How Generative AI Is Quietly Transforming Traditional Industries
AI That Just Works—Even If You’re Not a Technologist
At our recent GenAI Assembling meetup in the Bay Area—powered by Alibaba Cloud—we gathered leaders from across the AI ecosystem to ask a critical question: What happens when generative AI leaves the lab and lands in industries like real estate, education, legal, and retail? Not in theory, but in practice.
With panelists Nishant Agrawal (Alibaba Cloud), Charles Xie (Zilliz), and Xiao Zhang (Collov AI), moderated by Zhaoyang Wang, the discussion moved beyond technical hype into operational reality.
🔍 Key Takeaways
AI Adoption hinges on simplicity
Non-technical users—from realtors to designers—want tools that work out of the box. AI gains traction when it blends into their workflows, not when it requires prompt engineering.
Feedback loops drive product-market fit
Both Collov AI and Zilliz emphasized how user feedback reshaped their products. In Collov’s case, the shift from rule-based image generation to end-to-end learning dramatically improved quality and user satisfaction. Zilliz’s open-source strategy allowed unexpected use cases like pharmaceutical drug discovery.
Alibaba Cloud sees AI as ecosystem infrastructure, not just tech
With open-source projects like Qwen and ModelScope, Alibaba Cloud isn’t just building models—it’s enabling a broader AI ecosystem. By sharing core tools, they expand access, talent, and adoption. The goal: empower people, not replace them—making work faster, smarter, and more autonomous.
Agentic AI is reshaping the foundation of how retrieval works
Agentic RAG is a new form of retrieval-augmented generation where AI agents break down tasks, retrieve information recursively, and iterate like human researchers. This is driving up demand for vector databases, and signaling the next shift in how we structure enterprise knowledge.
The full transcript (lightly edited for clarity) follows. Whether you’re a founder, engineer, or decision-maker, we hope it offers a practical window into where GenAI is actually landing—and what still holds it back.
Nishant Agrawal: Very happy to be here with such a great audience. A little bit about myself—my name is Nishant, and I’m a Solutions Architect. This is a particularly interesting topic for me because I’m not an engineer in the traditional sense of building products. Instead, I work closely with customers, many of whom are business owners, marketers, or creative professionals. I listen to their needs and then communicate those insights to our product teams who are actually building the technology. I feel fortunate to serve as a bridge—translating everyday needs into technical solutions. Especially here in the Bay Area, as engineers, we often live in a bubble, always talking about the latest tools, vector databases, or agent frameworks. But it’s equally important to understand how this knowledge is perceived and adopted in daily life. So I’m really excited to be here. Thank you.
Xiao Zhang: Hi, my name is Xiao, and I’m the founder and CEO of Collov AI. A bit about my background, I pursued a PhD in Applied Physics at Stanford, where I worked on interdisciplinary research involving AI for science. However, I eventually realized that physics felt too far removed from real-life problems. So, after much reflection, I decided to apply the AI algorithms I had studied to practical challenges, and that’s how Collov was born.
At Collov, we’re building a visual financial model. We’ve introduced several innovations in model architecture, giving our models unique capabilities. For example, our system offers highly controllable visual outputs and ensures multi-view consistency during generation. We leverage these features in industries like real estate and home renovation, enabling AI-powered virtual staging and design to replace traditional workflows.
We’re also expanding into generative e-commerce, combining AI with supply chain operations to streamline processes and boost efficiency. We’ve been operating for just over three years and are currently in the Series A+ stage. I’m excited to share our journey and insights with all of you.
Charles Xie: My name is Charles, and I’m the founder and CEO of Zilliz. Our mission is to democratize unstructured data processing. To that end, we’re building the most advanced vector database system to manage all types of unstructured data—images, videos, documents, audio, and more.
Over the past few years, we’ve worked across a wide range of industries—from cutting-edge GenAI applications to more traditional sectors like e-commerce, real estate, legal services, and education. Our technology helps these industries build intelligent systems and manage massive volumes of unstructured data.
