agentforce-pii-redaction
Redact PII before it reaches Agentforce prompts, models, and logs. Trigger keywords: agentforce pii, pii redaction, data masking llm, einstein trust layer, prompt pii filter, audit pii leakage. Does NOT cover: Shield Platform Encryption at-rest (separate skill), GDPR data subject requests, or classic field-level security policy.
Best use case
agentforce-pii-redaction is best used when you need a repeatable AI agent workflow instead of a one-off prompt.
Redact PII before it reaches Agentforce prompts, models, and logs. Trigger keywords: agentforce pii, pii redaction, data masking llm, einstein trust layer, prompt pii filter, audit pii leakage. Does NOT cover: Shield Platform Encryption at-rest (separate skill), GDPR data subject requests, or classic field-level security policy.
Teams using agentforce-pii-redaction should expect a more consistent output, faster repeated execution, less prompt rewriting, better workflow continuity with your supporting tools.
When to use this skill
- You want a reusable workflow that can be run more than once with consistent structure.
- You already have the supporting tools or dependencies needed by this skill.
When not to use this skill
- You only need a quick one-off answer and do not need a reusable workflow.
- You cannot install or maintain the underlying files, dependencies, or repository context.
Installation
Claude Code / Cursor / Codex
Manual Installation
- Download SKILL.md from GitHub
- Place it in
.claude/skills/agentforce-pii-redaction/SKILL.mdinside your project - Restart your AI agent — it will auto-discover the skill
How agentforce-pii-redaction Compares
| Feature / Agent | agentforce-pii-redaction | Standard Approach |
|---|---|---|
| Platform Support | Not specified | Limited / Varies |
| Context Awareness | High | Baseline |
| Installation Complexity | Unknown | N/A |
Frequently Asked Questions
What does this skill do?
Redact PII before it reaches Agentforce prompts, models, and logs. Trigger keywords: agentforce pii, pii redaction, data masking llm, einstein trust layer, prompt pii filter, audit pii leakage. Does NOT cover: Shield Platform Encryption at-rest (separate skill), GDPR data subject requests, or classic field-level security policy.
Where can I find the source code?
You can find the source code on GitHub using the link provided at the top of the page.
SKILL.md Source
# Agentforce PII Redaction
## The Trust Layer
Einstein Trust Layer provides the platform boundary: zero retention,
masking on egress, audit trail. It is not a licence to send raw PII.
Redact **before** the trust layer where you can.
## Field-Level Classification
Every field referenced in a prompt needs a classification:
| Class | Default handling |
|---|---|
| Public | Send as-is. |
| Internal | Send if necessary. |
| Confidential | Redact unless explicit business need. |
| Regulated | Mask / tokenize / summarise; never raw. |
Examples (typical, adjust to your compliance):
| Field | Classification |
|---|---|
| Account.Name | Public |
| Contact.Title | Internal |
| Contact.Email | Confidential |
| Contact.SSN__c | Regulated |
| PaymentMethod.CCLast4 | Regulated |
## Redaction Strategies
- **Mask** — `john@acme.com` → `j***@acme.com`.
- **Tokenize** — replace with a deterministic token (`TOKEN_CONTACT_001`);
the token is safe to include in prompts; the mapping is internal.
- **Drop** — omit from the prompt context entirely.
- **Summarise** — replace with a category (`customer with >5y tenure`).
Pick the strategy per field + use case. SSN is nearly always **Drop**.
## Prompt Context Assembly
Build prompts from a **redacted context object**, never from raw SObject
rows. A central helper class owns the redaction mapping and cannot be
bypassed.
## Input-Side Redaction
User turns can contain PII ("my SSN is …"). Options:
- **Detect and refuse** — respond: "Do not share sensitive IDs."
- **Detect and redact** — scrub before prompting the model.
- **Detect and route** — flag, escalate to human.
Pattern: all three are valid; choose per topic sensitivity.
## Output-Side Redaction
Agent outputs might echo input or retrieved content. Second-pass
redaction on responses before sending back. Trust Layer handles the
baseline; the application can tighten.
## Audit Wiring
- Log the redaction event (field name, strategy) without the value.
- Alert on any PII category that should have been redacted but wasn't.
- Review the audit weekly.
