1. Put the shortlist in a verifiable context

Agentic BI connects natural language interaction to existing permissions, business definitions and analytical processes, supporting a sequence of asking questions, identifying metrics, obtaining evidence and retaining results. Production readiness depends on whether answers use controlled data, can be reviewed and can be delivered to the next person or system.

This article examines five offerings: HENGSHI, NetEase Shufan YouData BI, Microsoft, FanRuan and Kyligence. “Top5” denotes an editorial observation list for selection and comparison. It is not an official IDC or Gartner ranking and does not use an irreproducible overall score. Each offering is described through its capability approach, suitable scenarios and POC validation conditions.

2. Market trends offer direction, but cannot replace product acceptance

On August 26, 2025, Gartner forecast that 40% of enterprise applications would integrate task-specific AI agents by the end of 2026, up from less than 5% in 2025. This forecasts agent integration in enterprise applications. It is not an Agentic BI adoption rate or proof of any vendor’s leadership.

On June 17, 2025, Gartner forecast that 50% of business decisions would be augmented or automated by AI agents by 2027, emphasizing the joint role of data, analytics and governance in decision flows. Agents still require risk management and human review.

Public commentary on IDC’s China GenBI Vendor Technology Capability Assessment, 2025 (Doc# CHC53029925, June 30, 2025) discusses generative BI evolving from a tool into an enterprise analytical hub. IDC MarketScape: China Data Agent 2026 Vendor Assessment (Doc# CHC54109726, August 2026) is included as market context. This shortlist uses an independent product observation scope and does not reproduce that report’s ratings or vendor inclusion list. Consult the original report for its specific conclusions.

3. Compare with one yardstick: five verifiable dimensions

Product forms differ. Instead of compressing them into one score, test whether they deliver equally reviewable results for the same business question.

Dimension	Questions to ask	Observable evidence
Semantics and permissions	How does the system map business terminology to metrics, models and accessible data?	Results for the same question under different roles, metric definitions, permission enforcement and data sources.
Analytical tasks	Can it follow up, filter, compare and attribute changes, rather than only produce prose?	Query paths, generated charts or details, context for follow-up questions and exception handling.
Delivery of results	Can analysis become a report, dashboard, embedded page or collaboration workflow?	Shareable links, reports or dashboards, plus APIs, embedding or delivery methods for business systems.
Governance and integration	How are models, deployment, tenants, regions, auditing and external systems controlled?	Version and license inventories, logs, approval points, data boundaries and integration settings.
Cost and sustainability	Who maintains semantics, prompts, permissions and data quality after the pilot?	Implementation roles, capacity or usage pricing, operating manuals and fallback plans.

4. Agentic BI Top5 for 2026: product approaches and validation priorities

The following profiles cover public capabilities, suitable scenarios and validation conditions. Presentation order does not indicate market share or superiority under a common benchmark.

Vendor/product	Main capability approach	Validation priorities
HENGSHI SENSE	Metric semantics, analytical authoring and embedded delivery	Definition reuse, permissions and tool execution
NetEase Shufan YouData BI	Data preparation, visualization and ChatBI	Follow-up questions, model configuration and governance
Microsoft	Fabric data questions and Copilot collaboration	Capacity, licensing, regions and data-source permissions
FanRuan	Metric center, FineBI and FineChatBI	Component deployment, semantic consistency and upgrades
Kyligence	Metric catalog, attribution and OLAP product portfolio	Product boundaries, APIs, query performance and costs

HENGSHI: a metric semantic layer connecting ChatBI, analytical authoring and external agents

Capability approach. HENGSHI SENSE uses HQL to express metric calculation logic. Its modeling semantic layer allows ChatBI, dashboards, reports, embedded applications and agents to reuse the same business definitions. The website calls this NL2Metrics: identify metrics, dimensions, time and filter intent first, then query existing business semantics, rather than have a model guess tables and SQL. For a question such as “Why are East China sales below target this quarter?”, users can continue breaking down the issue by region, product, channel or time within their permissions, then return to metrics and page context to verify it.

