PranavNagrecha/AwesomeSalesforceSkills / skills/admin/ai-ethics-and-governance-requirements
ai-ethics-and-governance-requirements
VerifiedUse this skill when defining AI governance policies, designing human oversight workflows, documenting bias mitigation strategies, or meeting regulatory transparency requirements for Salesforce AI features. Trigger keywords: responsible AI, AI bias, AI audit trail, AI transparency, human-in-the-loop, AI risk inventory, AI disclosure. NOT for the MLOps pipeline or Policy-as-Code engine architecture — use architect/ai-governance-architecture. NOT for Trust Layer technical configuration — use agentforce/einstein-trust-layer.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 14.8 KB |
| references/examples.md | markdown | 5.5 KB |
| references/gotchas.md | markdown | 6.5 KB |
| references/llm-anti-patterns.md | markdown | 9.4 KB |
| references/well-architected.md | markdown | 5.4 KB |
| scripts/check_ai_ethics_and_governance_requirements.pyexecutable | script | 9.4 KB |
| templates/ai-ethics-and-governance-requirements-template.md | markdown | 5.2 KB |
references/gotchas.md
View on GitHub# Gotchas — AI Ethics and Governance Requirements
Non-obvious Salesforce platform behaviors that cause real production problems in this domain.
## Gotcha 1: Einstein Trust Layer Zero-Retention Does Not Apply to Salesforce-Native Model Calls
**What happens:** Practitioners assume that the Einstein Trust Layer's zero data retention guarantee applies to all AI calls in their org. In fact, zero-retention is enforced only for calls made to third-party LLMs (OpenAI, Anthropic, etc.) routed through the Trust Layer gateway. Calls to Salesforce-native Einstein models — Einstein Prediction Builder, Einstein Discovery, Einstein Next Best Action — are processed within Salesforce's own infrastructure under standard Salesforce data retention and processing terms, not the third-party zero-retention contract.
**When it occurs:** Any time an org audits its AI data flows and assumes a single policy covers all AI features. Also triggered when a DPA (Data Processing Agreement) review asks "where is AI inference data retained?" and the answer incorrectly covers only Trust Layer.
**How to avoid:** In the governance documentation, explicitly separate two categories: (1) native Salesforce Einstein model calls covered by Salesforce's standard data processing terms, and (2) third-party LLM calls covered by Trust Layer zero-retention. Obtain the correct Salesforce DPA addendum for each category and reference both in the governance policy.
---
## Gotcha 2: AI-Generated Content Disclosure Is Not Automatically Applied to Custom LWC Surfaces
**What happens:** Salesforce's standard AI-generated content disclosure marker (the sparkle icon and "Generated by AI" label) appears automatically on Salesforce-built UI components such as Einstein Email Composer and Agentforce chat UI. When a developer builds a custom Lightning Web Component that surfaces AI-generated text — such as a custom product recommendation card or a knowledge article summary rendered through an LWC — the disclosure marker is not automatically injected. The component ships with no AI disclosure unless the developer explicitly adds it.
**When it occurs:** Any org that builds custom UI on top of Einstein or Agentforce APIs. Very common in ISV and partner-built solutions where the Einstein API is called but the display layer is fully custom.
**How to avoid:** The governance checklist must include a specific line item: "For every custom LWC or Aura component that surfaces AI-generated text, confirm disclosure language is present in the component markup." The disclosure copy must be approved by legal and must not be removable by org configuration. Include a review of all custom components in the go-live QA checklist.
---
## Gotcha 3: Field History Tracking Does Not Capture the AI Model Version or Contributing Features
**What happens:** When an Einstein model writes a value to a Salesforce field (such as a Lead Score field updated by Einstein Prediction Builder), Field History Tracking records the before and after values and the record's last-modified user — but not the AI model version, the model's confidence score, or the contributing predictor fields that influenced the prediction. If the same field is later updated by a human or a Flow, the audit trail shows only a sequence of value changes with no indication of which changes were AI-generated.
**When it occurs:** Any org that relies on Field History Tracking as its sole AI audit log. This gap is especially dangerous for high-risk use cases where the regulator needs to know not just that a score changed, but which model version produced it and on what basis.
**How to avoid:** Supplement Field History Tracking with a custom platform event or a dedicated AI Decision Log custom object. Each time Einstein writes to the audited field, a Flow or Apex trigger writes a companion log record capturing: record ID, AI feature name, model version (if exposed), prediction score, top contributing features, and timestamp. Export this log to Data Cloud for long-term retention. Reference this logging design in the audit trail specification section of the governance policy.
---
## Gotcha 4: Bias Report Results Are a Snapshot, Not a Continuous Monitor
**What happens:** The Einstein Prediction Builder bias detection report runs at a point in time and produces a static result. When the model is retrained on updated data — whether by a scheduled retraining cadence or by a manual retrain triggered by a model performance alert — the bias baseline resets. The previous report findings no longer apply. Orgs that run the bias report once at initial deployment and never re-run it may be deploying retrained models with newly introduced bias, undetected.
**When it occurs:** Any org that retrains Einstein models on a regular cadence without a governance checkpoint between retraining and production promotion. Retraining is often triggered automatically when model accuracy drops — the bias evaluation is rarely included in that automated flow.
**How to avoid:** Add a mandatory governance gate to the model promotion process. The gate requires: (1) a new bias report run on the retrained model, (2) results reviewed against the accepted threshold, and (3) sign-off by the data governance owner before the model is activated. Document this gate in both the governance policy and any CI/CD or DevOps pipeline that manages model promotion.
---
## Gotcha 5: EU AI Act High-Risk Classification Can Apply to Salesforce-Powered Decisions Even for Non-EU Orgs
**What happens:** Org admins assume that EU AI Act obligations apply only to EU-headquartered companies. The EU AI Act applies to any system that produces outputs affecting EU residents, regardless of where the org is based. A US-based financial services firm using Einstein to automate credit or insurance decisions for European customers is subject to EU AI Act Article 10 (data governance), Article 14 (human oversight), and Article 26 (obligations of deployers) — irrespective of the firm's country of incorporation.
**When it occurs:** During a governance scoping exercise when the legal/compliance team asks "does EU AI Act apply to us?" and the org's Salesforce AI features are not included in the analysis because the org is US-based.
**How to avoid:** Include a jurisdiction mapping in the risk inventory. For each AI use case, document the geographic profile of affected users. Flag any use case affecting EU residents for EU AI Act scoping. Escalate the flagged items to legal counsel for a conformity assessment determination. Do not make the final classification call without legal input.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| rea-analysis-optimizationmorluto/reaOptimize REA providers and shared analysis workflows for agent context, latency, CPU, and memory while preserving correctness and Evidence. Use for analysis bottlenecks and cache or query design; CI/test-suite optimization belongs to rea-ci-tests-audit. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| rea-ci-tests-auditmorluto/reaAudit REA CI and test suites for correctness, fixture drift, false passes, and measured performance; use for suite or workflow investigations and efficiency changes, rather than ordinary feature testing or a single test failure. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| rea-tool-designmorluto/reaAudit, investigate, design, or change REA investigation tools, CLI/MCP contracts, provider capabilities, and Evidence semantics. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-auditDietrichGebert/ponytailQuality audit of the whole repo: bugs, security, real load, missing tests, speed, and what to delete. Most important first. | Stars: 196k | Updated | Claude CodeCodex | Verified | Install |