AI Explainability and Trust

This document explains how askEdgi uses AI, what data is processed, how results are generated, and what controls are in place to ensure trustworthy, compliant operations. It is designed for decision-makers responsible for evaluating AI systems within governance and risk frameworks.

How askEdgi Uses AI

askEdgi combines natural language processing with enterprise metadata to convert user questions into analytical operations. The system does not operate as a general-purpose chatbot. It is a constrained, governed assistant designed to work within predefined boundaries.

AI Model Architecture

askEdgi uses a hybrid architecture:

Governance Layer (Internal)

Operates within the OvalEdge platform. This layer interprets user questions, validates permissions, retrieves metadata, generates SQL queries, and applies access controls. It enforces governance policies before any data is processed.

Language Understanding Layer (External):

Uses large language models (OpenAI or Gemini) for natural language interpretation. This layer receives only metadata, summaries, and contextual information. It does not receive raw enterprise data.

What the AI Does Not Do

• askEdgi does not learn from enterprise data

• It does not train models on customer datasets

• It does not retain conversation history for cross-user learning

• It does not access data outside the user's permissions

• It does not bypass governance controls

Data Flow and Processing

Understanding what data moves through the system is essential for risk assessment.

What Stays Internal

The following data never leaves the secure environment:

• Raw enterprise data records

• Personally identifiable information (PII)

• Sensitive or regulated data values

• Table contents (row-level values)

• User credentials and authentication tokens

What Is Shared with External AI Providers

Only the following metadata and contextual information are sent to external language models:

• Table names and column names

• Column descriptions and data types

• Semantic summaries and aggregated data profiles (example: "90% of values are between 0 and 100")

• Business glossary definitions

• Tag associations and stewardship details

• User questions in natural language

No raw data values are included in these requests. The external model processes metadata to understand context and generate appropriate responses.

Data Residency by Edition

Public Edition:

Temporary workspaces operate in AWS-hosted environments. Uploaded files and analysis artifacts are stored in encrypted S3 buckets. Files are deleted after session completion per retention policies.

SaaS Editions:

All data remains within OvalEdge's secure AWS infrastructure. Enterprise connectors enforce data residency and access controls. No data is duplicated or moved outside governed boundaries.

On-Prem Edition:

All data remains within the customer's infrastructure. No external AI enrichments or cloud-based processing occur. Metadata analytics run entirely on local systems.

Access Control and Governance

askEdgi enforces access controls at multiple levels to ensure users only see data they are authorized to access.

Role-Based Access Control (RBAC)

Every query validates the user's role and permissions before execution. Users see only datasets and columns they have been granted access to via OvalEdge's catalog governance.

Validation points:

• Before adding catalog objects to the workspace

• Before executing queries on datasets

• Before displaying metadata search results

• Before allowing recipe execution

Data Masking

Masked columns remain masked during analysis. If a user does not have permission to view a column in its raw form, askEdgi will not expose it. Masking policies are enforced automatically without user intervention.

Workspace Isolation

Each user operates in a dedicated workspace. Uploaded files, analysis artifacts, and temporary datasets are not shared across users, even within the same organization. This prevents unintended data exposure and ensures audit trails remain clear.

Catalog Permissions

Only datasets with minimum Data Read access appear in the workspace selection modal. Non-permissioned datasets are hidden entirely. Users cannot circumvent catalog governance through askEdgi.

Explainability and Transparency

askEdgi delivers explainability through a layered explanation model that caters to both business and technical users.

Each analysis may include a high-level summary describing the outcome in plain language, followed by detailed technical explanations that outline transformations, filters, joins, and assumptions applied during execution.

Where applicable, executable artifacts such as SQL queries or Python logic are available to support full reproducibility and independent review.

askEdgi provides layered explanations:

• Summary: Plain-language result description

• Technical Detail: Transformations, filters, joins, assumptions

• Reproducibility: Executable SQL, Python, or formula steps

View Generated SQL

Users can view the SQL queries generated by the system. This allows technical reviewers to verify logic, inspect joins, and confirm that filters are applied correctly.

Access method

Result blocks include a "View Code" option. Clicking it reveals the generated SQL for that step.

Edit Generated SQL

If the generated SQL is incorrect or incomplete, users can edit it before execution. This provides a safety mechanism for refining logic while maintaining transparency.

Access method

Result blocks include an "Edit Code" option. Users modify the SQL and re-run the query.

Insights and Summaries

askEdgi generates narrative summaries alongside tables and charts. These summaries explain key observations, trends, and anomalies detected in the data. The text is derived from computed results and metadata context, not from speculative reasoning.

Query Validation

Before executing a query, askEdgi validates:

• Dataset availability in the workspace

• Column name correctness

• Syntax and structure of generated SQL

• Compatibility with the target connector

If validation fails, the system prompts the user for clarification or correction.

Read-Only Operations

askEdgi performs read-only analysis. It does not write data back to source systems, modify catalog metadata, or trigger workflows that alter enterprise data. This prevents unintended changes and ensures all operations are safe for exploratory analysis.

Session Context Limits

Conversation context persists within a single session. When a user closes a thread or starts a new conversation, prior context is cleared. This prevents context leakage across unrelated analyses.

Audit Logging

All user actions are logged, including:

• Queries submitted

• Datasets accessed

• Recipes executed

• Workspace operations (uploads, removals)

• Compute and token consumption

Logs are retained according to enterprise audit policies and support governance reviews.

