Why is it that the smarter AI gets, the more we need BI?

In August 2026, IDC released the first Market Scape vendor evaluation report for China's Data Agent sector — *China Data Agent Vendor Evaluation, 2026*. This report covers the full link of data management, governance and analysis, examining the comprehensive capabilities of Data Agent rather than a single function. The evaluation objects cover more than ten leading enterprises including internet vendors, cloud service vendors, and big data vendors, covering players across various tracks such as Alibaba Cloud, Volcano Engine, Tencent Cloud, and Transwarp.

A notable phenomenon on this list is that Smartbi, as a BI vendor, is shortlisted, ranked as a "Major Player", and takes the first place among major players on the capability axis, close to the boundary of the Leaders quadrant — it is the only BI vendor included in this evaluation.

This information itself forms an industrial proposition. The Chinese BI market is not short of players — IDC data shows that the size of China's BI software market reached 1.32 billion US dollars in 2025, maintaining a year-on-year growth rate of over 25% for 6 consecutive years. In such a crowded track, what kind of capabilities does a BI vendor rely on to enter the comprehensive capability evaluation system of Data Agent and obtain professional recognition?

Moreover, this is far more than the attribution question of "why Smartbi". What is more noteworthy is that Smartbi's path precisely answers a practical problem that plagues a large number of enterprises at present: in the AI wave, where should enterprises' BI go, and where should AI be implemented?

We found clues in the recent strategic business upgrade signals released by Smartbi. The core is not "shifting from BI to AI", but re-clarifying the relationship between BI and AI from the perspective of enterprise construction: BI is the foundation, which precipitates credible data, unified indicators, analysis models, permission systems and business applications; AI is the engine that improves the efficiency of calling, combining and analyzing these assets; the integration of the two constitutes the implementation foundation for AI to enter complex enterprise scenarios.

This relationship is exactly a key to understanding the competition logic in the Data Agent era.

Threshold of the first DataAgent evaluation: From "visible" to "trustworthy"

To understand why most BI vendors cannot meet the threshold of this report, we need to first break down IDC's evaluation logic.

According to IDC's public research framework, this Market Scape evaluation does not focus on whether a certain AI dialogue function is easy to use, but the comprehensive capabilities of Data Agent — full-link coverage from data integration, data governance, data discovery, to indicator development, knowledge management and analysis decision-making. In other words, the report does not examine "whether there is an AI function", but "whether AI capabilities can truly run through the entire data process".

This is exactly the threshold. According to IDC's definition in *China Data Agent Market Map, 1Q26*, Data Agent is "using Agent to manage and govern data, realizing precise query, analysis and decision-making through conversational or low-code entrances, lowering the threshold for obtaining insights" — it is not the Agent tool itself, but embedding Agent capabilities in a wide range of scenarios such as data integration, data governance, indicator development and knowledge management. This definition determines that Data Agent is not adding a dialogue window on BI reports, but requires vendors to have a complete capability stack from underlying data governance to upper-level intelligent decision-making.

In the past, the value of BI was more perceived by users through reports, dashboards and self-service analysis, and the product competitiveness was also concentrated on these visible presentation layers; while the indicator system, business models, permission governance and analysis methods have long been hidden in the background. After entering the Data Agent stage, this value structure has undergone a fundamental reversal: what was hidden in the background in the past begins to directly determine whether AI answers are credible. Every question users ask AI and every conclusion they get is based on whether the indicator definition is unique, the caliber is unified, the permissions are isolated, and the analysis method has been verified by business. When the evaluation standard is upgraded from "whether the report looks good" to "whether the AI answer is credible", vendors lacking underlying governance capabilities will naturally find it difficult to enter the system.

Gartner's judgment echoes this. In the 2026 *Magic Quadrant for Analytics and Business Intelligence Platforms*, the market trend is summarized as "Agentic AI, governed semantics and AI-augmented decision support", and "Agentic Insights" is listed as a necessary capability of the platform. Its subtext is very straightforward: having AI dialogue functions is only the entry point, and that AI conclusions are traceable, can remember enterprise business calibers, and analysis experience can be precipitated as organizational assets is the real "admission ticket".

And this admission ticket cannot be "developed" in the laboratory, but "accumulated" in large-scale complex scenarios. A BI vendor can spend three months catching up with the latest AI dialogue technology, but cannot spend three months making up for the pitfalls and gaps encountered in ten years of large-scale enterprise delivery. This also explains why IDC lists "revenue scale" and "product technology capability" as core dimensions: behind the revenue scale is the evidence that the vendor has verified its products in enough complex scenarios. Without the tempering of these scenarios, technical capabilities are only demo capabilities, not production-level capabilities.

In addition to the accumulation of engineering delivery, there is a more hidden threshold — industry know-how. For Data Agent to be truly "usable", general data governance is not enough: when AI faces a specific business problem, it must understand the reasonable caliber of indicators in this industry, where the regulatory red line is, and how to split business scenarios. This cannot be "learned" by the model through general corpora, but "achieved" in real industry projects one by one.

