> ## Documentation Index
> Fetch the complete documentation index at: https://docs.dqlabs.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# AI Stewardship

> The human + AI governance layer that ensures autonomous actions stay transparent, accountable, and under control

<script type="application/ld+json">
  {`{
            "@context": "https://schema.org",
            "@type": "TechArticle",
            "headline": "AI Stewardship",
            "description": "The human + AI governance layer that ensures autonomous actions stay transparent, accountable, and under control",
            "url": "https://docs.dqlabs.ai/platform/ai-stewardship",
            "publisher": {
              "@type": "Organization",
              "name": "DQLabs Inc",
              "logo": "https://media.brand.dev/332adc35-5bc4-4d2b-bf78-256aa4a5e414.svg"
            }
            }`}
</script>

## Overview

AI Stewardship is Prizm's governance layer for every action an agent takes. It ensures that as Prizm's agents act with increasing autonomy, every decision remains **transparent, risk-scored, and — where warranted — subject to human approval**. This is what makes Prizm's agentic execution trustworthy rather than a black box: humans stay in control even as agents do the work.

<Frame>
  <img src="https://mintlify.s3.us-west-1.amazonaws.com/dqlabs/images/prizm-ai-stewardship.png" alt="AI Stewardship governance workflow" />
</Frame>

## Foundational concepts

Understanding Stewardship starts with four building blocks:

| Concept      | Definition                                                                                                                                                                                                                     |
| ------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| **Metadata** | Raw information that describes an asset — titles, schemas, columns, tags. Metadata is a book's table of contents, chapters, and index.                                                                                         |
| **Context**  | Metadata + surrounding information + retrieval + relevance that the AI uses to make decisions. Context is interpreted information for intelligent decision-making — a book's summary plus a teacher's notes on why it matters. |
| **Action**   | A **change** to metadata, quality rules, metrics, schemas, lineage, or policy. Extracting metadata is not an action — but a change detected or applied *becomes* one. Actions can be generated by a human or by AI.            |
| **Log**      | Every action is logged. Application, API, and AI-driven events appear in the Audit Log; user sessions and activity appear in the Activity Log.                                                                                 |

## Action sensitivity and risk scoring

Every action carries a default **sensitivity** level, which combines with the asset's criticality to produce a quantitative risk score.

### Sensitivity weights

| Action Sensitivity | Weight | Meaning                                                                                      |
| ------------------ | ------ | -------------------------------------------------------------------------------------------- |
| **Low**            | 0.3    | Metadata-only, reversible, safe (e.g. updating a description)                                |
| **Medium**         | 0.6    | Context-dependent, moderate impact (e.g. creating a non-financial metric)                    |
| **High**           | 1.0    | Business-critical, schema/metric/PII-impacting (e.g. PII classification, KPI formula update) |

### Risk score formula

$$
\text{Action Risk} = \text{Action Sensitivity (Weight)} \times \text{Asset Criticality (Percentile)}
$$

This yields a single 0–100 score. The weighting is deliberate: **0.3** keeps low-risk actions low even on critical assets; **0.6** lets medium-risk actions scale meaningfully with asset importance; **1.0** means high-sensitivity actions fully inherit the criticality of the asset they touch.

<Accordion title="Worked examples">
  |                   | Example A — Low Sensitivity | Example B — Medium Sensitivity | Example C — High Sensitivity           |
  | ----------------- | --------------------------- | ------------------------------ | -------------------------------------- |
  | Action            | Update description          | Create a non-financial metric  | PII classification, KPI formula update |
  | Sensitivity       | Low → 0.3                   | Medium → 0.6                   | High → 1.0                             |
  | Asset criticality | 80                          | 80                             | 80                                     |
  | **Risk**          | 0.3 × 80 = **24 → Low**     | 0.6 × 80 = **48 → Medium**     | 1.0 × 80 = **80 → High**               |
</Accordion>

### Risk level bands

| Risk Score    | Risk Level    |
| ------------- | ------------- |
| 0 – below 40  | Low Risk      |
| 40 – below 75 | Medium Risk   |
| 75 – below 90 | High Risk     |
| 90 – 100      | Critical Risk |

## Autonomy levels

Prizm maps its AI autonomy to a familiar self-driving car analogy — each level represents how much the AI decides and executes versus how much a human must initiate or approve.

| AI Autonomy Level             | Self-Driving Equivalent | Behavior                                                                                                  |
| ----------------------------- | ----------------------- | --------------------------------------------------------------------------------------------------------- |
| **Human**                     | Manual Driving          | User initiates, decides, and performs all actions. AI only provides visibility (like a lane-assist beep). |
| **AI Guidance**               | Driver-Assisted Mode    | Human starts the action; AI offers recommendations, human still decides.                                  |
| **AI Assisted**               | Co-Pilot Mode           | AI actively detects and suggests actions; human approves, like a co-pilot with hands on the wheel.        |
| **AI Completed (Autonomous)** | Fully Self-Driving      | AI detects, decides, and executes automatically. Human only monitors high-risk events.                    |

