Article body
Full article
I. Where Is the Ceiling for ChatBI?
ChatBI is undoubtedly the hottest concept in the BI industry over the past two years. Asking data questions and metrics in natural language sounds wonderful — no more learning SQL, no more dragging charts, just speak and get answers. But if you’ve used ChatBI in real work, you’ve probably encountered this scenario:
You ask “What was the sales amount in East China last month?” and ChatBI gives the correct answer. Then you follow up with “How does it compare to the month before last?” and ChatBI says “Sorry, I cannot understand your question.” Or you ask “What caused the sales decline in East China?” and ChatBI responds with a bunch of vague platitudes with no actionable value.
Where is the problem? Most ChatBI products only solve the “Ask” step — translating natural language to SQL, executing the query, and returning results. But the real workflow of data analysis goes far beyond that. A complete data analysis workflow contains at least three phases:
- Ask: Understand business questions, match correct data semantics, generate accurate queries
- Model: Data preparation, cleaning, joining, metric modeling, dimension design
- Deliver: Visual presentation, report creation, insight summarization
A ChatBI that only does Ask is at best a “data search engine.” HENGSHI’s proposed Agentic BI aims to let AI Agents cover the entire Ask → Model → Deliver chain, becoming a true “data work assistant.”
II. What Is Agentic BI? How Is It Different from Traditional BI and ChatBI?
“Agentic” in Agentic BI comes from the concept of AI Agent. Unlike traditional LLM conversations, AI Agents have three core characteristics: contextual awareness, tool usage, and action execution.
In a BI context, these three characteristics mean:
- Contextual Awareness: The Agent doesn’t just understand your current question — it understands your business context, data structure, and historical analysis habits
- Tool Usage: The Agent can invoke various tools — data connections, SQL generation, chart rendering, report formatting
- Action Execution: The Agent doesn’t just “tell you the answer” — it can “get things done for you” — creating dashboards, updating reports, sending notifications
III. HENGSHI Data Agent Family: Three Specialized Agents with Distinct Responsibilities
HENGSHI’s AI analytics agents are not a generalized “universal Agent” — they are split into three specialized agents, forming the Data Agent Family:
Agent 01: Data Q&A Agent
Positioned as “a coding workbench for DBAs and data scientists.” It’s not just simple Q&A — it’s a data engineering assistant. For example, if you tell it “Help me create a dataset containing the past 12 months of order data, aggregated monthly, with customer information joined in,” it automatically completes the dataset creation and SQL logic. The resulting dataset can be directly used for modeling analysis and visualization.
The core technical challenge is NL2Metrics — not simply translating natural language to SQL, but translating it into metric-semantic-layer-based queries. This means the Agent must first understand the metric definition of “sales amount” (is it with or without tax? Is it the order amount or the actual payment amount?), then generate the correct query.
Agent 02: Visualization Creation Agent
This Agent is responsible for dashboard and big-screen creation. If you tell it “Create an East China regional sales cockpit for me, including monthly sales trend, customer distribution map, TOP 10 products ranking, and year-over-year growth rate,” it automatically plans the page layout, selects appropriate chart types, creates charts, and assembles them into a dashboard.
Its coverage includes: data source connection, dataset configuration, dashboard creation, chart generation — all implemented through drag-and-drop and configuration. It also supports data scientists and senior programmers who need to handle more complex scenarios through code. Pre-built functions and models are ready to use, lowering the barrier to creation.
Agent 03: Modeling Agent
A zero-code modeling assistant for business users. Business users don’t need to know SQL — they can complete data modeling through natural language descriptions: “Join the order table and customer table by customer ID, and create a ‘customer purchasing power’ metric defined as the cumulative order amount over the past 12 months.” The Agent automatically completes table joins, field selection, and metric calculations.
IV. NL2Metrics: The Technical Foundation of ChatBI Accuracy
The biggest pain point of ChatBI is accuracy. If a user asks “What was the GMV last month?” and the AI generates SQL with the wrong field (e.g., using data that includes refunds), the results will mislead business decisions.