Smartbi has long been deeply engaged in industries with complex data environments and strict caliber requirements such as finance, government, manufacturing, energy, and retail, and has precipitated sets of industry-oriented indicator templates, analysis models and business rules such as risk exposure, capacity utilization, people-goods-scenarios, and government indicators into "industry assets" that can be directly called by DataAgent. When industry know-how enters the product, it means that AI does not only give "a result", but "an answer that conforms to the industry context and can withstand the test of industry calibers".

IDC's evaluation of Smartbi as "taking strong BI capabilities as the base" is exactly the third-party proof of this route. Smartbi has served more than 6,000 enterprise customers for 15 years, repeatedly delivering in industries with complex data environments and strict compliance requirements such as finance, government, manufacturing and energy, and solidified the experience of solving cross-system caliber alignment, cross-level permission governance, and regulatory-level data verification, together with the precipitated industry know-how, into the product core. When engineering experience and industry know-how enter the product at the same time, the test of enterprise-level scenarios is no longer "whether it can pass", but "how many times it has passed".

Re-understanding BI and AI: Foundation and engine, not sequential replacement

The reason why Smartbi's shortlisting is worthy of separate interpretation is that it touches a widely misinterpreted relationship.

In the past few years, a narrative has been popular in the AI boom: large models can directly converse with data, and the "intermediate layer" of BI seems to be skipped. However, Smartbi, as a BI vendor, entered this evaluation covering the full data link, and gave the opposite answer with facts — AI not only does not skip BI, but even needs BI.

The reason lies in the different division of labor between the two, not the difference in status.

BI is the foundation. It precipitates the most scarce assets of an enterprise: credible data, unique indicator calibers, verified analysis models, hierarchically isolated permission systems, and business applications that are already running in daily operations. These assets seem to be "slow", but they are the solidification of the enterprise's years of operation logic, which is an unavoidable premise for any AI. It is particularly worth noting that these indicator calibers and analysis models often carry distinct industry characteristics — the risk exposure caliber in finance, the capacity utilization rate in manufacturing, and the people-goods-scenario indicator system in retail are not provided by general corpora, but are know-how grown from industry practices. Large models can "talk", but they do not recognize the enterprise's indicator calibers, permission boundaries and business semantics; no matter how smart they are, they cannot "guess" the completely different definitions of the same indicator by the finance department and the business department out of thin air.

AI is the engine. Its value is not to replace BI, but to improve the efficiency of calling, combining and analyzing these existing assets — compressing the attribution process that used to require manual multi-level reporting chasing and step-by-step drilling down into a series of continuous inquiries; arranging the analysis methods scattered everywhere into a complete analysis process. AI makes the assets precipitated by BI "rotate" faster, but it does not generate assets itself.

Therefore, it is a misinterpretation to describe BI and AI as "sequential replacement", and it is also a misinterpretation to describe AI as an independent capability that can be directly deployed without a data foundation. Industry data has provided footnotes for this: IDC expects global enterprise AI spending to reach 940 billion US dollars in 2026 and increase to 2.1 trillion US dollars in 2029, but at the same time it is expected that 50% of AI-driven scenarios will fail to meet the ROI target by the end of 2026, one of the reasons being the "weak data foundation". Behind the coexistence of high investment and high elimination is the same rule — when an enterprise has ununified indicator calibers, untraceable data sources, and lack of standardized definitions of business semantics, no matter how powerful the model is, it can only "guess" on chaotic data instead of "reasoning".

This is exactly the strategic significance of Smartbi's upgrade of the indicator platform to "Indicator Center" and the upgrade of Baize to Baize V5 this year: it is not "transforming to do AI", but to make the foundation more solid, and then make the engine run faster. The Indicator Center takes "unique indicators, credible data, efficient management, and easy data usage" as the core, eliminates the chaos of "one indicator with multiple calibers" through the global duplicate checking mechanism, which not only provides reliable support for BI, but also builds a solid data foundation for AI applications; Baize V5 adopts the two-wheel drive of "indicator system + multi-agent collaboration", allowing AI to conduct in-depth analysis on this credible base.

The foundation and the engine are indispensable.

Starting from the enterprise's stage: When to supplement BI first, and when to introduce AI

After the relationship is clarified, when it comes to enterprises, the problem becomes a more pragmatic choice: in limited resources, what to do first?

The answer does not lie in "choosing BI or AI", but in judging what stage the enterprise's own data foundation is at. The two are capabilities in the same data system that solve different problems at different stages, not mutually exclusive options.

Under what circumstances should we supplement BI first?

When the following signals appear in the enterprise, the priority should be placed on consolidating the base instead of rushing to deploy AI:

The indicator calibers are not unified, the same "sales revenue" has different algorithms in different departments, and the numbers cannot be matched across departments; the permission system is missing or chaotic, everyone can view the data, and unauthorized access cannot be isolated; the data source is untraceable, and the source cannot be located when a problem occurs; analysis applications are still imperfect, and even stable reports and self-service analysis have not run smoothly.

Introducing AI at this time is equivalent to building a tall building on a soft foundation — the "smarter" the model is, the more "confident" the wrong conclusions it outputs, and the greater the misjudgment it causes. Unifying calibers, governing data, and establishing permissions to make basic data credible first is the most cost-effective investment at this stage.