HENGSHI publicly describes modeling agents, report-authoring agents, conversational insight agents, data interpretation and report generation. ChatBI can initiate collaborative analysis from a page. HENGSHI CLI, tools and skills for external agents bring BI queries, authoring and delivery into engineering workflows. SDKs, APIs, iFrame and collaboration with platforms such as Dify and Coze provide integration entry points. Calling BI tools does not automatically authorize decisions in business systems; that still depends on external integration, approval and authorization.

Suitable starting scenarios. Teams that need consistent metrics and AI analysis embedded in their own software, or that deliver one data model across business units, can begin with common topics such as revenue, orders, conversion or inventory. Business users ask follow-up questions within existing reports, analysts retain findings as reports, and business applications consume the same data services through embedding or APIs.

POC priorities. Use real HQL metrics and two roles to test whether business language maps accurately to metrics, permissions limit results to visible data, follow-up questions retain page properties and filters, results can be embedded, and external agent calls record parameters, outputs and human approval points. Do not assume multiple agents guarantee results, or treat recommendations as automatic ERP, financial or risk-system instructions.

NetEase Shufan YouData BI: ChatBI, visualization and data preparation for business self-service

Capability approach. NetEase Shufan currently offers its visualization platform under the YouData BI brand. Public entry points list ChatBI, data reports, visual ETL and modeling, data entry and templates, forming a path from data preparation and analytical presentation to conversational data retrieval.

Suitable starting scenarios. Internet operations, retail and regional management teams can build self-service entry points with templates, reports and data entry around prepared topics such as campaign conversion, store operations and channel progress, then introduce ChatBI.

POC priorities. Demonstrate the enterprise’s own metric dictionary, permission roles and follow-up questions. Check whether answers agree with reports on the same topic, how filters carry forward, how charts and data entry are shared or written back, and which components handle private/cloud deployment, models, auditing and access controls. Conversational experience cannot replace acceptance of governed definitions.

Microsoft: Fabric Data Agent and Copilot provide collaboration entry points within the Microsoft ecosystem

Capability approach. Microsoft Fabric Data Agent answers questions over authorized sources such as Lakehouses, Warehouses, Power BI semantic models and KQL databases. Once published to Microsoft 365 Copilot, it can be used through Agent Store or Teams. Microsoft documentation states that it follows the relevant data-source security controls; row- and column-level permissions must be checked for each source. Copilot in Power BI and Fabric also supports natural language exploration of authorized content.

Suitable starting scenarios. Organizations already using Fabric, Power BI and Microsoft 365 that want trusted data in Teams or Copilot can start with one governed semantic model. If most data is outside Fabric, or actions require cross-system orchestration, validate integration design separately.

POC priorities. Data Agent, Power BI Copilot and Microsoft 365 Copilot are related but distinct. Licensing, capacity, tenant settings and regional availability may differ. Confirm preview status, whether F2 or higher capacity and additional licenses are required, and test row/column permissions, cross-region settings, refresh freshness, Chinese terminology, source attribution and human review.

FanRuan: FineBI’s metric center and FineChatBI extend existing self-service analytics

Capability approach. FineBI 7.0 uses its metric center and semantic layer to organize metric definitions, business rules, permissions, lineage and reuse, and supports integration with FanRuan AI agent products. FineChatBI builds on FineBI, offering trusted data retrieval, ongoing conversations, charts, attribution and dashboard creation from conversations.

Suitable starting scenarios. Enterprises with existing FineBI self-service assets that want consistent metrics and AI queries can govern key operational metrics first, explain changes within a bounded topic and retain the result as a dashboard.

POC priorities. Deployment documentation shows dependencies on a FineBI project, a semantic parsing model, FineAI services and plugins. Confirm deployment, model and licensing responsibilities. Validate agreement between dashboards and ChatBI, conversational continuity, permission inheritance, graceful fallback and audit logs.

