Original · Unique Research · 2026-07-11
Editor's note: This is the original Chinese author's interview with Han Qing and its framing. This English rendition retains the full text in source order, including all 16 Q&A items. The interviewee's product claims, customer figures, and market views are his self-reports, attributed to him and not independently verified. Person, company, and product names are preserved as source attributions.
AI Industry Observation
Han Qing (韩卿), Kyligence co-founder & CEO: without a semantic layer, AI can't even find the door.
This is what Han Qing heard verbatim from a banking business executive. The IT department spent heavily on AI systems, yet the business side still files a manual ticket to pull a simple data point. AI hasn't even found the door. The line stings, but Han Qing says he's seen too much of it on the front line.
Where's the problem? Not under-spending, not models too small. Plainly, without a semantic layer, security, quality, permissions and other "enterprise-grade" requirements are all castles in the air. AI doesn't even understand what you're saying, so what's the point of talking permission tiering with it?
Han Qing drew me an analogy: many enterprises building AI systems today are led astray by pure tech teams. They start with data security, permission control, quality governance — things that "sound correct" and "no one dares take responsibility for." Dare you say security doesn't matter? Dare you say permissions don't need managing? But the things business actually cares about — can AI understand my report, does it know how my department measures performance — are usually skipped near project completion. The result is that banking executive's feeling: the AI IT built, the business side simply can't use.
Semantic Layer: For AI to Understand Business, It First Needs a "Translator"
What is a semantic layer? Plainly, AI needs an intermediate layer to get data and business context. It doesn't directly connect the database; it first understands whether "customer" in this table means signed customers or leads, whether this "revenue" is tax-inclusive or not. Kyligence has done this for the past ten years; the industry now calls it the semantic layer (Semantic View), or more grandly the Ontology.
"AI can read data; that doesn't mean it understands the business."
What is truly understanding the business? Knowing which metrics an enterprise cares about, how performance is measured, which factors correlate, which changes are worth acting on. Han Qing gave an insurance-client example: this insurer has a team-leader force of over 1,000, each managing nearly 20 agents, nearly 20,000 people in total. Before, team leaders wanting to see performance had to log into a separate reporting system and flip three or four tables to figure out their team's ranking and gap. Most couldn't read complex cross-tabs, nor had time to study. Now Kyligence's AI system directly pushes performance progress, horizontal comparisons, and improvement suggestions; team leaders open their phone in the morning and see "where you rank, how far behind, who to talk to." Management radius nearly doubled. The key isn't how smart AI is, but that AI first figured out how "performance" is calculated in insurance, what "good performance" means, and which direction "improvement" goes.
Here Han Qing added: "Using data used to be a privilege, mainly for managers, especially decision-makers." This points out a reality many don't realize — the data barrier is so high it keeps most frontline people out. The semantic layer isn't only Kyligence talking about it. Snowflake this year pushed Snowflake Intelligence, core is adding a layer of semantic understanding on the data warehouse; Databricks' Unity Catalog moves the same way; even Claude recently stresses "structured-data understanding." The whole industry moves the same direction — let AI first understand business language, then talk about what it can do.
Four Departments Calculated "Active Customers," and No Two Numbers Match
OK, the semantic layer's importance is clear. But when you actually land it, a starker problem appears. The same "active customer count," pulled by sales, operations, marketing, and finance, is likely four different numbers. Sales may count "with a signing action in the last 30 days"; operations "logged into the app"; marketing "left a lead"; finance may only recognize "already paid." Normally no one nitpicks; each uses their own, peacefully. But once AI joins decisions, the first thing to standardize isn't data but metrics. Data standardization solves "can it be used"; metric standardization solves "are we really discussing the same thing."
"Core operating metrics can't be redefined every time they're asked. How revenue is calculated, how profit is calculated, how active customers are defined, how growth is measured — these are consensus formed over long-term operation, the anchor of decisions."
Natural language solves "how to ask"; the metric system decides "on what basis to answer." Plainly, the metric system won't be replaced by AI. Its relation to AI is more like a foundation and the building above it — the stronger AI gets, the more solid a foundation it needs. Otherwise the fancier you ask, the more spectacularly it collapses. The future picture is roughly: stable operating consensus plus AI's dynamic analysis ability. Enterprises use the metric system to hold "what's right," and use AI to explore "why it changed, what to do." Data and context become enterprise-grade AI infrastructure — not optional, but mandatory.
