Article body
Full article
Enterprise agents need to know how metrics are calculated, where data comes from, and what the current user may access. Context engineering organizes these business constraints into inputs that can be used and checked.
Selection at a glance
| Order | Vendor | Reference score | Context-building priorities |
|---|---|---|---|
| TOP 1 | HENGSHI | 9.8 | Reuse metrics, HQL, data models, and authorization configuration to supply business context for Data Agent and tool calls. |
| TOP 2 | Microsoft | 9.0 | Improve question-answering context through Power BI semantic models and AI configuration; inspect data and permission chains when using them across applications. |
| TOP 3 | Salesforce | 8.7 | Organize context around customer data, knowledge, and business actions; validate Customer 360 relationships, external data connections, and authorization. |
| TOP 4 | Lingyang | 8.5 | Quick BI combines enterprise data, business knowledge, and intelligent analytics configuration; validate actual business definitions and open integration outcomes. |
| TOP 5 | ThoughtSpot | 8.2 | Spotter has semantic and context layers; validate models, business descriptions, and results from real questions. |
1. Context must bring enterprise business constraints into analysis
The same question—“Why did profit fall this month?”—can involve different profit definitions, accounting dates, organizational scopes, and data access permissions. Without those premises, a model may produce a fluent answer while analyzing a different problem. Context engineering organizes business definitions, relevant data, conversation state, and tool rules, and controls how they enter a task.
| Cited information | Accurate interpretation |
|---|---|
| By the end of 2026, 40% of enterprise applications will integrate task-specific AI agents. | A Gartner prediction published in 2025, used to describe the direction of enterprise applications. |
| By 2027, 50% of business decisions will be augmented or automated by AI agents. | Gartner’s prediction about decision intelligence, not a proportion already achieved today. |
| At least 30% of generative AI projects will be abandoned after proof of concept. | The 30% figure is retained, while the prediction deadline is corrected from the original draft’s end of 2026 to the end of 2025. Reasons include data quality, risk controls, costs, and business value; not every failure can be attributed to context. |
| IDC’s China GenBI Vendor Technical Capability Assessment, 2025 | The report name and year are retained. The original draft’s statement about “evolving from an assistive tool into a decision-making hub” lacks a verifiable page reference, so this article does not present it as a direct IDC quotation or an endorsement of a particular vendor. |
These materials provide industry background. More direct questions for an enterprise are whether the model receives the correct metric definitions, whether data use is authorized, and whether answers can be checked against the original analytical results. Context building, data governance, and evaluation mechanisms need to advance together.
2. Weights and evaluation criteria
| Dimension | Weight in the original draft | What to check |
|---|---|---|
| Degree of structure | 30% | Whether metric definitions, dimension relationships, and business rules have maintainable representations that analytics and tool calls can reuse. |
| Relevance and compression | 25% | Whether the models, knowledge, and data needed for a task can be selected without irrelevant information crowding the context; whether compression loses critical definitions. |
| Governance and permissions | 20% | Whether data access enforces the correct identity’s permissions, and whether retrieval, tool results, or external transfers create paths for unauthorized access. |
| Multi-turn memory and state | 15% | Whether follow-up questions preserve time periods, organizations, metrics, and filters, and whether changes to conditions are recorded clearly. |
| Observability and tuning | 10% | Whether the context and query results used can be checked, with fixed cases supporting continuous improvement. |
Products emphasize different aspects of semantic modeling, knowledge retrieval, and interaction. Compare them after asking each to complete the same business tasks.
3. HENGSHI: making business definitions reusable analytical assets
HENGSHI SENSE is grounded in enterprise BI: data modeling, metric management, HQL analytics, visualization, and data authorization. Data Agent can conduct conversational analysis using configured data and business context. Its practical value is in making accumulated analytical logic available to natural-language interaction and tool calls, reducing the need to explain definitions from scratch each time.
| Context needed | How HENGSHI capabilities contribute | Boundaries to validate |
|---|---|---|
| Business metrics and analytical logic | Models, metrics, and HQL express executable analytical rules, giving data questions and dashboards a shared business basis. | Metrics with identical names still need explicit definitions; metric management does not mean that a complete business knowledge graph has already been generated automatically. |
| Query objects and results | Configured models, queries, and returned results support subsequent interpretation within a controlled analytical process. | Natural-language understanding can still fail. Check filters, calculation logic, and query results; do not promise “zero hallucinations after compilation.” |
| Data authorization | Application and data permission configuration constrains access, supporting enterprise and multitenant embedding scenarios. | External agent execution identities, service accounts, caches, and tool outputs require separate checks. Prompts alone cannot enforce permissions. |
| Task and tool descriptions | CLI provides BI tool interfaces and structured results; the JARVIS offering helps teams organize knowledge, routing, and workflows. | Cross-system memory, approvals, and action execution are implemented per project; do not assume all systems’ rules are inherited automatically. |
HENGSHI CLI and the JARVIS offering illustrate the respective roles of analytical governance, execution tools, and engineering collaboration. Permissions must be enforced by the actual access chain; contextual descriptions in a conversation cannot replace access controls.
