Original · Unique Research · 2026-06-07
Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the complete opening analysis, all five interview sections, and the full 10-question Q&A excerpt. All named speakers, companies, roles, numbers and claims are preserved. Company, personal and product names are transliterated where official English forms remain unverified. Market projections, performance claims and company-specific figures are source or speaker attributions, not independently verified findings.
AI Industry Observation
How Chaotic Your Company Really Is, Just Connect an AI and You'll Know
What Enterprises Truly Need to Solve Is Not Whether AI Is Smart Enough, but Whether AI Can Enter the Business Accountability Chain
"
The value of an industry agent platform is putting model capability into real enterprise scenarios, making it not just 'smart,' but 'usable, controllable, deliverable.'
Let me put a number here first: 40% to 95% of enterprise AI projects fail to deliver measurable business value.
The range is large, and different research institutions define it differently. But no matter which number you take, the conclusion points to the same thing. Most enterprises spent money, implemented systems, made presentation PPTs, and then AI lay there quietly, nobody using it.
It's 2026, and models are already strong enough. The problem isn't the model.
"What enterprises truly need to solve is not whether AI is smart enough, but whether AI can enter the business accountability chain." This is what Yang Ting, CEO of SOIN AI, told me. Indeed, the vast majority of enterprise AI project failures are not due to technical reasons, but organizational reasons.
Yang Ting was formerly VP of commercialization at Kunlun Wanwei's TianGong AI, and now leads SOIN AI in enterprise-level Agents and vertical industry model application implementation. Last year she was invited to attend AI Founders Night co-hosted by OpenAI and EPIC, and is also an inductee of the ISC.AI Hall of Fame. But compared to these labels, I'm more curious about her answer to one question: when everyone is talking about how smart AI is, what do you think enterprises are talking about?
Her answer was four words: business results.
The Ceiling of General AI Is at the Business Accountability Chain
General large models have become the basic intelligence entry point for many enterprises and individuals, but enterprise core processes require far more from AI than generation and analysis capabilities. What's truly difficult is whether AI can enter the business accountability chain.
Take manufacturing as an example. When an engineering requirements document reaches an engineer, they need to check whether terminology conforms to industry standards, whether content conflicts with existing requirements, whether quality meets compliance requirements, who should handle approval, and how to trace issues if results go wrong.
General large models can assist with part of these things, but it's difficult to independently, stably and controllably complete the entire business closed loop. The problem is not that models completely don't understand engineering, but that enterprise scenarios rely on terminology systems, permission boundaries, process rules, historical data, system interfaces, result verification and accountability tracing.
"The value of an industry agent platform is putting model capability into real enterprise scenarios, making it not just 'smart,' but 'usable, controllable, deliverable.'"
Put more bluntly: what SOIN AI does is not make AI smarter, but make AI more reliable.
Yang Ting gave four hard standards for vertical industry agents—
Can understand industry knowledge. Not just knowing a few professional terms, but truly understanding business rules and judgment standards.
Can enter business processes. Not standing outside the process giving advice, but embedded in it, executing tasks.
Can connect real systems. Not self-amusing in a dialogue window, but interfacing data with the enterprise's ERP, PLM and document management systems.
Can be accountable for results. Outputs are verifiable, traceable and explainable.
Together, these four are the real dividing line between general AI and enterprise-level agents.
A set of industry data corroborates this judgment: 98% of manufacturers are exploring AI, but only about 20% believe they are ready for scaled deployment. The gap is not in model parameters, but in the deployment environment. In those places models can't touch.
"Checkup"—What AI Reveals Is the Enterprise's Own Organizational Capability
Midway through the interview, Yang Ting said something that left a deep impression on me:
"AI implementation is often also an organizational capability checkup. For enterprises with clear processes, agents will amplify efficiency; for enterprises with chaotic processes, agents will first surface the problems."
This is precise to the point of being piercing.