That banking executive's words actually weren't finished. After "that AI our IT built is pretty useless," he added: "what I want is simple: I ask a number, and it gives me the right one." Sounds like a low requirement. But anyone who's been in this industry knows — this is exactly the hardest thing.
BI Won't Die, but That Wall Is About to Collapse
The semantic layer and metric system solved AI's "understanding." But a sharper question follows: if AI can really answer questions directly, does the BI industry still need to exist? Asked whether "AI will make BI disappear," Han Qing's answer has two halves. BI's professionalism and deep-analysis ability can't be replaced by AI short-term. Complex attribution, multi-dimensional cross-drill, regulatory-compliant fixed reports — things needing professional engineers, BI will keep doing. But BI's presence as a standalone tool will decline fast. In the past, to analyze you had to enter a dedicated BI system and view several or even hundreds of reports. Now, analysis capability is embedding into CRM, operating systems, office entries, and Agent workflows, happening naturally, even imperceptibly.
"Reports are still needed, but the entry will no longer be a standalone 'report system' or 'analysis system.' What truly disappears isn't BI, but the obvious boundary between analysis and business."
Soon You'll Either Command AI, or Be Commanded by It
In the past, using data was essentially a "privilege." Mainly managers, especially decision-makers, used it. Enterprises built massive reporting systems and dashboards for this, layer upon layer, ever more complex. But most people seeing these simply can't understand — you need expertise, business experience, even an analyst standing by to interpret "what this number means," "why it rose," "should we worry." Han Qing says AI data analysis should truly serve the 70%-plus non-technical staff in an enterprise. Regional managers, frontline operators, sales leads, store managers — these people rarely touched data before, not because they didn't want to, but because the barrier was too high. Now they can ask AI like a conversation, equivalent to giving everyone a professional data assistant.
The deeper change is at management. In the past, data reaching decision-makers passed through layer upon layer of aggregation. Regional manager reports to regional director, director to VP, VP to CEO. Each layer of human processing attenuates information, even deliberately whitewashes it. What the CEO finally sees and what really happened on the front line may already be two different things. AI breaks this chain. Decision-makers can directly ask AI, drilling through to any layer of detail. No need to wait for next Monday's meeting, no need to have a secretary fetch data — at 3 a.m. thinking of a question, open the chat window and get the answer. And the more detailed you ask, the less AI minds. What does this bring? People's roles polarize. You either become the decision-maker — giving the AI system business know-how, setting strategy and action direction; or become the decision-executor — landing based on AI's action plan. The middle "information transmitters" and "report processors" have less and less room.
The Enterprise's Brain Is Growing
If you think this is still too far off, Han Qing believes the biggest change in the next three years has already begun — Agent-ification. BI to AI analysis is a tool change; analysis to automatic decision is a capability change. But the deeper change is the redistribution of enterprise data capability. Basic analysis that used to rely on professional data teams — pulling data, making tables, reconciling definitions — gradually democratizes to the business front line. More complex analysis and judgment are done jointly by humans and AI. The data platform is moving from "storing history" to "the cognitive foundation of enterprise AI." There's a key distinction. Traditional data warehouses mainly answer one question: what data does the enterprise have? The future data platform must answer two more: what does this data mean? How should it be used? This means what the data platform settles isn't just data but metric consensus, business relationships, analysis logic, and permission boundaries. If AI can continuously understand and call these, the data platform is no longer just a warehouse storing what happened in the past. It settles how the enterprise understands business, forms judgment, makes decisions — plainly, this is the enterprise's brain.
"The AI-era data system no longer just answers questions; it reshapes the organization's whole management and operating model."
Selected Interview Q&A
Q1: Without an official intro, how would you explain to someone who doesn't know Kyligence what you do?
Han Qing: In one sentence, we provide enterprise-grade trusted data and context for AI. Today's AI is very powerful, but to truly understand enterprise business, there's still an accurate "translation" layer missing in between. AI needs an intermediate layer to get enterprise data and the business context behind it. This is actually the product and core technology we've built for the past 10 years. Now the industry calls it the semantic layer, Semantic View; more grandly, you can also call it the Ontology.
Q2: After AI enters the enterprise, what's the most essential change in the data system?
