Whitepaper source
Whitepaper: Governed Data Agent for Operational Intelligence
WP26 technical whitepaper: A governed data agent is not a SQL chatbot; it is an evidence-packaging layer for operations.
Governed Data Agent for Operational Intelligence
Executive Summary
Assumed audience: platform engineers, enterprise architects, integration owners, and technical reviewers evaluating governed data intelligence capabilities in hotel distribution.
TL;DR: A governed data agent is not a SQL chatbot; it is an evidence-packaging layer for operations.
Central claim: A governed data agent is not a SQL chatbot; it is an evidence-packaging layer for operations.
HotelByte’s Data Agent provides a governed conversational analytics layer for platform operations. It helps operators investigate questions that span operational MySQL data, TDengine telemetry, and repository-defined business logic while preserving strict permission, masking, visualization, and audit controls.
The key design choice is to treat the agent as an evidence assembly system rather than a raw SQL console. Natural-language prompts are converted into bounded query intent, source-specific read plans, repository evidence, masked result artifacts, and human-confirmable recommendations. This enables faster incident triage and data reconciliation without bypassing established operational controls.
The first implementation slice is intentionally narrow and verifiable: platform-only users can ask supplier reliability questions; the backend reads retained TDengine supplier reliability aggregates, uses raw hb_log only for drilldown inside log retention, attaches repository evidence, masks supplier identifiers, and returns typed artifacts for the UI. Unsupported or unavailable sources are reported as evidence gaps.
Scope
This whitepaper covers the architecture, controls, and verification model for the Data Agent capability:
- Natural-language investigation across operational data and telemetry.
- MySQL and TDengine read adapters with policy enforcement.
- Repository and documentation evidence retrieval.
- Sensitive-data masking before reasoning and rendering.
- Multi-turn conversation state.
- Visual and tabular result artifacts.
- Auditability of prompts, query plans, evidence, and recommendations.
The intended audience is enterprise technical decision makers, data leaders, platform operations teams, and security reviewers.
Objectives
- Operational Velocity: Reduce time from question to evidence-backed explanation during data-ops and incident investigations.
- Cross-System Reasoning: Join durable business state, time-series logs, and code-defined business rules in one workflow.
- Governed Access: Ensure platform-only visibility, backend authorization, source allowlists, and deterministic masking.
- Decision Quality: Render charts, structured tables, and evidence references instead of prose-only answers.
- Auditability: Preserve a reviewable trail from prompt to query plan, result artifact, and recommendation.
Design Principles
Evidence Before Narrative
Every answer is grounded in data and implementation evidence. The agent is expected to return query intent, source references, and repository or documentation context before producing interpretation or recommendations.
Read-Only by Default
The Data Agent is designed for investigation, not mutation. SQL and telemetry access are constrained to read-only adapters, schema allowlists, row limits, and bounded time windows. Operational actions remain human-confirmed drafts.
Masking as Code
Sensitive data is not protected by prompt instructions alone. Credentials, tokens, PII, commercial fields, order identifiers, trace identifiers, and supplier account details are masked by deterministic sanitizer code before they reach model context or browser rendering.
Visualization as a First-Class Artifact
Operational answers must be scannable. The agent emits typed chart and table artifacts so operators can compare trend, volume, confidence, and redaction impact before taking action.
Repository-Aware Interpretation
Hotel operations data often derives its meaning from code: state transitions, supplier fallback rules, mapping logic, credential policy, and integration-specific behavior. The Data Agent includes repository and documentation retrieval to prevent correct metrics from being interpreted with the wrong business rule.
Architecture
flowchart TD
OP["Platform operator"] --> UI["Agent Workbench data-ops profile"]
UI --> AG["Agent API"]
AG --> AUTH["Authorization and audience policy"]
AUTH --> INTENT["Intent parser"]
INTENT --> PLAN["Query and evidence planner"]
PLAN --> GUARD["Read-only policy guardrails"]
GUARD --> MYSQL["MySQL adapter"]
GUARD --> TD["TDengine adapter"]
GUARD --> REPO["Repository and docs retrieval"]
MYSQL --> SAN["Sanitizer"]
TD --> SAN
REPO --> SAN
SAN --> ART["Chart, table, and evidence artifacts"]
ART --> LLM["Reasoning and recommendation layer"]
LLM --> STORE["Conversation and audit store"]
STORE --> UI
Intent Parser
The parser transforms a natural-language question into structured intent:
- selected sources,
- time window,
- supplier/order/trace/session filters,
- requested output types,
- expected aggregation or reconciliation shape,
- safety requirements.
Query and Evidence Planner
The planner produces bounded read plans rather than arbitrary SQL strings. Plans declare the source adapter, allowed schemas, selected columns, joins, filters, grouping, row limits, and sanitizer requirements. SQL or TDengine statements are generated from these plans only after policy validation.
Source Adapters
The MySQL adapter reads durable operational entities such as orders, suppliers, credentials, mappings, wallets, and configuration. The TDengine adapter reads request and log telemetry with required time bounds and aggregation-first query shapes. The repository adapter retrieves code and documentation evidence in bounded chunks with source references.