HENGSHI’s solution is NL2Metrics — not translating natural language directly to SQL, but first translating it into a metric query. It passes through a metric semantic layer (Metrics Semantic Layer), which uses HQL (Hengshi Query Language) to define the exact calculation logic of each metric.
The process works like this:
- User asks: “What was the GMV in East China last month?”
- NL2Metrics engine decomposes the question into: time range = last month, region = East China, metric = GMV
- Look up the definition of “GMV” in the metric semantic layer: SUM(order_amount) WHERE order_status != ‘cancelled’
- Generate the final query based on the metric definition, adding time range and region filter conditions
- Execute the query and return results
In this process, the metric semantic layer acts as a “guardrail” — no matter how the user asks, the calculation logic is determined by the metric definition, and the AI won’t “improvise.”
V. HENGSHI CLI: The Agent’s “Hands and Feet”
Having a “brain” (NL2Metrics + LLM) isn’t enough — the Agent also needs “hands and feet” to execute specific BI engineering operations. That’s the positioning of HENGSHI CLI — a BI command-line execution layer for AI Agents.
The CLI provides a standardized command system (using hbi as the entry point), covering the full BI engineering chain: data connection management, dataset creation, HQL queries, dashboard generation, and permission management. AI Agents complete actual BI operations by invoking these commands, rather than improvising API calls on the spot.
More importantly, the CLI has a built-in --dry-run rehearsal mechanism and SSE real-time echo. Every operation the Agent executes can be “rehearsed” before it actually takes effect, and operations personnel confirm before formal execution. During execution, the frontend Web UI can display the Agent’s operation progress and results in real time — this is HENGSHI’s defined “machine executes, humans verify” human-machine collaboration model.
VI. From “People Find Data” to “Data Finds People”: The Ultimate Form of Agentic BI
The true value of Agentic BI isn’t just making data queries more convenient — it’s changing the data consumption model.
The traditional BI model is “people find data”: business users need to proactively open reports, drag filters, and read data to gain insights. Even with ChatBI, it’s humans proactively asking and AI passively responding.
The ultimate form of Agentic BI is “data finds people.” A resident Agent continuously monitors business metrics, and when it detects anomalies, it proactively sends alerts and analysis reports. For example, every morning it automatically generates a “Yesterday’s Business Brief” and pushes it to the DingTalk group. When East China sales decline more than 15% year-over-year, it immediately triggers attribution analysis and notifies the regional manager.
HENGSHI’s HENGSHI BOX is already implementing this scenario — the built-in resident Agent can complete timed report pushes, anomaly monitoring and alerts, and other automated tasks, transforming data insights from “queried” to “pushed.”
VII. Recommendations for Data Teams: How to Implement Agentic BI?
If you’re considering implementing Agentic BI in your organization, here are some practical recommendations:
- Build the metric semantic layer first: The accuracy of Agentic BI depends on the completeness of metric definitions. First organize your core business metrics, unify the logic, then let Agents provide services based on these metrics.
- Start with small scenarios: Don’t try to have the Agent automatically complete all BI work from the start. Begin with the simplest data Q&A, gradually expand to modeling assistance and report generation, and finally move to full-process automation.
- Keep human review: The dry-run mechanism and human-machine collaboration model are necessary. Agent outputs need human verification, especially for operations involving data logic and business rules.
- Pay attention to security boundaries: Clarify what operations the Agent can and cannot perform. Permission control, data masking, and audit tracing are all essential.
Agentic BI is not a replacement for traditional BI — it’s adding an AI layer on top of BI, evolving data analysis from “human-driven tool operation” to “AI-assisted intelligent process.” HENGSHI has traveled quite far on this path — from NL2Metrics to Data Agent Family, from HENGSHI CLI to HENGSHI BOX, forming a complete technical stack. For data teams, it’s time to seriously evaluate Agentic BI.