Our most well-known open-source project is Milvus, which we released in 2019. Since then, it has attracted over 10,000 enterprise users globally and surpassed 100 million installations. I’m happy to be here today and look forward to sharing more of our story. Thank you.
Zhaoyang Wang: Before the panel started, I noticed many people chatting about the interaction between AI and traditional sectors. When reviewing the event registrations, I saw that several attendees submitted questions—one of the most common was: What does the current landscape look like when it comes to AI’s interaction with traditional industries?
So, to begin, I’d like to ask all three of you: how are traditional industries adopting AI technologies today? How does this compare to the previous wave of AI, like computer vision? And how deeply has AI been integrated into their day-to-day operations? Maybe we can start with Charles?
Charles Xie: Before I answer, can I ask a quick question—how do you define a traditional business? For example, e-commerce has been around for over 20 years. Would you consider that traditional? What about telecommunications?
Zhaoyang Wang: In the age of large language models, it feels like everyone is becoming traditional. What’s your take?
Charles Xie: That’s a fair point. I think everyone has their own interpretation of what counts as traditional or not. In my view, it’s less about how long a business has existed and more about whether it’s a high-tech or non-tech company.
High-tech companies, with their engineering resources, naturally have an advantage in adopting modern AI—they can train their own models and manage large datasets. But non-technical companies often face more challenges. That, to me, is the promise of AI: democratizing technology so that every company and every individual can benefit.
That’s where the concept of “agentic AI” comes in—AI that acts as an agent on behalf of professionals. You shouldn’t have to be a programmer to benefit from AI at work. In fact, over the past three years, we’ve seen legal firms, real estate companies, and education businesses start to use AI in powerful ways. They’ve developed impressive products—AI for legal services, AI-enhanced education, AI-powered e-commerce—and we’re seeing that trend accelerate rapidly.
Zhaoyang Wang: So you mean that this acceleration is happening because generative AI is now more accessible to everyday users, making it easier to integrate into their work?
Charles Xie: Exactly. If we look back ten years, only a very small fraction of developers had access to modern AI. At that time, the field was dominated by research labs like DeepMind, and we were working with models like RNNs and CNNs. Even among technical companies, adoption was limited—less than 1% of developers globally had access to those tools.
But today, things have changed dramatically. Product managers, designers—people without a technical background—now have the freedom to use AI tools. I’d estimate that the reach of modern AI has grown by over a thousand times in the past five years.
Zhaoyang Wang: So coming back to my original question, it sounds like the boundaries are disappearing. Every company is becoming an AI company. There’s no such thing as a “traditional” sector anymore. Okay—let’s hear from Xiao next. I know many tech companies have been trying to enter the real estate industry over the past decade.
Xiao Zhang: Compared to the previous discussion, I’d say our customers are very much in the “traditional” industry category. Many of them are realtors, interior designers, or furniture manufacturers. They’re typically not familiar with AI tools—especially generative AI. In fact, many of them hadn’t even heard of ChatGPT or OpenAI. Our average customer is between 40 and 60 years old, so that’s the general demographic we work with.
This creates both challenges and opportunities. The challenge is that we have to invest a lot in market education. We can’t assume users already know how to interact with AI systems. Another big challenge is product design. Since our users aren’t familiar with prompt engineering or AI experimentation, we have to build tools that are nearly perfect—intuitive, reliable, and tailored exactly to their workflows.
They don’t want to spend time learning how to craft better prompts or explore different AI use cases. They want a simple flow—for example, take a photo, and then see an immediate, high-quality result. That’s the bar we have to meet. The upside of this, though, is that it gives us a strong product moat. So we’re not as easily disrupted if, say, OpenAI suddenly launches a new feature that could otherwise wipe us out.
Let me share an example related to product-market fit. We’ve spoken with a lot of other AI founders, and one case that stood out was the team behind Cursor, an AI coding tool. Originally, they had steady growth—but then they developed a feature where users could write part of a function, hit “tab,” and the tool would accurately complete the entire line of code. Once they perfected this feature to meet customer expectations, their growth rate accelerated.