## Action lifecycle

Every action progresses through a defined lifecycle, whether it's fully autonomous or awaiting human sign-off:

```text theme={null}
Detected → Proposed → Pending Approval → Executing → Completed / Partially Completed / Failed / Overridden → Reverted (if needed) → Logged / Archived
```

| Stage                | When it happens                           | Example                                    |
| -------------------- | ----------------------------------------- | ------------------------------------------ |
| **Pending**          | AI detects an opportunity for improvement | "Description unclear → improvement needed" |
| **Ready for Review** | AI proposes a specific change             | "AI draft ready for steward review"        |
| **Verified**         | Steward approves the update               | "Steward validated updated description"    |
| **Completed**        | AI executes the update                    | "Description updated successfully"         |
| **Deprecated**       | Steward rejects the AI suggestion         | "Suggested description not applied"        |

<Note>
  **Overridden** means a human rejected an AI-Assisted suggestion. **Reverted** means an action was rolled back after a negative trust impact was detected. Both states are preserved for audit and learning.
</Note>

## The Stewardship hub

The **Stewardship** page is the central hub for monitoring every action in the system — autonomous, AI-assisted, or human — giving stewards a single place to review what the platform has done, is doing, or is waiting on.

### Action record schema

Every action is captured with a structured record, including:

| Field                                | Description                                                                                         |
| ------------------------------------ | --------------------------------------------------------------------------------------------------- |
| `action_id`                          | Unique identifier                                                                                   |
| `action_type`                        | e.g. `schema_correction`, `quality_rule_validation`, `business_term_mapping`, `remediation_trigger` |
| `autonomy_level`                     | `Autonomous`, `AI-Assisted`, `Human`, `System Triggered`                                            |
| `domain` / `product` / `application` | Business context the action belongs to                                                              |
| `severity_level`                     | `Low`, `Medium`, `High`, `Critical`                                                                 |
| `confidence_score`                   | AI's confidence in the decision (0–1)                                                               |
| `initiated_by` / `approved_by`       | Who or what triggered and validated the action                                                      |
| `trigger_source`                     | `Anomaly Detection`, `Quality Monitor`, `Lineage Drift`, `Manual Input`                             |
| `result`                             | `Success`, `Failed`, `Partially Successful`, `Skipped`                                              |
| `trust_impact_score`                 | Net effect on the platform's trust index                                                            |
| `learning_flag` / `audit_flag`       | Whether the action feeds retraining and/or compliance workflows                                     |

### Example actions by category

<AccordionGroup>
  <Accordion title="🧠 Autonomous Actions — AI-driven, no human input">
    * Schema drift auto-corrected
    * Missing value patch generated
    * Data product ownership updated
  </Accordion>

  <Accordion title="🤖 AI-Assisted Actions — AI suggests, human approves">
    * Suggested quality rule confirmed
    * Business term mapping validated
    * Threshold adjusted via co-pilot
  </Accordion>

  <Accordion title="👤 Human Actions — manual, tracked for learning">
    * Steward manually resolved a quality issue
    * Metadata enrichment via human input
    * Rule exception documented
  </Accordion>

  <Accordion title="🔁 In Review / Next Actions — queued or pending">
    * Pending approval for a schema fix
    * Awaiting steward confirmation
    * Under confidence-threshold recheck
  </Accordion>
</AccordionGroup>

### Stewardship UI

The Stewardship page is organized into two sections:

1. **Key Metrics** — scorecards for Autonomous, AI-Assisted, Steward, and Pending Review counts
2. **Table Component** — every action listed with `action_id`, `action_type`, `autonomy_level`, `description`, `domain`, `product`, `severity_level`, `status`, `start_time`, `end_time`, and `duration`

<Accordion title="Example weekly metrics snapshot">
  * **Total Actions This Week**: 1,248
  * **Autonomous**: 742 · **AI-Assisted**: 336 · **Human**: 132 · **In Review**: 38
  * **Avg Confidence Score (Autonomous)**: 92%
  * **Avg Time to Resolution**: 3.2 hours (↓48% YoY)
  * **Trust Index Improvement**: +27%
  * **Steward Adoption Rate**: 81%
  * **Actions per Domain**: Finance (310), Marketing (228), Ops (186), HR (92)
  * **Insight**: 34% of manual tasks converted to AI-assisted over Q3; 12 recurring data patterns now resolved autonomically
</Accordion>

## Stewardship across the platform

AI Stewardship isn't confined to one module — it's applied consistently across every part of Prizm:

| Area              | What Stewardship does                                                                                             |
| ----------------- | ----------------------------------------------------------------------------------------------------------------- |
| **Catalog**       | Metadata enrichment, classification governance for sensitive data, drift detection, knowledge graph evolution     |
| **Quality**       | Adaptive rule generation, explainable quality scores, bias detection, improvement recommendations                 |
| **Observability** | Contextual anomaly explanations, predictive health monitoring, impact analysis, self-healing with human oversight |

## Related pages

<CardGroup cols={2}>
  <Card title="Autonomous Intelligence" icon="brain" href="/platform/autonomous-intelligence">
    See how Stewardship fits into Prizm's broader autonomy architecture
  </Card>

  <Card title="Multi-Agent Architecture" icon="robot" href="/platform/multi-agent-architecture">
    See the agents whose actions Stewardship governs
  </Card>
</CardGroup>