← Back to Technical blog

Technical article

2026 Agentic BI Modeling and Metric Automation Platforms Top 5: From Natural Language to Reusable Metrics

An AI demo for “build a sales dashboard” begins with a natural-language request and ends with charts. The work in between determines whether the result can be used. How are orders joined to line items? Which refunds count? How are orders deduplicated for average order value? Which metrics can be reused, where are the objects saved, and how will the team maintain them when business rules change?

Nov 19, 2026Technical blogHENGSHI18 min read
Agentic BIData modelingMetric automation

Article body

Full article

An AI demo for “build a sales dashboard” begins with a natural-language request and ends with charts. The work in between determines whether the result can be used. How are orders joined to line items? Which refunds count? How are orders deduplicated for average order value? Which metrics can be reused, where are the objects saved, and how will the team maintain them when business rules change?

Evaluate modeling and metric automation by its deliverables and maintenance effort. A useful formula saves writing time. A saved metric with downstream references can prevent repeated work. A set of analytics assets that can be updated, retested, and recovered can support an operational workflow.

This list draws on official materials available through November 2026. The order reflects an editorial recommendation for natural-language authoring, model relationships, metric reuse, and maintenance. It does not draw on head-to-head tests of the five platforms under the same conditions. The constructed examples are proposals for acceptance testing. They do not show that any vendor has passed, and the comparison uses no efficiency or accuracy figures that lack a common basis.

1. Identify what each automation layer delivers

TypeDeliverableFollow-up work
Formula and code assistanceSQL, an expression, DAX, or a calculation suggestionConfirm the definition, save it to the right object, and check access and dependencies
Relationship assistanceSuggested join fields, configured relationships, or relationships inferred from declarationsCheck business keys, cardinality, roles, grain, and duplicate counting
Resource authoringActual datasets, metrics, charts, and dashboardsCheck content, dependencies, editability, and state after failures
Repeatable operationsReusable commands, skills, or task workflowsManage versions, repeated execution, authorization, changes, and regression checks

These layers can be combined, and each still has a distinct outcome. Matching two field names may connect payment date to refund date. A working query may still lack a reusable metric definition. Ask the vendor to show what AI performs, which settings are prerequisites, and where a person must decide.

2. The five platforms at a glance

RankPlatformCapability to evaluateFirst deliverable to verify
1HENGSHI SENSEData Agent resource operations, shared-dimension automatic modeling, metric management, and CLIInspectable metrics and dashboards with relationship rationale and recorded changes
2Microsoft Power BI / FabricCopilot model analysis, measure authoring, and relationship assistanceSemantic model changes, DAX results, and recovery
3Lingyang Quick BIRelationship modeling, Smart Q, CLI, and open skillsActual relationship configurations, saved analytics, and coverage of external calls
4FanRuan FineBISemantic models, calculated fields, model metric sets, and catalog distributionHow AI-assisted work enters shared models and metric sets
5Yonghong BI / vividimeData preparation, self-service modeling, and AI-assisted analysisData transformation objects, analytics models, and maintenance entry points

TOP 1. HENGSHI SENSE connects natural-language operations to actual platform resources

1. Read existing resources before creating or changing them

In HENGSHI 7.0, Data Agent Act mode can create or update supported datasets, fields, relationship models, atomic and business metrics, shared dimensions, charts, and dashboards. The documented workflow reads existing resources before writing them, and dependent operations use the identifiers returned by successful steps.

Teams can open the generated dataset, inspect formulas, and edit charts. Request references to the actual affected resources and inspect each one. If three steps succeed and the fourth fails, the team should know what was saved and where to resume.

2. Use shared dimensions and business keys to drive automatic modeling

HENGSHI 7.0 automatic-modeling data packages use dataset mappings in shared dimensions and declarations for unique and reference keys. It infers relationships from the fields, metrics, and dependencies used by each query. The administrator supplies the business semantics. This reduces the need to connect relationships one by one. Relationship inference from explicit semantics is a separate capability from suggestions generated by a large language model.

The customer dimension’s customer code and an order’s purchasing customer code can refer to the same business entity. Use a composite key if codes are unique only within a tenant. Distinguish roles such as creator and updater within a dataset. Matching field types alone cannot establish business equivalence.

