Domestic market. B2B. Customization. Each term sparks debate in today’s artificial intelligence startup landscape. But that’s exactly where Dai Zonghong, co-founder of 01.AI, has decided to stake his new venture.
On March 24, after leaving 01.AI, Dai launched AlignBase, a company focused on harnessing AI for B2B customization. Unlike traditional enterprise projects, his approach seeks to automate much of the manual groundwork.
Typically, enterprise customization requires teams of specialists to conduct interviews, gather and clean data, analyze it, and build digital models of workflows before companies can run simulations. AlignBase aims to hand this entire process to AI, from data collection to workflow mapping and beyond.
At the center of its platform is an operating system that is said to automatically map enterprise workflows. Base models and industry-specific models parse raw business data to identify production factors. These are then modeled one by one into workflow nodes through a production factor toolchain. A reinforcement learning toolchain uses the digital twin of the workflow to build customized enterprise models and AI software.
With this system, companies can reportedly finish data collection, cleaning, and workflow construction in a single day, with clients unlikely to find errors.
Industries from steelmaking to renewable energy are already testing the platform for cost reduction, efficiency gains, and supply chain optimization. For example, a company seeking to cut supply chain costs by 15% can connect its supply chain API to AlignBase’s system, which then generates a business model outlining the best path to reach that target.
AlignBase recently closed an angel funding round, raising more than RMB 100 million (USD 14 million) from a slate of investors including CapitalNuts, Empyrean Venture, CAS Star, Mellanox, Puhua Capital, Sinovation Ventures, Unity Ventures, and Yinqiao Fund.
Skepticism remains. “Some investors worry that Chinese enterprise clients are reluctant to pay,” Dai told 36Kr. The concern is not unfounded. Since the AI boom began, many firms have moved from customization to standardized products, pointing to slow payment cycles and limited client willingness to pay. Baidu founder Robin Li has even urged companies to avoid project-based B2B work in favor of standard offerings.
Yet Dai disagrees. Before AlignBase, he managed AI infrastructure at Alibaba’s Damo Academy and was later the CTO for AI at Huawei Cloud, where he oversaw more than 100 customization projects. From that, he drew a key lesson:
“It’s not that Chinese companies don’t want to pay, they just pay differently. Overseas, clients pay for tools. In China, they pay for results.”
That insight, Dai said, signals sustained demand for customization. He now spends about 40% of his time with clients, many of whom resist turnkey solutions that fail to integrate with existing workflows. Instead, they prefer end-to-end customization and, according to Dai, are willing to pay generously for it.
For Dai, the greater risk is not competition or capital but the erosion of client trust if customization fails to deliver. Poor outcomes damage confidence. “Clients don’t care about models; they care about results,” Dai said.
To address this, AlignBase doesn’t push companies to adopt its operating system upfront. Instead, it first delivers solutions built through the OS, showing value before asking for deeper adoption.
“If successful case studies increase, clients’ confidence will grow, and customer acquisition will be easier,” Dai said.
The following transcript has been edited and consolidated for brevity and clarity.
Turning data into workflows
36Kr: How would you describe AlignBase’s business?
Dai Zonghong (DZ): I want to provide an operating system that bridges the gap between industry and AI, so core business workflows can quickly adopt AI tools and methods for wide-ranging change, ultimately improving business value.
36Kr: That sounds a bit abstract. Can you give an example?
DZ: For one client, we fed raw data from its various systems into our platform. The models automatically learned from it, with no human intervention. With just a few machines running for about a day, we mapped the company’s entire workflow.
Traditionally, mapping operations may require months of interviews, technical reviews, and data governance. We can now complete it in a day. After detailed client review, no errors have been found.
36Kr: So how is this different from traditional customization?
DZ: Most companies rely on manual customization and modeling, which can bump into problems with data governance and workflow learning. Our technology can independently learn from complex enterprise data, reconstruct workflows, and create a new system for unsupervised data governance.
36Kr: What’s the advantage of this approach?
DZ: First, it generalizes. Second, it penetrates core business workflows, not just office tasks. Third, it scales, not just beyond one or two enterprises, but across entire industries.
36Kr: You co-founded 01.AI in 2023. Did you already see an opportunity for customization then?
DZ: From ChatGPT itself, no. Its strength lies in language and logic. Only with deep reasoning can models truly grasp workflows and support decision-making. That changed when OpenAI released o1. At that point, the question was no longer about viability, but effectiveness.
At 01.AI, we also explored frameworks like ReAct. The results were meaningful, though not as robust as o1. Still, it was progress, and achieved entirely on our own technology.