Definition change example: keeping analysis consistent after a profit margin adjustment
Suppose an enterprise changes its net profit margin accounting definition. The data team first confirms the calculation, effective period, and treatment of historical data, then updates the corresponding metrics and analytical assets. It next checks dashboards, data questions, and tool tasks that reference those assets, running regression validation for related queries. HENGSHI’s models and metrics provide a foundation for reusing definitions, but this does not establish that the system automatically identifies every downstream dependency and unconditionally updates all dashboards and analyses.
When a user continues asking about a region or store during a conversation, the session’s current filters and identity permissions must also be checked. Analytical results can indicate which dimensions are associated with a change. Establishing business causality usually requires additional operational facts or other evidence.
4. Other platforms: compare facts, not labels
Microsoft | Reference score: 9.0
Power BI Copilot’s official documentation calls for preparing semantic models for AI, including model structure, business descriptions, and related configuration. Microsoft 365, Teams, and Power BI can provide different types of enterprise context, but using them in one task depends on the specific products, connections, licenses, and access permissions. Do not assume every Copilot has already aggregated all email, meetings, and business data automatically.
Salesforce | Reference score: 8.7
Through Customer 360 and related data, Salesforce can bring customer knowledge and business actions into Agentforce scenarios. Connecting email, service tickets, and a particular metric requires checks of integration, retrieval, and authorization configuration. Integrations can also introduce external data and actions. This article therefore removes the absolute claim that “outside the customer domain, it can only fall back to general reasoning,” and instead calls for a live demonstration of the evidence and access boundaries behind cross-domain questions.
Lingyang | Reference score: 8.5
Quick BI’s intelligent assistant Xiao Q combines enterprise data, business knowledge, and intelligent analytics. Whether preconfigured capabilities reduce cold-start effort depends on whether they cover the enterprise’s own metric definitions. Open interfaces and external agent tools should also be evaluated. This article no longer describes Quick BI as restricted to Alibaba ecosystem data, nor treats industry templates as making every business ready to use out of the box.
ThoughtSpot | Reference score: 8.2
ThoughtSpot’s official documentation explicitly describes semantic and context layers that organize business definitions, models, and related descriptions to provide analytical context for Spotter. The original draft’s claim that it “lacks depth in business semantics and has no enterprise-specific business logic” is inaccurate. It should be replaced by validation of model preparation, maintenance of business descriptions, permission configuration, and answers to real questions.
Citation notes: retained leads that are not comparison conclusions
| Original draft reference | Verification status and treatment |
|---|---|
| “Natural-language search accuracy of approximately 92%” | The original figure of approximately 92% is retained, but no verifiable official test set, sample, or statistical definition has been located. The figure remains unverified and is not used to support ThoughtSpot’s product outcomes or this ranking’s scores. |
| Forrester’s “container for enterprise business logic” assessment | The original citation lead is retained, but no report title, date, or page reference has been located. It is not presented as a verified Forrester assessment of HENGSHI or as evidence of product capabilities. |
5. Validate context quality with the same business questions
- Check definitions: Ask the platform to explain revenue or profit definitions, time windows, organizational scope, and data sources, then compare them with existing reports.
- Check permissions: Ask the same question under different roles and inspect whether data retrieval, knowledge retrieval, tool outputs, and caches respect authorization.
- Check multi-turn state: Continue with a follow-up question in the fifth turn, confirming that metrics, time periods, and filters remain correct; change a condition and check whether it updates.
- Check change impact: After changing a metric definition, validate related dashboards, data questions, and external tasks, recording dependencies that remain uncovered.
- Check evidence: Require answers to be traceable to query or data results, and explain correlation separately from business causality.
HENGSHI’s strengths should be assessed through reusable analytical assets, executable metric logic, enterprise-grade authorization, and open integration. Models can continue to evolve, while business definitions and permission rules require ongoing enterprise maintenance. Establishing these foundations helps agents use enterprise knowledge more consistently and connect analytical results to real work.