For many years, gray areas in enterprise processes were compensated for manually. Where boundaries were unclear, someone made a phone call and it was solved; where data口径 was inconsistent, someone manually reconciled it and it passed; where judgment standards were模糊, an experienced master made a call and it was decided.
But AI doesn't make phone calls, doesn't manually reconcile accounts, doesn't make decisions by intuition. For AI to run stably, there must be clear rules and boundaries. So when enterprises try to connect AI to core processes, those problems previously covered up by humans—will be ruthlessly revealed.
Many enterprises didn't anticipate this step. They thought introducing AI was upgrading a tool, only to discover they were performing organizational surgery.
Yang Ting divided enterprise assets into two categories. Data assets are what enterprises already have—documents, spreadsheets, system records, historical materials. Intelligent assets are capabilities extracted from data that can enter business and continuously produce results.
"Data assets are more like raw materials; intelligent assets are more like capabilities that can continuously create value."
An enterprise may have terabytes of documents. But if these documents can't be understood, called upon, or used to support business judgments by AI, then they're just static materials taking up hard drive space. Yang Ting gave a specific example: an enterprise has大量 R&D documents—this is data. Extracting requirement decomposition rules, quality inspection standards, historical issues and expert experience from the documents—this is knowledge. Letting AI automatically check requirement quality, discover consistency issues, and assist engineers in making judgments based on this knowledge—this is what's called intelligent assets.
But from data to intelligent assets, what's the biggest obstacle lying in between?
"The hardest part is making tacit business experience explicit," Yang Ting said. "The truly valuable experience of an enterprise is often not in standard documents, but in experts' judgments, team habits, lessons from historical projects, and long-formed business intuition."
Indeed, an enterprise's most precious knowledge is usually locked in the minds of a few key people. When these people leave, the knowledge is断. When they retire, the experience is gone.
And the process of building agents is essentially forcing enterprises to say these things out, write them down, and structure them. Not just for AI to use—but for themselves to use.
Yang Ting's original words: "For an enterprise to build agents, it's not just technical integration, but also a knowledge engineering and business梳理 exercise."
This is important—she didn't say "using our product will solve it," but was saying "you first need to figure out how you actually do things."
From Demo to Production—"Stunning Demos" Can't Save You
"Enterprises don't pay for a stunning demo—they pay for long-term stable business results."
This sentence points directly at the most common trap in the industry. The demo is cool, the boss nods and approves, then the project slowly bleeds to death on the road to launch.
Three typical death modes circulate in the industry:
"Silent delisting"—after launch nobody uses it, and it's quietly removed during budget review.
"Perpetual PoC"—the pilot validated value, but lacks budget, engineering capability and持续 leadership support, forever stuck in the pilot stage.
"Negative ROI deployment"—the model went live, but the human cost of handling exceptions and fixing hallucinations反而 exceeds the benefits.
Yang Ting believes that from Demo to production, the hardest part is controllability and sustainable delivery.
"A Demo only needs to prove effectiveness once, but a production environment requires it to work every day, work for different teams, work with different data, and the results must be traceable, explainable and correctable."
This is also why she defines SOIN AI's self-developed AWE (Agentic Workflow Engine) as a "continuously growing engine" rather than a static framework. AWE combines task decomposition, knowledge retrieval, tool invocation, process orchestration, memory precipitation and result verification into one system. It absorbs OpenClaw's ideas on tool use and real-environment execution, and also borrows Hermes's design on experience learning and skill precipitation. But Yang Ting emphasized that after these cutting-edge capabilities enter an enterprise, they must be叠加 with permission control, audit tracking and exception handling—"otherwise Agents can run, but it's hard to truly go live."
When talking about the most common mistake enterprises make when introducing Agents, Yang Ting's answer was crisp: setting the goal too large from the start, without clearly defining what counts as success.
"Many enterprises hope AI will cover many departments and many processes from the start, but without clear business owners, data foundations and value metrics, the project easily becomes a concept demonstration."
Her advice is to do it in reverse. Find a scenario that is specific enough, important enough, and easy enough to verify. First run results in a small closed loop, then expand.