Under what circumstances is it suitable to introduce AI?

When the enterprise has already precipitated data models, indicator systems and permission systems, and daily reports and self-service analysis have been running stably, the value of AI emerges — it can upgrade the interaction mode from "viewing reports" to "asking for answers" on the premise of not overthrowing existing assets, and upgrade the analysis mode from "manual drilling down" to "intelligent attribution".

This is exactly the implementation logic of Smartbi Baize V5: different from building a separate independent AI analysis system, it can reuse the enterprise's existing indicators, data models, permissions and analysis assets, so that AI analysis is built on the enterprise's existing business calibers. For enterprises that have already built BI, there is no need to start from scratch; for enterprises whose data foundation is still being improved, they can also gradually introduce AI capabilities starting from the most urgent report, indicator or analysis scenarios at present.

In other words, BI and AI have never been two paths of sequential replacement, but two capabilities of the same data system at different maturity levels, and finally run collaboratively in the same system.

A implementable path: Starting from business problems, verifying with real data

Whether it is supplementing BI or deploying AI, the implementation should not pursue one-step perfection, but advance gradually along a verifiable path: business problem identification — data foundation assessment — high-value scenario selection — real data POC — production launch — continuous operation.

The first step is business problem identification. The starting point is not to choose which model or product, but to first think clearly about "what is the most worthy business problem to solve". Is the multi-level reporting chasing for performance attribution too slow? Or is the production cycle of business analysis reports too long? The more specific the problem is, the clearer the path will be.

The second step is data foundation assessment. Judge whether the data needed to solve this problem is already credible, whether the indicators are unified, and whether the permissions are in place. This assessment directly determines whether the next step is to "supplement the base" or to "launch the engine".

The third step is high-value scenario selection. Cut in from scenarios with clear indicator foundations, clear business value and high usage frequency, instead of pursuing large-scale coverage from the very beginning. If the choice is correct, value and trust can be seen faster.

The fourth step is real data POC. Use the enterprise's real data for verification, instead of using demo data to prove itself. The value of this step is to expose real problems such as caliber conflicts and permission boundaries as early as possible, and keep risks out of the production environment. The precipitation of industry know-how allows this step not to start from scratch — mature industry indicator templates and analysis models enable POC to quickly verify based on "proven industry experience", instead of repeatedly trial and error on "general capabilities".

The fifth step is production launch. After the verification is passed, enter the real business process, and support compliance capabilities such as localized deployment and data not leaving the domain to ensure that the results are governable and traceable.

The sixth step is continuous operation. AI analysis does not end after one launch, but needs to precipitate experience, iterate calibers, and expand scenarios in continuous use, so that analysis capabilities become organizational assets.

This path has been verified in Smartbi's customer cases. In a large insurance group, Baize sorted out 50 dimensions and 400 core indicators around financial management, changing performance attribution from "multi-level reporting chasing" to "step-by-step drilling down, instant query and instant answer"; in a provincial government affairs center, the generation cycle of government affairs reports was shortened from 2 to 3 days to the minute level. One of these two cases is in the finance industry and the other is in the government affairs industry, which is a direct embodiment of industry know-how — neither of them is rigidly applied with a set of general templates, but relies on the industry indicators and rules precipitated by Smartbi in the finance and government affairs fields to quickly enter vertical scenarios. The common point of these two cases is not how dazzling the technology is, but that they are all products starting from clear business problems, verified with real data, and finally entering the real business process.

From "Can AI complete analysis" to "Can AI analysis be truly launched in complex enterprises"

Following this path, the competition standard in the Data Agent era is undergoing an imperceptible transition.

In the past, the market cared about "whether AI can complete analysis" — whether it can converse, generate charts, and give a decent conclusion. This tests the model capability. But when AI moves from the exhibition booth to the main business process, the real problem becomes: "Can AI analysis be truly launched in complex enterprises?" This tests the data foundation and delivery capabilities beyond the model.

This difference is essentially not the strength of the model capability, but the thickness of data assets. From the perspective of the demand side, whether a Data Agent vendor can be truly "usable" in complex enterprises usually depends on having four core capabilities at the same time: the BI base, which precipitates credible data, unique indicator calibers, verified analysis models and hierarchical permissions, is the underlying foundation for AI to see the real enterprise context; unified indicators, so that AI's answers have the same traceable caliber source; industry know-how, which productizes the long-precipitated indicator templates and business rules in finance, government, manufacturing, energy, retail and other industries; enterprise-level delivery capability, which enables AI projects to be privately deployed, data not leaving the domain, and enter the production environment of highly sensitive scenarios. The common point of these four capabilities is that they are difficult to make up through short-term technology catch-up, which is also the reason why pure AI vendors generally encounter setbacks in enterprise-level scenarios. Smartbi's entry from BI means that it does not build a building on open ground, but installs an engine on the existing foundation.

In other words, for AI to be truly "usable" in complex enterprises, three problems must be solved first: what data it bases its judgment on, what permission boundaries it follows, and what business rules it understands.