We had a similar experience. At first, we used open-source models and APIs, and tried various fine-tuning approaches. But the generated images often had flaws—the room layout would change after generation, and the items didn’t look photorealistic. That meant our tool couldn’t be used for real estate staging or interior design. Once we crossed that “customer satisfaction line”—meaning the generated rooms kept their structure and the visuals became photorealistic—our adoption rate surged, and so did our subscription numbers.
So yes, it’s both a challenge and a big opportunity for us.
Zhaoyang Wang: That’s interesting. In terms of market education, it sounds like your users aren’t even starting from ChatGPT—they might not know it or care about it.
Xiao Zhang: Exactly. While some of our users are aware of ChatGPT, many aren’t comfortable using AI tools in general, which presents both obstacles and opportunities for us.
Zhaoyang Wang: Very cool. Now, Nishant—Alibaba Cloud works with a broad range of customers. What does the landscape look like from your side?
Nishant Agrawal: I want to respond to Charles’ question about what defines a “traditional” industry. My perspective might be a bit different. One thing we learned in business school is that a company isn’t just a product or a logo—it’s a collection of people, a culture, a history. A company could be as old as ExxonMobil or as new as something a founder in this room launched today. What sets them apart is the accumulated infrastructure, data, and decision-making patterns.
So in my view, “traditional” is not about age, but about the velocity at which a company can adapt—how fast they can launch new products or iterate on existing ones. If you have a large employee base and a massive customer footprint, you can’t just pivot overnight. Change takes time, coordination, and cultural buy-in.
A great example is BMW. A few weeks ago, Alibaba announced a partnership with them. Tesla changed the market by introducing a minimalist design and software-driven experience. Now BMW is adapting by developing its own AI copilots—but they carry decades of legacy manufacturing and design decisions. The challenge for them is not just technology, but aligning a global organization to move in a new direction while staying true to what their customers expect.
At Alibaba Cloud, we see ourselves as ecosystem builders. We provide the infrastructure and tools, but we also help bring together regulators, industry players, and customers. For example, within Alibaba, our Banma team collaborates with regulators in China to define self-driving standards. Making a car isn’t just about the software—it involves thousands of suppliers and tightly integrated ecosystems. We help enable that.
Zhaoyang Wang: That’s a great point—When it comes to AI and traditional industries, there are two sides: how tech companies build tools, and how traditional businesses adopt them. I’ve noticed that once you give users a tool, the way it’s used is no longer defined by the maker. Your customers take over. They bring their domain knowledge and end up creating use cases you never anticipated.
Can you share any experiences where your customers used your product in ways you didn’t expect—or in interfaces where you weren’t involved directly?
Nishant Agrawal: This ties back to our theme—what happens when you’re not a technologist? How do you work with these tools?
Let me give you a recent example. I’ve been helping a customer in the media industry. At Alibaba, we’re actively contributing to the open-source AI community, and our model, Qwen, consistently ranks near the top of leaderboards—sometimes slightly ahead or behind DeepSeek. We’re also preparing to release a new quantized version soon, hoping to reclaim the top spot.
In this case, we were using Qwen to process video content. Media companies invest heavily in video—actors, camera crews, production—so there’s real value in repurposing that content. Maybe you want to trim a 45-minute video into a 30-minute cut, or create short clips that are optimized to go viral on social media. AI can help with all that.
I was building a demo, and when I shared it with the customer’s video team, their feedback was, “It works… but not quite.” What they meant was, yes, the AI sped things up—but it didn’t capture the human side of storytelling. Being a director or a creator is about understanding emotion, pacing, narrative—things that, at least for now, AI can’t fully grasp.
So yes, the tech can make the process more efficient, but it’s not about replacing the human. If you try to remove the human judgment and creative touch entirely, the final product might lose its meaning or impact.
Xiao Zhang: In our case, there aren’t many completely unexpected use cases—probably because we’re a tech-driven team, and we often start with the tool first, then look for the use case, like holding a hammer and looking for nails.