"Enterprise agents are not implemented through one grand plan—they grow out of real scenarios one by one."
There's another new trend worth noting in 2026: more and more enterprises are starting to pay attention to the ROI cycle of AI Agents. The typical AI ROI cycle is 2 to 4 years, but many enterprises' financial assessment periods are only 7 to 12 months. Under this expectation mismatch, many projects are cut before they mature. This is also why Yang Ting recommends starting from small scenarios—not because ambition is small, but because only by surviving do you have a chance to grow.
Two Seemingly Different Battlefields
What SOIN AI currently focuses on is enterprise scenarios with high knowledge density, high document density and high process complexity. Advanced manufacturing R&D processes are one typical direction, while the enterprise-level marketing brand content production tool platform is another scenario that can reflect the value of Agentic Workflow.
The former faces supply chain Datasheets, requirements, documents, standards, verification and quality inspection; the latter faces creativity, scripts, materials, storyboards, generation and review. They look different on the surface, but at the core neither is single-point generation—both are complex workflows.
Manufacturing R&D, from requirements to documents to standards to design to verification to quality inspection, every step relies on knowledge understanding and process collaboration. AI video generation, from creativity to scripts to materials to storyboards to generation to review to distribution, is also not a single-point action of clicking "generate."
On the manufacturing R&D side, what Yang Ting is more optimistic about as an entry point is not hot topics like production scheduling or equipment maintenance, but the R&D and engineering collaboration环节—requirement quality inspection, similar requirement lookup, R&D document review, knowledge reuse, consistency analysis.
Why? Because these scenarios have high knowledge density, occur repeatedly, rely heavily on expert experience, and their effects are easy to measure. SOIN AI's ALA product targets ALM / PLM scenarios, focusing on requirement management, engineering document processing, R&D quality inspection and knowledge reuse in advanced manufacturing R&D processes.
In the AI video direction, Yang Ting's judgment is equally sober: "Enterprises don't make content to generate a video—they do it to serve brand, product and conversion goals."
SOIN AI's "Zhihui Changjuan" product supports professional-grade generation of 5 to 15 minute long videos, breaking through the limitation of traditional AI tools that can only make a few seconds of clips. But what Yang Ting cares more about is not the duration breakthrough itself, but organizing the entire content production chain—creativity, scripts, materials, storyboards, generation, review—into a manageable, reusable and optimizable workflow.
The commonality of the two directions is here: what's truly valuable is not some beautiful output, but the set of working methods behind it that can continuously operate.
When talking about cross-industry replication, Yang Ting made it clear: the underlying engine and operating mechanism can be standardized, but industry knowledge and business rules must be deeply customized. "If you don't understand industry differences and only copy the technical framework, it's hard to truly land."
Globalization: Opportunity Is in Industry Scenarios, Not Foundation Models
Yang Ting appeared at multiple international conferences in Singapore and Dubai last year. On the globalization opportunity for Chinese AI Agent companies, her judgment is:
"The biggest opportunity is not in foundation models, but in industry scenarios and application implementation."
The logic is simple. Chinese and Asian enterprises have accumulated a large number of real and complex scenarios in manufacturing, supply chains, content production and cross-border business. These scenario experiences themselves are competitiveness. The global market is now all answering the same question: how does AI go from a personal efficiency tool to an enterprise business capability?
Whoever can solve specific problems in specific industries, adapt to local environments, and continuously deliver results—whoever has the opportunity.
One Last Question
Before the interview ended, I asked Yang Ting: what do enterprise-level agents truly change?
Her answer:
"What enterprise-level agents truly change is not that enterprises have one more AI tool, but that enterprises have the opportunity to沉淀 business capabilities that previously relied on personal experience, departmental collaboration and human judgment into organizational capabilities that can be continuously executed, continuously replicated and continuously optimized."
I'd like to add one sentence myself. If this can be done, Agentic AI is not just a technology trend—it's a fundamental restructuring of how enterprises operate.