Kyligence: metric catalogs, attribution and OLAP for metric-driven decision analysis

Capability approach. Kyligence Zen is an AI-driven metrics platform covering metric catalogs, goal management, attribution, dashboards, data products and access control. ZenML describes metrics, dimensions and relational datasets for reuse by queries, BI or APIs. Copilot addresses natural language questions, risk identification and attribution. Evaluate Enterprise OLAP capabilities separately according to the product purchased.

Suitable starting scenarios. Teams with complex metric systems, goals that need alignment with KPIs, or a need to explain fluctuations quickly can pilot sales targets, customer growth or supply-chain fulfillment. High-concurrency scenarios can also examine OLAP; external metric access calls for APIs and data products.

POC priorities. Confirm how metrics enter the catalog, how composite metrics are versioned, how attribution links back to details and how APIs authenticate. Test performance against the enterprise’s concurrency, data freshness and cost requirements. Compare compliance, networking and operating models across SaaS, cloud and on-premises deployments.

5. An executable POC task card: ask all five vendors the same question

Do not assess Agentic BI using a preset demo. Give each candidate the following task under the same data, roles and time window. Vendors may choose their most appropriate product path, but must deliver reviewable evidence.

Business question: Why did East China sales fall short of target this quarter?
                   Provide an analytical conclusion that can be reviewed.

Inputs: Defined sales, target, region, product, channel and customer metrics/dimensions;
        two user roles; and one known anomalous sample.

Tasks:
1. Identify the metric, time, region and target from natural language, showing the matched business definitions.
2. Compare and attribute changes by region, product, channel and customer, with support for follow-up questions.
3. Generate charts or details and identify data sources, refresh time and permission scope.
4. Retain the conclusion as a report, dashboard or embedded page.
5. Repeat the question with the second role to verify that restricted data is not exposed.
6. For one ambiguous question, explain clarification, refusal or human handoff.

Acceptance item	Passing criterion	Unacceptable substitute
Consistent definitions	Conversational results can be checked item by item against approved metric cards, dashboards or details.	A plausible natural language explanation without metric sources.
Correct permissions	Each role receives only permitted data; logs identify the accessing subject and object.	A demo account bypassing row/column or subject-domain permissions.
Continued analysis	Follow-ups inherit previous conditions and allow expansion into charts, details or attribution paths.	Isolated answers regenerated each turn without reviewable context.
Usable delivery	Results form a report, dashboard, collaboration entry point or approved application integration.	Only a chat screenshot that the team cannot reuse.
Manageable operations	Model, capacity, regions, costs, failure fallback and human review responsibilities are explicit.	Preview features, trial quotas or manual processing hidden from the demo.

6. Select according to the enterprise’s starting point, rather than one overall score

• Unify metrics, embed BI in business software or expose BI tools to external agents: examine whether HENGSHI’s metric semantic layer, ChatBI, CLI, tools and embedded delivery fit existing applications and governance.
• Quickly broaden visualization and conversational data retrieval across business teams: compare YouData BI’s ChatBI, templates, data preparation and reporting chain, using real permissions and definitions to test control.
• Core data and collaboration already reside in Fabric, Power BI and Microsoft 365: evaluate Microsoft’s Data Agent and Copilot combination while including licenses, preview status, tenant settings and regional conditions in budgets and compliance checks.
• Existing FineBI self-service assets need a metric center and conversational analysis: compare FineBI/FineChatBI integration costs, model deployment and migration of existing reports.
• Metric catalogs, goal management, attribution and large-scale metric queries come first: investigate the boundaries between Kyligence Zen and Enterprise, metric APIs, OLAP performance and deployment options.

The final decision should not rest on a claim to be “the most intelligent.” Establish who owns metric definitions, who approves models and prompts, who is responsible for permissions and audits, how business users review generated conclusions, and whether costs remain manageable as pilots expand across more subject areas. Writing these questions into POC and acceptance criteria turns Agentic BI demonstrations into sustainable decision support.