Responsible AI Trust Framework¶
AWS Bedrock Guardrails for FTR Compliance¶
Document Type: Partner First Call Deck Reference FTR Category: AI Agents and Tools Last Updated: December 2025
Executive Summary¶
Vellocity implements a three-tier Responsible AI Trust Framework using AWS Bedrock Guardrails to ensure all AI-generated content meets AWS Marketplace standards for accuracy, safety, and compliance.
Why This Matters for Partners¶
| Partner Concern | How Guardrails Address It |
|---|---|
| "Will AI hallucinate about my offering?" | Contextual grounding blocks unverified claims |
| "Will sensitive data leak?" | PII detection anonymizes or blocks sensitive information |
| "Will AI give bad advice?" | Topic policies deny legal, financial, and pricing guidance |
| "Is this FTR compliant?" | 100% AWS-native, deployed via CloudFormation |
Trust Tier Architecture¶
┌─────────────────────────────────────────────────────────────────────────────┐
│ VELLOCITY TRUST FRAMEWORK │
├─────────────────────────────────────────────────────────────────────────────┤
│ │
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────────────┐ │
│ │ DEV │ │ PROD │ │ ENTERPRISE │ │
│ │ TIER │ │ TIER │ │ TIER │ │
│ ├─────────────┤ ├─────────────┤ ├─────────────────────┤ │
│ │ Grounding: │ │ Grounding: │ │ Grounding: │ │
│ │ 0.40 │ │ 0.25 │ │ 0.15 │ │
│ │ (Flexible) │ │ (Balanced) │ │ (Strictest) │ │
│ ├─────────────┤ ├─────────────┤ ├─────────────────────┤ │
│ │ PII: │ │ PII: │ │ PII: │ │
│ │ Anonymize │ │ Anonymize │ │ BLOCK │ │
│ │ Email/Phone │ │ + AWS IDs │ │ All identifiers │ │
│ ├─────────────┤ ├─────────────┤ ├─────────────────────┤ │
│ │ Use Case: │ │ Use Case: │ │ Use Case: │ │
│ │ Internal │ │ Customer- │ │ Regulated/ │ │
│ │ Testing │ │ Facing │ │ Sensitive │ │
│ └─────────────┘ └─────────────┘ └─────────────────────┘ │
│ │ │ │ │
│ └──────────────────┼─────────────────────┘ │
│ │ │
│ ▼ │
│ ┌─────────────────────────────┐ │
│ │ AWS BEDROCK RUNTIME │ │
│ │ (Guardrail Enforcement) │ │
│ └─────────────────────────────┘ │
│ │
└─────────────────────────────────────────────────────────────────────────────┘
Guardrail Priority Resolution¶
When an AI agent executes, guardrails are selected using this priority order:
┌─────────────────────────────────────────────────────────────────┐
│ GUARDRAIL RESOLUTION │
├─────────────────────────────────────────────────────────────────┤
│ │
│ 1. AGENT-LEVEL Has agent-specific ──► Use agent's │
│ GUARDRAIL guardrail configured? guardrail │
│ │ │
│ │ No │
│ ▼ │
│ 2. CAPABILITY-LEVEL Is capability mapped ──► Use mapped │
│ GUARDRAIL to a guardrail? guardrail │
│ │ │
│ │ No │
│ ▼ │
│ 3. PLATFORM DEFAULT Is platform default ──► Use default│
│ GUARDRAIL configured? guardrail │
│ │ │
│ │ No │
│ ▼ │
│ 4. NO GUARDRAIL Execute without ──► No safety │
│ guardrail filtering │
│ │
└─────────────────────────────────────────────────────────────────┘
Trust Tier Comparison¶
Content Safety Filters¶
| Filter Type | Dev Tier | Prod Tier | Enterprise Tier |
|---|---|---|---|
| Hate Speech | Medium/High | High/High | High/High |
| Insults | Medium/High | High/High | High/High |
| Sexual Content | High/High | High/High | High/High |
| Violence | Medium/Medium | Medium/Medium | High/High |
| Misconduct | Medium/Medium | Medium/Medium | High/High |
| Prompt Attacks | None | None | None |
Contextual Grounding (Hallucination Prevention)¶
| Metric | Dev Tier | Prod Tier | Enterprise Tier |
|---|---|---|---|
| Grounding Threshold | 0.40 | 0.25 | 0.15 |
| Relevance Threshold | 0.40 | 0.30 | 0.15 |
| Strictness Level | Permissive | Balanced | Maximum |
Lower threshold = stricter enforcement. Content below threshold is blocked.