That said, customer feedback has been critical in pushing us to innovate. When we first launched our AI design tool in 2023, the results were often disappointing—the generated photos didn’t preserve room structure and didn’t look realistic. No matter how much we marketed it, subscriptions stayed low.
After listening to users, two key things became clear: the output needed to preserve the room layout and look photorealistic. That feedback pushed us to rethink our approach. Originally, we combined open-source models with rule-based logic—like always placing two nightstands by a bed or hanging art next to windows. But this couldn’t handle the wide range of room types and personal styles.
Around that time, we spoke with someone from Tesla’s Full Self-Driving (FSD) team. They had a similar challenge—rule-based systems couldn’t account for every driving scenario. Tesla spent years building those systems but eventually switched to end-to-end models, letting AI learn from vast data instead of relying on predefined rules. That shift dramatically improved performance.
Inspired by that, we also moved to an end-to-end approach. We trained our models on diverse room images and floor plans, letting the AI learn how to generate layouts flexibly. Once we got to the point where the outputs were reliable—room structures were preserved, and visuals looked realistic—our user adoption and subscription rates went up significantly.
In short, even though we start with a technical mindset, it’s the customer feedback that truly shapes and improves the product.
Zhaoyang Wang: How about you, Charles? Have any of your customers used your product in ways that surprised you?
Charles Xie: Yes, absolutely. One surprising example was when we found out that companies in the pharmaceutical industry were using our technology for new drug discovery. They’re analyzing the 3D structure of molecules and proteins, using our system for virtual drug screening. It turns out they can significantly speed up that process with our tools—which we never originally anticipated.
At the end of the day, we’re a database company. As a data infrastructure provider, our goal is to build a product that serves as many use cases and industries as possible—not something limited to a single vertical.
But that’s also a challenge. As a startup, you don’t have the resources to deeply engage with every industry. Our solution was to go open source from day one. We made every line of our code public, and that decision paid off. Now, we get feedback daily from users across all kinds of industries. They might find a missing feature or a new way to use the system. That collective input helps us identify patterns and build infrastructure that’s adaptable to a broad range of needs.
Zhaoyang Wang: Let’s move into the second part with some individual questions. Starting with Charles—you’ve mentioned several use cases already. Quick question to the audience: how many of you know what a vector database is?
Wow, more than expected. If we had this panel last year, probably far fewer hands would’ve gone up. So Charles, how have you seen this space grow?
Charles Xie: It’s grown a lot. When I started eight years ago, very few people knew what a vector database was. I remember pitching to investors—they knew I was one of the founding engineers of the Oracle Cloud Database, and many told me, “Charles, if you’re building a relational database, I’ll write you a big check.” But when I said I wanted to build a database for AI, they thought I was crazy.
Even around 2019 or 2020, probably less than 1% of developers knew what vector databases were. Now, I’d say that number is closer to 30%, and it’s still growing fast.
From a business perspective, we’ve tripled our revenue two years in a row, and expect to do it again this year. What’s exciting is how widespread the adoption has become, across three major segments:
1. GenAI startups – They use vector databases as a core part of building generative AI applications.
2. High-tech companies – Internet, e-commerce, ride-sharing, food delivery—many are integrating vector databases to power smarter experiences.
3. Traditional industries – Legal firms are using our tech for contract management, summarization, and document drafting. In education, it’s being used for tutoring and language learning. Even retailers and commercial banks are analyzing user behavior to grow revenue.
And this is just the beginning. I believe in the next 10 to 20 years, every company will be powered by AI—just like how computers and keyboards became universal after the IT revolution. We’re at that kind of inflection point again.
Zhaoyang Wang: The story of Zilliz is actually quite typical in Silicon Valley. New concepts emerge almost daily. If you build your company around one that hasn’t been widely adopted yet, it’s harder to raise funding—but you also gain more time to refine your product before the competition heats up.
Now, Xiao, you mentioned that you come from a physics background, and now you’re building a startup in real estate. That’s another classic Silicon Valley story—tech founders disrupting a sector few thought was ready for innovation. How did you go from studying physics to launching a startup in real estate?