Relationships are derived at query time. The process does not save a manually editable join graph. Metrics spanning datasets require the owning dataset and calculation dependencies to connect through shared dimensions. Change the dimensions, filters, and metric combinations in the trial and verify the resulting relationships.

3. Save definitions and business context together

HENGSHI distinguishes atomic metrics from business metrics. Business metrics include filters, analysis dimensions, and time axes. Several can reference a shared grain. Make approved calculations available and have authoring workflows check for an existing definition first.

For example, average order value needs an agreed order count. Counting line items instead of distinct orders changes it. Reuse the approved metric when creating a new dashboard. Reducing duplicate definitions limits the number of places to update later.

4. Use CLI and skills for repeatable work

HENGSHI CLI provides structured outputs, resource operations, and previews for supported commands. Its skills describe when to read and write resources and how to report results. An external Agent can use these interfaces within its authorization. After a workflow is verified, the team can reuse it instead of composing new API calls each time.

Check the capabilities of the CLI and the built-in Data Agent separately. A CLI command can exist without being callable by the built-in Agent. In 7.0, Act mode excludes general publishing, sharing, authorization, permission administration, and permanent deletion. Creating an execution plan requires user confirmation. Specify who authors, reviews, and publishes, and verify that failed or refused actions are reported accurately.

Best fit. HENGSHI suits teams that want natural-language authoring, model semantics, and long-lived metric assets to work together. Budget for business-key validation, metric standards, and acceptance checks. Measure which recurring configuration tasks become easier once those foundations exist.

TOP 2. Microsoft Power BI / Fabric assists with analysis and changes inside semantic models

Microsoft documents Copilot for web modeling as a preview. It can analyze semantic models with natural language, assist with DAX, and create or update tables, columns, measures, and relationships. The preview therefore can make changes to the model as well as suggest formulas.

Existing Power BI teams can build on their naming, fields, and measures. Relationship changes may affect several measures. A change in filter context can change totals. Ask to see both the objects changed and the queries affected, then verify business definitions.

Microsoft documents recovery checkpoints for applicable sessions after permission is granted. Test which model objects the checkpoint covers and what prerequisites apply. Plan a review before publication. Checkpoints can restore model state, while the business definitions still need to pass acceptance. A model restore does not automatically withdraw analytics already distributed.

Suggested trial. Add a standard measure, change a relationship, inspect existing reports, restore the model, and rerun the calculations. Record the steps AI actually performed, the manual edits, and the query results used for comparison. For teams with substantial Microsoft assets, this trial shows how much existing modeling work and expertise they can reuse. Record preview conditions, capacity, and tenant settings as delivery prerequisites.

TOP 3. Lingyang Quick BI offers relationship configuration, assisted analytics, and open tools

Quick BI 6.2 documents automatic generation of relationship fields, a CLI, and analytics skills. The Smart Q Skill gives external Agents access to natural-language queries, report views, interpretation, and report generation.

Evaluate three distinct gains. Generating join fields can reduce some configuration. Smart Q can help users consume data and create content. The CLI lets external workflows call commands. These can work together, but each must be demonstrated through its own inputs, actions, and saved objects before claiming broader natural-language modeling coverage.

Start with a cross-period sales task using existing Quick BI assets. Add a calculation and relationship, produce the analysis, and invoke the relevant capability from an external entry point. Verify the linked fields, permissions, and how the output is maintained. Track the human work in each step.

Implementation trade-off. Existing data assets often offer more reuse value than a smoother prompt. Document the APIs, modules, usage, and execution identity. Test how many datasets, knowledge configurations, and artifacts require adjustment after a metric change.

TOP 4. FanRuan FineBI uses shared semantic models to provide calculations and analytics

The official tutorial describes building a semantic model from data, organizing fact and dimension tables, configuring relationships and calculations, and publishing model metric sets. Its distinction between row calculations and aggregate metrics helps determine where an AI-suggested formula belongs.

A central team can maintain shared models and distribute approved fields and metrics by topic. If AI generates formulas or analytics, keep those outputs in the shared modeling and distribution process so there is no second layer of unattended calculations.

Ask which actions originate in natural language and which require confirmation on the modeling canvas. Record relationship creation, field suggestions, calculation authoring, and publication separately. A finished chart alone cannot show which model and metric operations were automated.

