm. So as you kind of put all this together though, um, who is gonna take the lead on these projects? 'cause to your point, everybody and his brother is gonna have the same kind of capability Yes.

On the tooling. Yes. So who should be at the forefront of these initiatives?

So Mainly business leaders who actually have to show business outcomes, right? So this is not a, I have a technology, let me try to find a use for it. This is about saying, look, I have a business goal today.

I cannot meet that business goal. Most of that business goal is about increasing your revenue, increasing your margins, whatever that might be. Because you have to separate things into bottom line and top line.

And I look at, uh, how leaders have to implement this as bottom line has a ceiling. So you can only get so much productivity outta the system. Top line actually is significantly improving the business.

You cannot just do one, you'll have to do both. And it is about where the businesses at that point in time, in terms of what you do first. But not doing either will be a significant disadvantage.

Mm-hmm. Now I'll give you some examples, right? Take a simple case of, okay, if you are not somewhere like giving your, uh, people the ability to automatically summarize meetings automatically, like identify action items automatically send that into your systems today.

That's a significant disadvantage because not only that you are having somebody in the, the meeting figuring out how to take notes and then figure out where the the action items come from. But also the way to track it manually is always gonna be slower than what the systems can do today. Today it's trivial.

Whether you use Zoom, whether you use teams, anything, it's nearly trivial to be able to use it. But the number of companies who don't use it still surprise me in a safe way. Right?

That's table stakes. What is that one thing that you see organizations doing today that just makes you shake your head a little bit and say, folks, we need to just be a little bit smarter than that. So if it's one thing, it's still a significant portion of, uh, enterprises, especially operating like we are still in 2024.

What I mean by that is they're assuming that the world is very similar to what it was in 2024 and 2025. Um, I can tell you working in an environment where every two days what we thought was our significant differentiator becomes stable stakes. It, it feels like every day is something new.

But even in very traditional industries, like we operate in manufacturing, we operate in energy, we operate in oil and gas in aerospace, but traditionally a change like this would've taken 20 years to go through. We are seeing customers change in a matter of weeks. And they're not changing because it's a fad.

They're changing because they can get to an outcome that they can measure that is fundamentally different, right? And the number of, uh, enterprises that we even show saying, look, this is what is happening out there, and we can actually show you evidence to look at that from a distance and go, but that's not me. Uh, I, I won't get affected by that.

I will continue to operate the way I am operating. The, the tide for that is going to change significantly in 2026. Alright, cool.

Just because the tools are out there, Folks, you heard it here. Change is coming. It's already here.

The only issue is figuring out how to operationalize this whole new set of technology in a way that gives you something that looks like a competitive advantage. But don't assume that whatever you created is gonna be there for long. 'cause everybody else is gonna do the same thing really quickly.

Hey, doctor, thanks for being on the show. Thanks a lot, Mike. Thank you so much for having me.

All right. ai Leadership Insight series. You can find this episode and others on our website.

We invite you to check all those out. Until then, we'll see you next time.