Sensitive Information Protection¶
| Data Type | Dev Tier | Prod Tier | Enterprise Tier |
|---|---|---|---|
| Email Addresses | Anonymize | Anonymize | BLOCK |
| Phone Numbers | Anonymize | Anonymize | BLOCK |
| Physical Addresses | — | Anonymize | BLOCK |
| AWS Account IDs (12-digit) | — | Anonymize | BLOCK |
| Deal/Opportunity IDs | — | — | BLOCK |
Denied Topics (Hard Blocks)¶
| Topic | Dev | Prod | Enterprise |
|---|---|---|---|
| Legal Advice | ✓ | ✓ | ✓ |
| Financial Predictions | — | ✓ | ✓ |
| Off-Marketplace Pricing | — | ✓ | ✓ |
| Regulatory Interpretation | — | — | ✓ |
| Pricing Recommendations | — | — | ✓ |
| Co-Sell Commitments | — | — | ✓ |
Blocked Phrases (All Tiers Block These)¶
- "approved by AWS"
- "endorsed by AWS"
- "AWS guarantees"
- "unlimited" (Enterprise only)
- "risk-free" (Enterprise only)
Capability-to-Guardrail Mapping¶
Different AI capabilities require different trust levels:
| Capability | Recommended Tier | Rationale |
|---|---|---|
| Marketplace Awareness | Enterprise | Prevents hallucinated AWS Marketplace claims |
| Deal Influence Tracking | Enterprise | Blocks sensitive deal IDs and PII |
| Content Generation | Prod | Balanced safety and creativity |
| Social Media Publishing | Dev | Lighter filtering for creative content |
| Internal Analysis | Dev | Maximum flexibility for exploration |
FTR Compliance Mapping¶
How Guardrails Support FTR Requirements¶
| FTR Requirement | Guardrail Feature | Implementation |
|---|---|---|
| AWS-Native Architecture | Bedrock Guardrails API | CloudFormation-deployed |
| Data Protection | PII Detection & Blocking | Email, phone, address, AWS IDs |
| Content Safety | Content Policy Filters | Hate, violence, sexual, misconduct |
| Accuracy Standards | Contextual Grounding | Threshold-based hallucination prevention |
| Compliance Boundaries | Topic Policies | Legal, financial, regulatory denials |
| Audit Trail | Trace Logging | Full intervention logging to CloudWatch |
FTR Category: AI Agents and Tools¶
Vellocity's guardrail implementation addresses these AI-specific FTR considerations:
┌────────────────────────────────────────────────────────────────────────────┐
│ FTR AI AGENTS & TOOLS COMPLIANCE │
├────────────────────────────────────────────────────────────────────────────┤
│ │
│ ┌──────────────────┐ ┌──────────────────┐ ┌──────────────────┐ │
│ │ RESPONSIBLE AI │ │ DATA HANDLING │ │ AUDIT & LOGGING │ │
│ ├──────────────────┤ ├──────────────────┤ ├──────────────────┤ │
│ │ • Grounding │ │ • PII anonymize │ │ • Intervention │ │
│ │ prevents │ │ or block │ │ logs │ │
│ │ hallucinations │ │ • No external │ │ • CloudWatch │ │
│ │ • Content │ │ data egress │ │ integration │ │
│ │ filters block │ │ • AWS-native │ │ • Trace mode │ │
│ │ harmful output │ │ storage only │ │ for debugging │ │
│ │ • Topic denials │ │ │ │ │ │
│ │ prevent bad │ │ │ │ │ │
│ │ advice │ │ │ │ │ │
│ └──────────────────┘ └──────────────────┘ └──────────────────┘ │
│ │
└────────────────────────────────────────────────────────────────────────────┘
Deployment Architecture¶
AWS Services Used¶
| Service | Purpose | FTR Status |
|---|---|---|
| Amazon Bedrock | Foundation model inference | ✅ Native |
| Bedrock Guardrails | Content safety enforcement | ✅ Native |
| CloudFormation | Infrastructure as code | ✅ Native |
| CloudWatch | Logging and monitoring | ✅ Native |
| IAM | Access control | ✅ Native |
CloudFormation Stack Outputs¶
Outputs:
DevGuardrailId: vellocity-marketplace-trust-dev
ProdGuardrailId: vellocity-marketplace-trust-prod
EnterpriseGuardrailId: vellocity-marketplace-trust-enterprise
Partner Integration Options¶
Option 1: Platform-Managed (Recommended for Most Partners)¶
- Vellocity manages guardrails in platform AWS account
- Partners select tier per agent or use platform defaults
- Zero AWS configuration required from partner
Option 2: BYOC (Bring Your Own Credentials)¶
- Partners deploy guardrails in their own AWS account
- Partners provide IAM role for cross-account access
- Full control over guardrail configuration
- Syncs to Vellocity for agent assignment
Option 3: Hybrid¶
- Use platform guardrails for standard capabilities
- Use partner guardrails for sensitive/custom workflows
Quick Reference Card¶
For Partner First Calls¶
What are Bedrock Guardrails?
AWS-native content safety layer that filters AI inputs/outputs in real-time
Why three tiers?
Different use cases need different safety/flexibility tradeoffs
How does this help FTR?
100% AWS-native, auditable, configurable per-agent or per-capability
What gets blocked?
Hallucinations, PII, legal/financial advice, false AWS claims
What's the overhead?
~$0.75 per 1,000 content units (text) — minimal latency impact
Tier Selection Guide¶
| If your partner needs... | Recommend |
|---|---|
| Maximum flexibility for internal tools | Dev |
| Customer-facing content generation | Prod |
| Regulated industry or sensitive data | Enterprise |
| Unsure / First deployment | Prod (safe default) |
Technical Documentation References¶
For implementation details, see:
/docs/BEDROCK_GUARDRAILS.md— Full implementation guide/docs/CAPABILITY_GUARDRAILS_INTEGRATION.md— Capability mapping guide/cloudformation/application/bedrock-guardrails.yml— CloudFormation template/cloudformation/application/deploy-guardrails.sh— Deployment script
Contact & Support¶
For FTR Questions: Review AWS FTR requirements at AWS Partner Central
For Guardrail Configuration: Admin → Bedrock Guardrails → Sync from AWS
For Custom Tier Requirements: Contact Vellocity platform team
Vellocity Responsible AI Trust Framework — Designed for AWS Marketplace Partners