Limitations and Failure Modes

No AI system is perfect. Understanding where askEdgi may fail helps set appropriate expectations.

Ambiguous Questions

If a question is vague or lacks necessary context, askEdgi may select the wrong dataset, misinterpret intent, or generate incomplete results. Users should provide clear, specific prompts for best results.

Mitigation: Pin datasets to focus analysis. Rephrase questions with explicit column names and filters.

Incorrect SQL Generation

The system may generate syntactically correct SQL that does not match user intent. This occurs when questions are ambiguous or when dataset relationships are not well-defined in metadata.

Mitigation: Review the generated SQL using the "View Code" option. Edit and re-run if necessary.

Dataset Structure Mismatch

Recipes created on one dataset may fail when executed on a different dataset with different column names or structures. This is not a system error but a compatibility issue.

Mitigation: Document expected dataset structures in recipe descriptions. Test recipes on representative data before sharing.

Compute and Performance Limits

Large datasets, complex joins, or resource-intensive queries may exceed workspace capacity. This results in timeouts or incomplete results.

Mitigation: Upgrade workspace containers (Public edition) or contact administrators to adjust resource allocations (Enterprise editions).

Compliance and Regulatory Alignment

askEdgi is designed to operate within regulated environments.

GDPR Compliance

• Users retain full ownership and control of data

• OvalEdge operates as a data processor, not a data controller

• No customer data is used for AI model training

• Data subject rights (access, deletion) are supported via platform-level controls

CCPA Compliance

• askEdgi does not sell or share personal information

• Data is used only for analysis within the platform

• Users can request the deletion of uploaded files and workspace artifacts

ISO 27001 and SOC 2

askEdgi inherits compliance controls from AWS and OvalEdge platform architecture:

• AES-256 encryption at rest

• TLS 1.2+ encryption in transit

• Regular vulnerability assessments and penetration testing

• Continuous monitoring and audit logging

Industry-Specific Regulations

Organisations in healthcare, finance, or other regulated industries can deploy askEdgi in configurations that enforce additional controls:

• On-Prem deployments for full data residency

• Metadata-only editions to avoid data movement

• Custom masking and access policies via catalog governance

Responsible AI Principles

askEdgi follows responsible AI practices to ensure ethical and accountable use.

Transparency

Users can inspect generated SQL, review recipe steps, and access audit logs. System behavior is documented in detail, and limitations are disclosed openly.

Accountability

askEdgi operates within OvalEdge's governance framework. Administrators control access, configure connectors, and enforce policies. Users are responsible for validating results and using insights appropriately.

Human Oversight

askEdgi assists decision-making but does not replace human judgment. Users review results, confirm accuracy, and apply business context before taking action.

No Autonomous Actions

askEdgi does not execute workflows, send alerts, or trigger downstream processes without explicit user instruction. All operations require human initiation.

Fairness and Bias

askEdgi processes data according to user-defined logic and catalog governance. It does not introduce bias through model training on customer data. However, biases present in source data or business rules are reflected in results. Organizations are responsible for reviewing data quality and governance policies.

Risk Mitigation Summary

Risk	Mitigation
Unauthorized data access	Role-based access controls are validated before every query
Data leakage to external AI	Only metadata and summaries sent to external models
Incorrect results	Generated SQL is reviewable and editable by users
Unintended data modification	All operations are read-only; no writes to source systems
Session data exposure	Workspace isolation ensures users cannot access others' data
Lack of auditability	Comprehensive logging of all queries, executions, and access
Non-compliance with regulations	Edition-specific controls support GDPR, CCPA, ISO, SOC 2
Model hallucination	RAG architecture grounds responses in enterprise metadata
Compute overruns	Spend limits and budget alerts enforce consumption controls

Governance Recommendations

Organizations deploying askEdgi should implement the following practices:

Define Access Policies:

Configure role-based access and data masking rules in the OvalEdge catalog before enabling askEdgi. This ensures users cannot bypass governance through the AI interface.

Establish Usage Policies:

Document acceptable use cases, prohibited actions, and escalation procedures. Communicate these policies to users during onboarding.

Monitor Usage:

Review audit logs regularly. Identify anomalous activity, excessive consumption, or policy violations. Use logs for compliance reporting and governance reviews.

Validate Outputs:

Encourage users to review generated SQL and cross-check results against established KPIs. Implement quality checks for high-impact analyses.

Control Marketplace Access:

In SaaS editions, decide whether to enable Public Marketplace recipe consumption. Policy-based controls allow administrators to restrict external recipe usage.

Test Before Production:

Conduct proof-of-concept deployments to validate behavior, performance, and governance enforcement. Use test data to simulate production scenarios.

Limit On-Prem Scope:

For On-Prem deployments, restrict askEdgi to metadata analytics only. Avoid enabling features that require external AI enrichments or cloud connectivity.

Summary

askEdgi operates as a governed AI assistant within the OvalEdge platform. It enforces access controls, respects data masking, and provides transparency through reviewable SQL and audit logs. Enterprise data remains secure. Only metadata is shared with external AI providers.

Organizations retain full control over data, policies, and usage. Deployment flexibility ensures askEdgi can operate within strict compliance and residency requirements.

Responsible use requires clear policies, regular monitoring, and user training. When deployed correctly, askEdgi accelerates insight generation while maintaining governance and trust.