Han Qing: I think the most essential change is who uses it. Two changes here. First, from management using it, to grassroots also starting to use it. Second, from humans using it, to Agents also starting to use it. Using data used to be a privilege, mainly managers, especially decision-makers. So enterprises built lots of reporting systems and dashboards. But the problem is most people seeing these reports can't directly understand them, often needing expertise, experience, even an analyst to interpret. Today this "interpretation" work can already be done by an Agent, with broader knowledge, lower cost, higher efficiency. Humans only supplement part of information through Skills or a knowledge base. For example, an insurance client we serve: a team-leader team of 1,000-plus manages nearly 20,000 agents. Before they found reports a headache, because many couldn't understand them and had no time to study. But today, the Kyligence AI system can directly push performance progress, comparisons, and how to improve. This change is very clear; grassroots management radius nearly doubled.
Q3: Data systems used to solve "how to compute results faster"; now AI cares more "what does this result mean." Does this hold?
Han Qing: It holds. Past data systems first had to compute results fast and accurately, because they had to leave time for the "analyst" middle role to interpret, analyze, and summarize. But now these middle processes are all or mostly done directly by an Agent, which can even continue following up, reframe questions, handle ad-hoc problems. This used to be very costly to achieve. For example, sales down 10%; computing that number is only the starting point. What truly affects operations is whether you can further identify whether the problem comes from region, store, or product, and combine SOPs to form an actionable plan. AI's change is taking data analysis from "giving numbers" further to "understanding business," even "giving action plans." This greatly reduces middle layers and raises business agility.
Q4: How do you see the judgment that "data systems are moving from compute engines to decision infrastructure"?
Han Qing: I strongly agree. As more business people, managers, and Agents judge based on data, what the data system carries isn't just computation but how the enterprise defines problems, measures operations, and forms consensus. Through knowledge accumulation and AI's continuous learning, such a system directly becomes the enterprise's brain, even directly commanding other systems and human behavior. It's no longer just a data-providing layer but the enterprise's decision and operations infrastructure.
Q5: After AI enters enterprise analysis, is the key question "semantic consistency"?
Han Qing: Semantic consistency is very important, but more important is what this "semantics" actually means, what it means to different roles. That is business context, the most core. In most enterprises, the same "active customer count" is likely understood and defined differently by sales, operations, marketing, and finance. Especially when this KPI is used to calculate performance, all kinds of situations get very complex. So if the AI world can't solve this, AI results can't be trusted, and disagreements get amplified faster. It's also impossible to use one definition entirely, because each department's perspective naturally differs. So how to ensure semantic consistency and business flexibility isn't just a data-governance problem but whether an organization can form a common language, common cognition, and collaborate on the same set of facts. And this system needs to be translated to AI; that's the semantic layer's core value.
Q6: Kyligence always stresses metric systems and semantic-layer capability. In the AI era, is this layer's importance growing or being restructured?
Han Qing: It's being significantly enhanced. Recently Snowflake, Databricks, even Claude are all launching various semantic-layer capabilities, which shows the industry increasingly values this layer. And the semantic layer is foundational technology we started 10 years ago. Moreover, on top of the past semantic layer, the AI era needs more context. Before, one metric only needed dimension definitions plus some annotations. But today our system already supports various tags, expert semantics, knowledge bases, and more. The direction and way the semantic layer strengthens are getting clearer.
Q7: If the enterprise analysis entry becomes natural-language questions, does the metric system still matter?
Han Qing: I think it'll be extremely important instead. Natural language solves "how to ask," but the metric system decides "on what basis to answer." AI can dynamically generate analysis paths and help enterprises discover and reorganize metrics, but core operating metrics can't be redefined every time they're asked. Behind revenue, profit, customers, and growth is the operating consensus enterprises formed over the long term. So the future isn't the metric system being replaced by AI, but stable operating consensus combined with AI's dynamic analysis ability. Data and context become enterprise-grade AI infrastructure.
Q8: AI can read data directly, but does it truly "understand the business"? What's the key to understanding?
Han Qing: Reading data isn't understanding business. True business understanding is knowing what an enterprise cares about, how it measures, which factors correlate, which changes are worth acting on. So what AI must understand isn't one table or one data point, but how an enterprise recognizes, manages, measures, and runs its own business.
Q9: For AI to join decisions, what should be standardized first — data, metrics, or business processes?
Han Qing: If I must choose a starting point, I choose metrics. Data standardization solves "can it be used"; metric standardization solves "are we discussing the same thing." For AI to join judgment, it first needs to use the same business language as the organization. Humans already talk past each other; if AI applications don't solve this basic problem, they only bring more chaos. For example, two AI conclusions with inconsistent data results — how does the enterprise audit? How do you determine which is right, or whether both are wrong? If used directly without verification, the chaos is unpredictable. But metrics aren't the end. Ultimately it must enter flows, connecting problem discovery, cause analysis, judgment formation, and action push to form a business loop.