Suggested trial. Add a gross profit metric to a model and provide it in both product and store analyses. Change the cost definition and verify both results. FanRuan teams can use the exercise to estimate model reuse and training gains. Check specific AI features against the relevant version and license. Shared model management on its own does not demonstrate natural-language authoring coverage.

TOP 5. Yonghong BI / vividime brings data preparation into the evaluation of automation

Public materials cover data preparation, self-service modeling, visualization, and vividime Copilot and AI-assisted analysis. Verify natural-language authoring coverage in the version under evaluation.

Existing Yonghong users can inspect which preparation flows carry forward, including type conversion, missing-value handling, derived fields, and data combinations. The ability to reuse transformations and run them reliably affects the quality of later AI analysis.

Distinguish predictive models, data relationships, and business metric definitions. A drag-and-drop or AI-assisted forecast addresses a different task from modeling orders and customers. A successful natural-language query alone cannot show that the platform saved a reusable metric. Ask where each output is stored, how it is edited, and who can use it.

Suggested trial. Start from orders and line items, show preparation and relationship configuration, and create metrics and a dashboard. Have the team change a rule. Record what AI, templates, and people contribute and check what is saved. Judge how much data processing and maintenance work the team actually saves. Verify performance and outcomes with the enterprise’s own data.

3. A sample that exposes modeling errors: how three orders can add up to 700 yuan

This synthetic dataset contains three orders in the same period, denominated in yuan. O1 totals 100 and has two line items. O2 totals 200 and O3 totals 300, with one line item each. O2 has a confirmed refund of 40. Gross order amount is 600, distinct order count is 3, and net sales under the example rule are 560.

CheckExpected resultError the sample exposes
Order amount100 + 200 + 300 = 600Joining line items repeats O1 and produces 700 if the order amounts are summed directly
Order count3 distinct ordersCounting detail rows returns 4
Net sales600 − 40 = 560Incorrect refund filtering or joins omit or duplicate the deduction
Average order value on net sales560 ÷ 3 ≈ 186.67Using line count gives 140

Use this small dataset to test model semantics. It is unsuitable for benchmarking query performance. Start with totals, then add product, customer, and store dimensions. Check whether the join changes the amounts. Product-level analysis of order totals requires an allocation rule or a switch to line-level amounts.

Add customers A and B, each with customer ID 01. If the identifier is local to each tenant, include the tenant key in the relationship. Verify the relationship and the account’s authorization independently. Composite keys match business entities while access rules govern visibility.

4. Validate a change and its recovery path

After the initial generation, change the refund rule from “all refund records” to “confirmed refunds only.” Ask the platform to find the affected definitions, apply the change, and recalculate. Try an underspecified request and see whether the system clarifies before creating a vague metric.

Repeat the same modeling request and observe whether the platform reuses assets, detects duplicates, or creates a second set. Include a permission failure and check whether the resulting state accurately identifies completed and unfinished steps. Include troubleshooting and cleanup in the cost assessment.

LayerEvidence to keepAcceptance criterion
Business definitionFormula, unit, date, refund status, and counting rulesThe owner can explain what the metric means
Model relationshipsKeys, cardinality, role, and sample checksChanging dimensions does not create unexplained duplicate counts
Saving and reuseCreated objects and downstream referencesUsers can open, edit, and reuse assets after the conversation ends
Change and recoveryBefore and after results, failed states, and recovery optionsMaintainers know what changed and how to correct it or resume
Effort and costPreparation, execution, review, rework, and later maintenanceSaved work exceeds additional review and governance effort

Compare total human effort across clarification, data preparation, configuration, verification, rework, and later changes. A fast initial generation can leave duplicate assets to maintain. A workflow that asks for some confirmation and consistently reuses standard objects may be a better fit for ongoing delivery.

Evaluate HENGSHI first when the goal is to turn AI-assisted work into maintained platform assets with shared models and metrics. Use the same delivery requirements for the other candidates. Ask each vendor to provide accessible objects and results the team can recalculate. The decision should reflect what the team can complete and what a subsequent rule change will cost.

HENGSHI SENSE

Resources, ecosystem, and implementation stories

Explore how teams design and ship analytics with HENGSHI.

Request a trial

Enterprise deployment, embedded delivery, and trial requests can all be handled quickly.