The premise is that enterprises must first be honest with themselves.
Must first pass that checkup.
Interview Highlights Q&A
The following 10 Q&A pairs are excerpted from the original interview, preserving Yang Ting's complete expressions.
Q1: Without using the official introduction, how would you explain in one sentence what SOIN AI does?
SOIN AI turns business capabilities that enterprises have沉淀 in documents, systems, processes and expert experience into industry agent applications that AI can understand, execute, verify and continuously optimize. What we focus on is not whether AI can generate a piece of content, but whether AI can enter the enterprise's real business chain and continuously advance professional work.
Q2: Enterprises are already using ChatGPT, Claude, Gemini—why do they still need an industry agent platform?
General large models solve basic intelligence capabilities, but what enterprises truly need to solve is the business accountability chain. A model can understand language, generate content and give suggestions, but enterprise scenarios also face terminology consistency, permission boundaries, process rules, historical data, system interfaces, result verification and accountability tracing. If these problems aren't solved, AI's capability can hardly enter core business.
Q3: How do you view the difference between "data assets" and "intelligent assets"?
Data assets are more like raw materials; intelligent assets are more like capabilities that can continuously create value. An enterprise may have a large number of documents, spreadsheets, system records and historical materials, but if this content can't be understood, can't be called upon, can't support business actions, they're still just static materials, not intelligent assets.
Q4: When enterprises build agents, is the hardest part technology or making tacit experience explicit?
Often, the hardest part is making tacit business experience explicit. The truly valuable experience of an enterprise is often not in standard documents, but in experts' judgments, team habits, lessons from historical projects, and long-formed business intuition. These experiences are very important, but they're usually not clearly expressed or structured.
Q5: If an enterprise's processes themselves are unclear, can AI Agents truly land?
AI Agents can help enterprises optimize processes, but they can't replace enterprises in completing basic business definition. If a process itself has unclear boundaries, unclear accountability, unclear data口径, unclear judgment standards—after AI enters, the problems won't disappear, but will be exposed faster. AI implementation is often also an organizational capability checkup. For enterprises with clear processes, agents will amplify efficiency; for enterprises with chaotic processes, agents will first surface the problems.
Q6: From Demo to production environment, what is the hardest gap to cross?
From Demo to production environment, the hardest gap to cross is controllability and sustainable delivery. A Demo only needs to prove effectiveness once, but a production environment requires it to work every day, work for different teams, work with different data, and the results must be traceable, explainable and correctable. Enterprises don't pay for a stunning demo—they pay for long-term stable business results.
Q7: After enterprise-level Agents go live, how should value be measured?
Metrics differ by scenario, but must return to real business results. Enterprises can't just look at how much content AI generated or how many questions it answered, but whether it made processes faster, quality more stable, risk lower, collaboration smoother, or revenue conversion clearer.
Q8: What is the most common mistake enterprises make when introducing Agents?
The most common mistake is setting the goal too large from the start, without clearly defining the business closed loop. A better approach is to start from a scenario that is specific enough, important enough, and verifiable enough. First let Agents produce results in one business closed loop, then expand to more scenarios. Enterprise agents are not implemented through one grand plan—they grow out of real scenarios one by one.
Q9: Manufacturing and AI video generation seem completely different—why are both suitable for Agentic AI?
Manufacturing and AI video generation look very different, but at the core both are complex workflows. Every step of manufacturing R&D relies on knowledge understanding and process collaboration; AI video generation is also not a single-point generation action, but a content production chain. The advantage of Agentic AI is organizing multiple tasks, tools and knowledge around a goal to continuously advance.
Q10: One-sentence summary—what do enterprise-level agents truly change?
What enterprise-level agents truly change is not that enterprises have one more AI tool, but that enterprises have the opportunity to沉淀 business capabilities that previously relied on personal experience, departmental collaboration and human judgment into organizational capabilities that can be continuously executed, continuously replicated and continuously optimized.