Xiao Zhang: During my PhD in Applied Physics, I worked on using AI to solve physics problems. Specifically, I developed models with strong spatial reasoning and path-planning capabilities. For example, helping control and guide particle beams through 3D space.
After graduation, I realized that physics felt too far removed from real-world impact. I wanted to apply what I had learned to solve practical problems. That’s how I discovered the inefficiencies in real estate and home renovation. The market is huge, but the workflows are still slow and expensive.
Realtors spend thousands hiring staging companies to rent furniture and take photos. Furniture companies invest weeks in interior design proposals. There’s clearly a lot of room to improve efficiency. Given my experience in spatial AI, it made sense to apply those models here—to improve planning, layout, and visualization in real estate. After that, we also began exploring applications in generative e-commerce.
Zhaoyang Wang: That makes sense. You mentioned earlier that feedback plays a big role. How did you actually learn about the real estate industry and its needs?
Xiao Zhang: Transitioning from physics and AI into real estate meant I had a lot to learn. The most important thing has been listening to customers. Every day, my team and I talk to 5–10 users to get feedback and improve the product. We also brought in advisors from the industry. One of them is a former regional president at Engel & Völkers, one of the world’s largest real estate firms. His insights on go-to-market strategy and product fit have been incredibly helpful. In short, listening to your users and learning from industry insiders has been key to building something that actually works.
Zhaoyang Wang: Nishant, Charles mentioned Zilliz’s open-source roots, and you brought up Qwen. Qwen is actually one of the hottest open-source models right now—even Dr. Fei-Fei Li from Stanford is a big fan. How does Qwen connect to Alibaba Cloud’s broader product ecosystem? You have so many different offerings for business customers—what role does Qwen play?
Nishant Agrawal: I joined Alibaba in 2016, and I remember our founder, Jack Ma, talked about ideas like this even back then. But at the time, no one was really discussing generative AI yet.
Our research team started in 2017, initially focused on industrial solutions—things like improving agriculture, helping people mostly in China. Then COVID hit, and suddenly there was a lot of focus .
Over the last five years, one of our core strategies has been building an ecosystem. But the cloud market in the U.S. is very different from the cloud market in China. As Charles mentioned, the adoption curve looks completely different. One of the most effective ways to drive adoption—especially in China—is through open source.
That’s where Qwen comes in. The team has done an incredible job, with strong support from company leadership. They’ve committed billions of dollars to training these models—and crucially, to open-sourcing them. Because in a market like China, building a thriving AI ecosystem requires collaboration, scale, and access to compute. Open-sourcing Qwen helps make that possible.
Also, unlike in the U.S., there’s no Hugging Face equivalent in China. So Alibaba created ModelScope, a platform to host and share open-source models. It’s part of our broader effort to build community around AI development.
And it taps into a cultural strength—Chinese engineers are already major contributors to global open-source projects. Just look at conferences like KubeCon—many of the speakers and contributors are from China. So open source isn’t just a strategy; it’s deeply tied to how innovation happens in the region.
Zhaoyang Wang: Alibaba Cloud is interesting because on one hand, you have strong technical products like Qwen and other advanced tools. You have significant influence in the tech stack. But on the other hand, as you mentioned earlier, you’re often the bridge between the technology and the customer. Many of your users may not be familiar with all the new trends and buzzwords. So how do you help them? Do they need to understand every new concept—or should they just focus on solving their own problems?
Nishant Agrawal: That’s a great question. I’ll share something that recently stuck with me—our founder Jack Ma spoke at our internal company kickoff a few weeks ago. One of the things he said was: AI isn’t here to replace humans or humanity. It’s a tool. And it’s our job, as engineers, to help AI understand humans—and help it do the things we can’t. Take a simple example: if you’re a farmer working 12-hour days, and AI tools can help you get the same work done in 8 hours—that’s powerful.
I also want to give a shoutout to Zilliz. Their blog posts are fantastic. They do a great job breaking down complex technical topics in a way that’s really accessible. I often recommend their content to people trying to understand the basics of vector search and AI infrastructure.