Current verified behavior:
- TDengine UAT read path returned real
hb_logrows for the last-day hotelRates/rateCount aggregation on May 17, 2026. - MySQL source inventory showed a deployment split: the checked UAT backend connects to
hoteldev, wheresupplier_daily_snapshotis absent and legacyinventory_quality_daily_supplier_snapshothas only2026-04-25rows. - The supplier reliability path no longer probes MySQL snapshot candidates. It retains TDengine
hblog_ns.hb_log_supplier_reliability_dailyas the historical supplier SLA source, useshblog_ns.hb_logfor raw drilldown inside retention, and treats the snapshot tables as deleted/stale sources, not fallback truth. - The frontend renders only returned
data_ops_resultartifacts; it does not render static sample result rows.
Sanitizer and Artifact Builder
Raw source output is normalized into typed frames, sanitized, and then assembled into artifacts:
- chart series,
- masked result tables,
- evidence references,
- SQL intent summaries,
- recommendation drafts.
The model explains artifacts; it does not own the artifact schema.
LLM Operating Model
The LLM is downstream of retrieval and policy. It receives a governed evidence package: intent, bounded query plan, masked rows, repository references, and source gaps. It is not allowed to execute arbitrary SQL directly or decide whether a secret can be shown. The model can interpret the artifact, ask for a narrower follow-up scope, and draft recommendations for human review.
Knowledge Catalog
The agent knows where to look through a maintained source catalog that maps question families to data sources, schemas, dimensions, metrics, and repository anchors. For example, hotelRates supplier reliability maps first to TDengine hblog_ns.hb_log_supplier_reliability_daily for retained supplier history, with hblog_ns.hb_log reserved for recent drilldown. This is deterministic adapter routing, not LLM table inference. The catalog must be updated alongside migrations and business-rule changes.
Implemented Control Summary
| Control | Customer Value |
|---|---|
| Platform-only agent surface | Limits broad operational analysis to authorized platform users. |
| Backend authorization boundary | Prevents UI visibility from becoming the final security decision. |
| Read-only query planning | Reduces risk of accidental production mutation from generated SQL. |
| Source and schema allowlists | Keeps investigations inside approved operational data domains. |
| Default sensitive-data masking | Protects credentials, tokens, PII, commercial fields, and operational identifiers. |
| Repository evidence references | Grounds explanations in the business logic that actually runs the platform. |
| Typed visual and table artifacts | Enables deterministic rendering and review before follow-up action. |
| Multi-turn context retention | Lets operators refine time windows, suppliers, and traces without losing evidence. |
| Human-confirmed recommendations | Keeps operational judgment with the operator, not the model. |
| Audit trail for prompts, plans, and artifacts | Supports incident review, compliance checks, and model-output accountability. |
Auditability
Every investigation should be reconstructable:
- who asked the question,
- which profile and audience were used,
- what scope and sources were selected,
- which permission decisions were made,
- what query plan was generated,
- which source adapters executed,
- what sanitizer classes were applied,
- which evidence references were retrieved,
- which artifacts were rendered,
- what recommendations were produced,
- what the operator confirmed or rejected.
Audit records must avoid storing raw secrets or unmasked sensitive values.
Safety and Failure Behavior
The Data Agent fails closed:
- Missing or ambiguous permission results in denial.
- Unsafe SQL is rejected before execution.
- Unclassified sensitive columns are suppressed or masked.
- Unbounded time-series queries are rejected.
- Repository retrieval failures are reported as evidence gaps, not filled with guesses.
- Recommendations remain drafts until a human confirms an operational action.
Verification Model
External reviewers can evaluate the capability through:
- route and audience checks,
- profile availability checks,
- generated query plan inspection,
- read-only adapter tests,
- schema allowlist tests,
- sanitizer test vectors,
- chart/table artifact schema validation,
- audit log completeness checks,
- non-platform denial tests.
Differentiation
Traditional dashboards are fast for known questions but weak for new operational investigations. Raw SQL consoles are flexible but expose too much risk. General chatbots can summarize but usually lack data permissions, masking, and implementation evidence.
The Data Agent combines these surfaces: dashboard-like visualization, SQL-like flexibility, repository-aware explanation, and governance controls designed for production operations.
Authoritative Source References
| Reference Area | HotelByte Control Mapping |
|---|---|
| OWASP guidance for LLM application security | Prompt input, generated output, and tool execution are constrained by policy checks, structured artifacts, and deterministic sanitization. |
| NIST AI Risk Management Framework | The design emphasizes governance, traceability, human oversight, and measurable risk controls around AI-assisted decisions. |
| ISO/IEC 42001 AI management systems | Audit records, performance evidence, and decision traceability support AI management review. |
| Database least-privilege practice | Read-only credentials, schema allowlists, bounded queries, and DML rejection reduce blast radius. |
| Data minimization practice | The sanitizer and artifact builder expose only the fields needed for the operational question. |
Technical Whitepaper Governance Reading
Read Governed Data Agent for Operational Intelligence through the technical whitepaper governance loop: intent, evidence, bounded execution, verification, and durable governance.
| Plane | What to inspect in this paper |
|---|---|
| Intent | Which operational or integration risk the design removes. |
| Evidence | Which logs, metrics, records, traces, tests, or replay artifacts prove the behavior. |
| Execution boundary | Which layer owns the decision and which layer only adapts or transports data. |
| Verification | Which failure modes are tested beyond the happy path. |
| Governance memory | Which rules, dashboards, audit trails, or test cases make the lesson reusable. |
Conclusion
Governed Data Agent for Operational Intelligence matters because it turns a fragile implementation concern into a governed platform capability. The durable value is not that the component exists, but that its boundaries, evidence, failure semantics, and verification path can be reviewed after the fact.
A governed data agent is not a SQL chatbot; it is an evidence-packaging layer for operations.
Comments