Theresa Bui on SymphonyAI Eureka

<p><strong>Russ Fordyce (00:01)</strong><br>Welcome back to the winner circle, everybody. I am here today with Theresa from SymphonyAI. They have just cleaned up in the AI Excellence Awards, having been made a finalist last year. This year they have won three awards in the AI categories under products. So Theresa, congratulations on all the wins and all the success you're having.</p> <p><strong>Theresa Bui (00:21)</strong><br>Hey, thank you so much. This has been a great cap to a great year for us. And it's a crowded year with AI because it seems like every company is an AI company. And so I think to be able to be acknowledged for actual operational AI and what we're doing at scale — at our customer sites around the world — I think is a real coup to our team. We're so proud.</p> <p><strong>Russ Fordyce (00:54)</strong><br>Yeah, and I said you guys are actually putting it to work. That's what I really loved about seeing what your nominations were about — you're really building that foundation that allows enterprises to sensibly put AI to work. Before we get into the guts of all the pieces, walk us through SymphonyAI and what the company's all about.</p> <p><strong>Theresa Bui (01:23)</strong><br>Yeah, so we are a pure play AI company. We're private, running at about $500 million in revenue across 40 countries, 2,000 customers around the world. The top 25 CPG companies in the world use us, the top 30 banks in the world use us, very large consumer product manufacturing companies use us. And what makes that possible at SymphonyAI is we call ourselves a vertical AI company — and that's not a marketing term. Most enterprise AI is horizontal, meaning you bring in a platform and spend the first nine months training their agents and models on your industry and company. We come in with product platforms in retail, financial services, manufacturing, enterprise IT, and media — and we already have prepackaged the agents that understand your use cases, the ontologies that understand how your industry works, and the models that are already trained on data specific to your industry. It's not trained on the internet. It's trained on how a manufacturing plant works, how a bank works. That's why we can stand up our solutions inside six weeks and start running your use cases — not wait nine months.</p> <p><strong>Russ Fordyce (03:46)</strong><br>Yeah, and you guys won for this Eureka platform. From what I was reading, it's kind of three-layered — a knowledge graph, then an orchestration layer, and then governance. Walk us through what a typical deployment looks like for Eureka, say, in an industrial plant.</p> <p><strong>Theresa Bui (04:20)</strong><br>Yeah, let me talk about Eureka and then give examples of how that plays out. The foundation of our Eureka technology is threefold. First, shared context — we build a domain knowledge graph for every vertical, a map for how a business in that industry actually works. We already come in understanding it; we just need to map it from your data to our knowledge graph. Second, adaptive orchestration — we assign the right AI tool for the right process. We'll use a predictive model if we're trying to do detection. We'll use a language model when you need explanation. We'll use rules-based logic when you need an audit trail. It's not one set of technology — it's adapted to the workflow you're trying to build. Third, governance from day one. Every decision point is logged: what went in, what model ran, what action was taken, what happened. That audit trail is what makes our AI trustworthy.</p> <p><strong>Russ Fordyce (03:46)</strong><br>In an industrial manufacturing context, what does a real Eureka deployment look like?</p> <p><strong>Theresa Bui (06:42)</strong><br>In our industrial manufacturing business unit, we have a product called Iris Foundry. Imagine you have a bottling line and there's a worn capsule — you get an alert from our system. A second agent checks if that part is in inventory. A third agentic AI workflow creates a work order to order that part. A fourth agent looks: there's a 45-minute window between production runs — let's schedule a maintenance window. The plant manager gets that work order in Microsoft Teams in 10 minutes. Previously that would have taken one or two days of root cause analysis, checking systems, creating the work order, checking the schedule. That's true orchestrated agentic AI at work.</p> <p><strong>Russ Fordyce (10:08)</strong><br>Yeah, and I remember reading in the nomination you branded it as perceive-reason-act — giving that agent a path to go find a solution. Because without governance and guardrails, AI will run down a wrong path before you know it.</p> <p><strong>Theresa Bui (10:48)</strong><br>Yeah, what's interesting too about orchestrated AI is that it's learning and ours always has human in the loop. You have a tolerance threshold with Eureka for how much human-in-the-loop you want. Take SymphonyAI Risk Intelligence, our platform in financial services — it helps level one, two, three investigators do transaction monitoring, sanctions monitoring, know-your-customer, etc. A typical investigation might take 200 minutes manually. With AI, we'll reduce that by 60-70%. But companies at the beginning put that human-in-the-loop slider very high — every time we flag something and make a recommendation, a human looks at it and says yes or no. Over time, as they get more comfortable with the AI and find the human is agreeing 100% of the time with the AI's decision, they lower that threshold. And in the background, if a human is overriding a decision, the AI is noting that for next time.</p> <p><strong>Russ Fordyce (13:33)</strong><br>The thing I also read in your nomination was all about AI sovereignty. That seems to be popping up everywhere now. Walk us through what AI sovereignty means and why it's so important when deploying in regulated environments like banks and healthcare.</p> <p><strong>Theresa Bui (14:11)</strong><br>Yeah, at the 35,000-foot level, some people interpret AI sovereignty as where your data centers are geographically. But when I talk about AI sovereignty, I really mean: who controls and who owns your data? Think about the last 20-30 years of SaaS — you put your data on SaaS platforms, but you're running the same workflows as your competitors. You're not doing anything new or different. And with AI, you can now literally mimic workflows from public data sources in an hour. What companies really want is a wedge that is truly competitive. And where that wedge really is: you should own your data in an AI environment, whether that's in a private tenant or on-prem, so the AI models are running just on your data, learning just on your data. In our industrial business unit, we take it even further — we won't train our models on your data unless you give us permission. This is truly your IP. This is sovereign AI data that you own, learning on your own data that no one else has, in your own custom environment, with your own specific workflows. And it's compounding every day. That's what creates the competitive moat.</p> <p><strong>Russ Fordyce (18:05)</strong><br>You're on the front line of enterprises deploying AI. This is the year where they're actually implementing — last year was investigate and test. What's the biggest misconception they have when they first approach AI?</p> <p><strong>Theresa Bui (18:36)</strong><br>I think companies underestimate what it takes to scale AI. What it takes to be successful in a pilot in one plant on one fill line is very different from what it takes to be successful in 200 plants where each plant has 20 fill lines and there are 10 different flavors of process optimization plus five different predictive asset management use cases. And each plant, even though it's the same company, has totally different configurations of their equipment and data stores. That's number one. And the foundation of what it takes to scale is the normalization of your data across all those plants — using the same ontologies and the same way of thinking about the data across all 153 plants, without it turning into a five-year data normalization exercise.</p> <p><strong>Russ Fordyce (20:48)</strong><br>Yeah, the only thing that relates to CEOs when faced with this is a lot of them have outsourced customer service to another country — you're effectively doing the same thing. And I know when I went through that it was very painful, because you had to document every process you'd never documented before, and the people learning them didn't have the context.</p> <p><strong>Theresa Bui (20:48)</strong><br>Thank you so much. Yeah, thank you so much. I appreciate it, Russ.</p>