So no, I don’t think customers need to understand every buzzword. Instead, we focus on what they need to do. They’re the experts in their domain. My job is to help translate their needs into the right tools—not to teach them how every model works. And hopefully, as the tools evolve, they’ll become so intuitive that translation won’t even be necessary. AI will just work the way people need it to—naturally.
Zhaoyang Wang: Let’s wrap up by looking toward the future. Charles, you mentioned earlier that Zilliz’s story is built around a new concept. You also brought up a trend called Agentic RAG—and how vector databases play a key role in it. Could you explain what that means and what the future might look like?
Charles Xie: Sure. In my view, agentic AI is about building systems that are more capable and more autonomous. Unlike traditional one-shot interactions—where you ask a question and get a single retrieval—agentic systems break down tasks into multiple steps and iterate to get better results.
Take RAG (Retrieval-Augmented Generation), which became popular in the past few years. Traditionally, RAG performs a single retrieval from a knowledge base and passes that to the LLM to generate an answer. But Agentic RAG introduces recursion: the language model can decompose a complex task into subtasks, retrieve knowledge multiple times for each, and make decisions across several steps.
This shift is powerful. It means we’re moving from “ask once, answer once” to a dynamic, multi-step reasoning process. Agents can now plan, retrieve, evaluate, and refine—similar to how a human researcher might work.
From a vector database perspective, this is exciting. Each subtask triggers multiple retrievals, and that increases both query volume and the size of the knowledge base. So demand for vector search infrastructure grows significantly.
Of course, we’re still early. Today’s agentic systems can handle moderately complex tasks made of multiple steps, but we’re far from building agents that can pursue long-term goals. For example, creating an AI agent that could guide a child’s learning journey all the way to college—that’s still a major challenge.
The future will depend on how well we can build systems not just for tasks, but for goals—and that’s where the next breakthroughs will happen.
Zhaoyang Wang: Xiao, in your opinion, how will Collov change the real estate industry?
Xiao Zhang: Whether it’s Collov or another company, I believe the direction is clear—AI will dramatically improve both efficiency and cost in real estate.
Take real estate agents, for example. Right now, they hire staging companies that rent furniture, set it up in homes, and then bring in photographers. It’s time-consuming and costs thousands of dollars per listing. With AI-powered virtual staging, we can now generate photorealistic images for just 10 to 20 cents per image—compared to thousands for traditional staging. What used to take weeks can now happen in 15 seconds. The same applies to furniture vendors. Today, they hire interior designers who spend weeks creating design proposals. With AI, this process can be shortened dramatically and done at a fraction of the cost.
Another trend we’re seeing is the rise of individual professionals empowered by AI. Previously, designers might have needed a full studio team of 5 to 10 people. Real estate agents had to work with multiple vendors. Now, one person can handle staging, marketing, and even building a presentation or website—with AI tools supporting each step.
So yes, I think we’re moving toward a more efficient, cost-effective, and empowered industry—where even solo professionals can do the work that once required entire teams.
Zhaoyang Wang: Nishant, I know Alibaba is planning to do more marketing in North America. Is there anything you can share with us about what you’re currently working on?
Nishant Agrawal: Sure. In the U.S., we already operate a few cloud regions and support American companies—especially those doing business in China. Due to export control restrictions, tools like OpenAI aren’t available in China, so we often help bridge that gap.
One of our top priorities right now is continuing to build a world-class large language model. Qwen is at the center of that effort, and our engineers—along with partners in the open-source community, like those from companies such as Yama and Index—are working hard to improve it.
We’re also encouraging more experimentation. Technology adoption isn’t just about capability or infrastructure—it’s also about culture, talent, and timing. That’s why it’s important to have a diverse ecosystem of tools.
Today, both the U.S. and China are driving innovation from different angles. Take Kling, for example—it’s not from Alibaba, but it’s an impressive video generation model coming out of China. Having access to tools like that gives everyone—developers and users alike—more options to find what works best for their needs.
Zhaoyang Wang: Thank you. That’s all from me.