36Kr: Why did you leave Huawei to join 01.AI in 2023?
DZ: At Huawei, I worked on AI projects across industries. The experience was invaluable, but it was difficult to dive deeply into large model research. 01.AI had both the technical and financial foundation. After discussions with Lee Kai-Fu and Xue Mei, I decided to join.
At 01.AI, I systematically learned the full spectrum of large model technologies. Previously, I treated AI as a supporting tool. There, I had to execute across every metric, and that was a very different experience.
36Kr: What technical developments triggered your decision to start AlignBase?
DZ: I’m not a large model researcher by training. I worked on them because I believed they could help me achieve three things: generalization, workflow integration, and scalability.
The turning point was deep reasoning. When OpenAI released o1 and o3 in 2024, I realized I didn’t need to train my own foundation models. I could build on others’ work. That’s when I started AlignBase.
A 20-person team doing the work of 100
36Kr: What’s the structure of your operating system?
DZ: It’s a platform that integrates models and applications. At the base are general-purpose and industry-specific models that interpret business data. On top, company-specific data builds a hybrid digital twin. These models don’t hallucinate, because they are grounded in enterprise data.
36Kr: How do you “map an entire workflow” with AI?
DZ: Enterprise data is messy, spread across multiple sources and formats. We first mine it deeply, including hidden patterns, then extract production factors to build workflows. Each node is modeled, and gaps are automatically filled with missing data.
In short, we reconstruct real workflows. It sounds simple, but automation is hard. Some firms have millions of nodes, and learning all their logical and developmental relationships is a major challenge.
36Kr: How do you ensure the data aligns with real business logic?
DZ: It doesn’t need to be exact, only correct. We refine accuracy through reinforcement learning.
In AI’s 1.0 era, models were trained on labeled data, which required massive manual effort. We no longer rely on labeling. Instead, we design systems that create reinforcement learning environments, generating continuous real-world data to train our models.
36Kr: Why should enterprises care about modeling workflows with AI?
DZ: That enables them to simulate and make decisions. It’s like a mirror of their real operations. They can optimize costs, balance production, expand capacity, or refine supply chains.
36Kr: Does this still require significant human involvement?
DZ: Almost none. Once we have API access to the data, the process is automated.
36Kr: Does the system deliver the same results as manual customization?
DZ: We’re not replacing enterprise functions, but the AI experts who would normally model workflows. Traditionally, ten experts might build ten models. With thousands of factors, you’d need thousands of models. Our system replaces that with computers.
36Kr: B2B customization companies often end up bloated in headcount. How many team members does AlignBase have?
DZ: We’re running seven or eight projects with just 20 people. Traditionally, that would require more than 100. And we’ve never missed a delivery timeline. Technology solves most of the problems.
36Kr: What stage is your OS at now?
DZ: We delivered the first core features in early August, after three to four months of development.
36Kr: How is it performing in the field?
DZ: Clients already find it useful. With access to their historical data, we can quickly map workflows. They have used the digital twins for cost reduction, supply chain optimization, and energy management.
36Kr: How does a client use the platform to hit business goals?
DZ: We have yet to automate the full end-to-end process. For now, we manually connect the automatically built nodes. Clients set a goal, such as fastest production speed or maximum output, and our system searches for the optimal path.
Client trust comes first
36Kr: Are you focused on the domestic market or overseas?
DZ: Mainly domestic. China’s supply chain gives us a natural advantage. The longer the chain, the more room for optimization. Overseas companies may excel in specific areas, but China’s end-to-end industrial chains are stronger.
36Kr: Given you’re B2B- and China-focused, is fundraising harder?
DZ: We’ve met investors who share our philosophy. Those who have seen factory operations or worked on the ground understand our approach quickly. Often, one meeting is enough.
Before we proved the technology, some investors doubted workflow modeling could be automated. That was natural skepticism. But we were fortunate to find backers who understood our vision, and we closed our angel round of more than RMB 100 million quickly.
36Kr: Are you concerned that Chinese enterprises have low willingness to pay for B2B software?
DZ: I don’t think they are unwilling to pay, it’s just a matter of different preferences. Overseas, companies pay for tools. In China, they pay for outcomes. The clients we serve budget for value. If we can deliver tenfold returns, they would be eager to work with us.
Our current clients include both state-owned enterprises and private firms.
36Kr: How many clients do you have now?
DZ: Close to ten projects are underway across multiple industries. Clients are highly willing to pay, often with large upfront commitments.
36Kr: How should enterprises measure the impact of your OS? How do you price it?
DZ: Traditional customization software was essentially a dashboard. Its value depended on humans interpreting data, which made pricing difficult.