Q10: Who will be the users of AI analysis systems — business leads, management, or the system itself?
Han Qing: We believe AI data analysis should first serve the 70%-plus non-technical staff in an enterprise, the broad business people and functional-line colleagues. The real structural change isn't making existing analysts 20% faster, but letting people who rarely used data directly start using it. A regional manager, frontline operator, sales lead, store manager can all directly benefit from AI. It's like giving them a professional assistant, able to interact with an Agent, ask questions, follow up causes, verify judgments. Future it'll go one step further: the AI Agent itself also becomes a user of the data system. So the data system ultimately serves professionals, ordinary business people, and AI at the same time.
Q11: How do you see the view that "AI will make BI disappear"?
Han Qing: I don't think BI will disappear, but BI's presence as a standalone tool will decline fast. In the past, to analyze, users had to enter a dedicated BI system and view several or even hundreds of reports. Now analysis will happen more naturally in CRM, operating systems, office entries, and Agent workflows, even imperceptibly. Business people may not realize they're "using BI," but they're actually using data capability all the time. So what may disappear isn't BI, but the obvious boundary between analysis and business.
Q12: After AI enters the data system, what's the hardest problem to solve?
Han Qing: Definitely semantics. Without semantics, talking about security, quality, and permissions is empty talk. This is also a common ailment of enterprises building enterprise-grade AI systems today: often led astray by pure tech teams. Before things even start, a bunch of people raise various "enterprise-grade" requirements, especially security, permissions, quality, because these all sound correct and no one dares take responsibility. The result is enterprises spend lots of time solving these, while what business really cares about gets skipped. I recently met a banking business executive who said directly: "That AI our IT built is pretty useless." Because IT habits first build things valuable to them, but for business even a simple data pull still runs the manual-ticket model. Data quality, real-time, security have relatively mature engineering paths, but semantics involves how an enterprise defines customers, growth, risk, and success. Harder still, these definitions are scattered across departments, systems, and people's experience, needing large professional teams to build. So AI-system construction must start from the business and iterate fast. Better to abandon a system and rebuild than to put a bunch of shackles on the enterprise's real AI innovation and use from the start.
Q13: Why are many enterprises' data systems strong, but AI applications still don't run deep?
Han Qing: After touching many enterprises, one clear feeling: many enterprises have settled lots of data but haven't really settled "how to understand and use this data." Analysts know which table to query; business leads know which changes to watch; managers know when to act. But this knowledge largely lives in people's experience. As people change, this knowledge fades. So AI applications don't run deep often not for lack of data or models, but for lack of the ability to connect data, business semantics, and analysis logic. What enterprises most need now is distilling experience from people's heads into the AI system, while connecting information and data scattered across systems into the AI system.
Q14: If the AI system makes decisions directly on enterprise data, how to avoid the risk of misreading data semantics?
Han Qing: The key is making AI's judgment evidence-based, bounded, explainable, and traceable. We propose "four trustworthiness" to avoid this risk. First, trusted data — data results must be accurate. Second, trusted permissions — different roles see only data within their permissions. Third, trusted process — metric lineage, model input/output, and auditable records. Fourth, trusted delivery — results must be recomputable and human-verifiable. Enterprise AI can't only pursue "can answer"; more important is "trustworthy."
Q15: Over the next three years, what's the biggest change in enterprise data systems?
Han Qing: The biggest change is Agent-ification. From BI to AI analysis is a tool change; from analysis to automatic decision is a capability change. But the deeper change is that the enterprise's data capability begins to redistribute. Basic analysis that used to rely on professional teams gradually democratizes to the business front line. More complex analysis and judgment are done jointly by humans and AI. This changes not only the data system but the enterprise's decision efficiency and organizational operating mode. AI will definitely reshape the enterprise's organization and management model.
Q16: Will the data platform future become "the cognitive foundation of enterprise AI"?
Han Qing: I think so. The data warehouse mainly answers "what data does the enterprise have." The future data platform must also answer: "what does this data mean, how should it be used." What it needs to settle isn't just data but metric consensus, business relationships, analysis logic, and permission boundaries. If these capabilities can be continuously understood and called by AI, the data platform is no longer just storing what the enterprise did in the past, but settling how the enterprise understands business and forms judgment. This is what I understand as "the cognitive foundation of enterprise AI."