AI Impact Assessment (AIIA)
AI Impact Assessment template for AgentTrust Edge Gateway.
AI Impact Assessment (AIIA)
Standard: ISO/IEC 42001:2023 Clause 6.2 and Annex A.5.2 Document owner: AI Governance Lead Review cadence: Before each major release; annually; after any incident involving adverse impact Classification: Internal — Restricted
1. AI System Identification
| Field | Value |
|---|---|
| System name | AgentTrust Edge Gateway |
| Version | 1.x |
| Assessment date | [DATE] |
| Assessor | [NAME / ROLE] |
| Approved by | [NAME / ROLE] |
2. Purpose and Intended Use
AgentTrust Edge Gateway is a runtime governance platform for autonomous AI agent executions. It evaluates agent outputs against deterministic checks, policy rules, confidence scoring, and optional LLM judge scoring before returning an approve/block/escalate decision to the calling system.
Primary intended uses:
- Software development tooling governance (code generation, code review agents)
- Customer service agent quality assurance
- Financial analysis agent oversight
- Research and document summarisation governance
- Multi-agent orchestration trust chain enforcement
Supported agent frameworks: LangGraph, CrewAI, OpenAI Agents, Claude Agents, MCP, Custom, REST
Deployment environments: On-premise (Docker Compose), Kubernetes edge cluster, cloud-hosted SaaS
3. Stakeholder Analysis
| Stakeholder | Role | Potential Impact |
|---|---|---|
| AI agent operators | Deploy and configure gateway | Operational dependency; incorrect config → wrong decisions |
| End users | Receive outputs from governed agents | May receive blocked or delayed responses |
| Organisations deploying agents | Commercial users | Liability if governance fails; reputational risk |
| Individuals whose data is processed | Data subjects | Privacy; PII processed in audit ledger |
| Regulators | Compliance oversight | Non-compliance with AI Act, GDPR, sector regulations |
| Third-party AI providers (Anthropic, Ollama) | LLM judge backends | Data sharing; model behaviour changes |
4. Risk Categories — AI-Specific
4.1 Safety and Reliability
| Risk | Likelihood | Severity | Mitigation |
|---|---|---|---|
| False positive blocks (legitimate outputs blocked) | Medium | High | Human review queue; override mechanism; feedback API |
| False negative approvals (harmful outputs approved) | Medium | Critical | 6-check validation; policy packs; circuit breaker |
| Fast-path latency SLA breach (>50ms P99) | Low | Medium | Latency monitoring; alert rules |
| LLM judge returning hallucinated scores | Low | High | Pydantic schema validation; score range clamping; sampling only |
| Audit ledger corruption | Very Low | Critical | SHA-256 hash chain; CAS locking; backup policy |
4.2 Fairness and Non-Discrimination
| Risk | Likelihood | Severity | Mitigation |
|---|---|---|---|
| Systematic bias in tool severity ratings | Medium | High | Regular bias audit; configurable severity tables |
| Disproportionate blocks for certain agent types | Low | High | Approval rate monitoring by agent_id and framework |
| Policy rules encoded with cultural bias | Low | Medium | Policy review board; diverse stakeholder input |
| Historical reliability penalising new agents | Low | Medium | Minimum window threshold; cold-start exception |
4.3 Privacy and Data Governance
| Risk | Likelihood | Severity | Mitigation |
|---|---|---|---|
| PII in audit ledger (request_json, output_json) | High | High | PII detection policy; masking before storage; retention limits |
| Audit data shared with third-party LLM | Medium | High | Opt-out for judge; data minimisation in prompts |
| Audit records retained indefinitely | High | Medium | 90-day hot retention; archive + anonymise older records |
| Data residency violations (Claude API cross-border) | Medium | High | DPA with Anthropic; region configuration |
4.4 Transparency and Explainability
| Risk | Likelihood | Severity | Mitigation |
|---|---|---|---|
| Operators unable to explain a block decision | Medium | High | governance_disclosure field; confidence_rationale in response |
| End users unaware AI governance evaluated their agent | High | Medium | API disclosure field; SDK documentation |
| Confidence score not interpretable | Low | Medium | 7-signal breakdown in response; rationale string |
4.5 Security and Adversarial Inputs
| Risk | Likelihood | Severity | Mitigation |
|---|---|---|---|
| Prompt injection bypassing policy checks | Medium | Critical | Adversarial detection in ValidationEngine; policy cap |
| JWT token forging for tier elevation | Low | Critical | Server-side signature verification; client-side FREE default |
| Rate limit bypass via distributed IPs | Low | Medium | Redis sliding window; IP-based auth rate limiting |
| Audit ledger replay attack | Very Low | High | Hash chain integrity; CAS locking |
5. Adverse Impact Assessment
5.1 High-Risk Use Cases
The following use cases require additional review before deployment:
- Medical diagnosis or triage agents — a false negative could result in patient harm
- Financial trading or payment authorisation agents — a false positive or negative has monetary consequence
- Legal document generation agents — incorrect approval of legally invalid output
- Child safety or safeguarding contexts — extremely low tolerance for false negatives
For these use cases, operators must:
- Complete this AIIA with domain-specific risk assessment
- Configure policy packs with domain-specific rules
- Set block threshold for risk tier ≥
high - Ensure all
escalateandblockdecisions route to qualified human reviewers
5.2 Prohibited Use Cases
AgentTrust Edge Gateway must NOT be used as:
- The sole automated decision-maker for high-stakes irreversible actions (e.g., medical procedures, criminal justice)
- A tool for surveillance or profiling without explicit legal basis and consent
- A system to make employment decisions without human oversight
6. Mitigation Measures — Residual Risk Acceptance
| Risk ID | Residual Risk Level | Accepted By | Date |
|---|---|---|---|
| Safety-01: false positive blocks | Medium → Low | [NAME] | [DATE] |
| Fairness-01: systematic bias | Medium → Low | [NAME] | [DATE] |
| Privacy-01: PII in ledger | High → Medium | [NAME] | [DATE] |
| Security-01: prompt injection | Medium → Low | [NAME] | [DATE] |
7. Review History
| Version | Date | Author | Changes |
|---|---|---|---|
| 1.0 | [DATE] | [NAME] | Initial assessment |