We deliver automated AI. Value is measurable and predictable. We also optimize full workflows rather than single points, so we can quantify the contribution of each improvement within the larger system.
36Kr: Can this model break even?
DZ: Without additional R&D investment, we have a chance of breaking even this year.
36Kr: Would you consider taking on projects with quick returns to sustain the company while pursuing your vision?
DZ: That approach is probably wrong. Commercialization should mean creating value for clients, then sharing in it. We aim to deliver tenfold returns, then capture our portion. Otherwise, it’s just short-term projects.
36Kr: But if those projects sustain the team, isn’t that also of value?
DZ: That harms the industry. If I promise RMB 30 million (USD 4.2 million) in value but deliver less than RMB 10 million (USD 1.4 million), I’m likely to lose a client who might never invest again.
Confidence is critical in B2B work. This year’s results shape budgets for years.
36Kr: Have you met clients lost in the past? How do you rebuild confidence?
DZ: Some, yes. Their tolerance for sunk costs is lower. Others may give us six months to prove value. Increasingly, clients want results in three months or less.
That means precise calculations and business value assessments that stand up to scrutiny.
36Kr: How do you explain such a new system to clients?
DZ: Many AI companies pitch “intelligent assistance” or promise to build large models. Clients don’t care about models, they care about results.
If Chinese AI companies can clearly show business value, winning clients becomes much easier. We’ve benefited from China’s industrial AI push. Our system shows value directly, so client acquisition hasn’t been difficult.
36Kr: Which industries or companies show the best results with your OS?
DZ: It’s universal, a general operating platform. We already have clients in metal smelting, steel, environmental management, renewable energy, and electronics manufacturing. The range is broad.
Competition builds confidence
36Kr: Do you think your business has a moat?
DZ: First, we understand industries deeply. Second, we understand model development just as deeply. Third, our technology and products are forward-looking. And finally, we execute well. It’s the combination of knowledge, design, and practical ability that sets us apart.
36Kr: Many companies claim to have industry know-how. What does that really mean?
DZ: At Huawei, I worked on more than 100 projects related to AI and industry. Even with traditional AI, we learned how companies operate, how executives think, what their data looks like, and how value chains function. That insight came from years of hands-on work.
So today, a few points matter:
- Understanding how industries run and how decision-makers think.
- Understanding AI’s current limits, and where it will be in six months, a year, or two. If you only design for today’s AI, you’ll soon have to rebuild.
- Understanding costs, compute needs, business setups, and data sensitivity. It’s always a complex system.
36Kr: How do you forecast AI’s capabilities one or two years ahead?
DZ: We don’t need precise forecasts, just direction.
First, I don’t think large models will break the data barrier soon. Most natural data has already been consumed. Progress is more like standing on your own shoulders. So general capabilities will improve linearly, not exponentially.
Second, in specific scenarios, models will improve quickly in reasoning, self-correction, and reinforcement learning. That’s where real breakthroughs will be.
36Kr: How do these judgments shape AlignBase’s work?
DZ: It means we can’t rely on AI improvements alone. We need to solve problems through business workflows. That’s why our models are not just industry-specific, but DAG (directed acyclic graph)-based compositions and logic graphs. That’s the foundation of our full-factor modeling.
36Kr: In your two years of entrepreneurship, do you regret any technical decisions?
DZ: At 01.AI, the team was excellent. In mid-2024, we explored deep reasoning. The models demonstrated strong reasoning ability, but each inference took more than ten minutes, which we judged too slow for complex tasks.
Looking back, we could have invested more in chain-of-thought and reinforcement learning.
36Kr: What lesson does that hold for your current startup?
DZ: Tie tech closely to real and future scenarios. That’s the most direct measure of value. And judge more from technical principles, not just a handful of metrics.
36Kr: Any recent technical progress you find critical?
DZ: Not yet. I’m waiting for models that learn from feedback in real scenarios without relying on short-term memory. That’s possible, and we’re working on it too.
36Kr: Are you worried about competition?
DZ: Not at all. We’ve built knowledge, understanding, and technology over nearly a decade. Other teams may match us in some areas, but as a whole, we’re rare.
The market is also huge. China has more than 500,000 major industrial companies. Most have the ability and willingness to pursue intelligent transformation.
We’re the first in China to fully automate workflow modeling. I hope more companies join us. That will help traditional enterprises upgrade faster.
Competition, overall, is positive. More successful case studies build client confidence and make acquisition easier. If the market were small, one or two failures could kill it. But here, it’s large enough for many players to thrive.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Zhou Xinyu for 36Kr.