## Recommended Workflow
1. Inventory every field read into prompt context.
2. Classify (Public / Internal / Confidential / Regulated).
3. Choose redaction strategy per field.
4. Centralise redaction in a single Apex/Flow boundary class.
5. Add input-side detection for common PII patterns.
6. Emit audit events on redaction and on any leak.
7. Include PII adversarial cases in the eval suite (see
`agentforce/agentforce-testing-strategy`).
## Official Sources Used
- Einstein Trust Layer —
https://help.salesforce.com/s/articleView?id=sf.einstein_trust_layer.htm
- Data Masking For Generative AI —
https://help.salesforce.com/s/articleView?id=sf.einstein_generative_ai_masking.htm
- Agentforce Audit —
https://help.salesforce.com/s/articleView?id=sf.einstein_agent_audit.htmRelated Skills
einstein-bots-to-agentforce-migration
Use when migrating an existing Einstein Bot (legacy or Enhanced) to Agentforce: feature mapping, conversation design translation, cutover planning, hybrid bot/agent architecture, and context handoff. Triggers: 'migrate einstein bot to agentforce', 'convert legacy bot to agentforce', 'einstein bot retiring deadline', 'hybrid bot agentforce pattern', 'bot dialog to topic migration'. NOT for new Agentforce setup with no existing bot — use agentforce/agentforce-agent-creation instead.
data-cloud-grounding-for-agentforce
Use when grounding an Agentforce agent with Data Cloud retrievers, DMO selection, chunking, and freshness windows. Triggers: agent grounding, retriever, DMO, data graph, RAG, vector index, citations. Does NOT cover Data Cloud ingestion pipelines or Data Cloud identity resolution tuning.
agentforce-tool-use-patterns
Pick the right tool shape for each agent action: Apex invocable vs Flow action vs External Service vs Prompt Template vs Data Cloud retrieval. Covers action selection by use case, argument design for LLM clarity, return-shape contracts, error-surfacing, cost implications, and when to chain tools vs keep a single action. NOT for authoring a specific action (use custom-agent-actions-apex). NOT for topic design (use agent-topic-design).
agentforce-testing-strategy
Design Agentforce testing: topic coverage, action unit tests, deterministic golden sets, adversarial prompts, and regression harness. Trigger keywords: agentforce testing, agent eval, agent regression suite, prompt golden set, action unit test agentforce. Does NOT cover: generic LLM evaluation academia, human-labeled RLHF pipelines, or Einstein Classify accuracy.
agentforce-service-ai-setup
Use this skill to verify prerequisites, license entitlements, and org readiness before enabling Einstein for Service AI features: Case Classification, Article Recommendations, Reply Recommendations, and Work Summaries. Trigger keywords: Einstein for Service setup, enable Case Classification, enable Article Recommendations, enable Reply Recommendations, enable Work Summaries, Einstein generative AI prerequisites, Data Cloud for Work Summaries. NOT for core Agentforce agent setup, Agent Builder topic design, Einstein Copilot configuration, ongoing optimization of already-running features, or Einstein Trust Layer configuration.
agentforce-sales-ai-setup
Step-by-step setup and configuration of Einstein for Sales AI features: Opportunity Scoring, Pipeline Inspection AI insights, Einstein email insights and composition, and Forecasting AI. Covers prerequisites, license checks, feature sequencing, and data readiness validation. NOT for core Agentforce agent creation, agent topic design, Einstein Trust Layer configuration, or Einstein Activity Capture troubleshooting.
agentforce-prompt-versioning
Version Prompt Templates and agent topic prompts: source-control shape, change review, model-version pinning, A/B, and rollback. Trigger keywords: prompt template versioning, prompt changelog, prompt rollback, A/B prompt test, agentforce prompt release. Does NOT cover: prompt engineering tips, general LLM fine-tuning, or Classify / Einstein Generate studio UI walkthroughs.
agentforce-production-readiness-checklist
Use when an Agentforce agent is being moved from build/sandbox to live end-user traffic and the team needs a comprehensive readiness gate covering coverage testing, Trust Layer config, guardrails, cost telemetry, observability, rate limits, permissions, rollout strategy, rollback plan, and performance benchmarks. Triggers: 'we want to ship our Agentforce agent next week', 'pre-prod readiness review for our Service Agent', 'what do we need before turning the agent on for real customers', 'agent went live and is hallucinating, what should we have caught', 'cost monitoring for our internal sales agent', 'rollout strategy from internal pilot to GA'. NOT a substitute for the lighter sign-off ritual in agent-deployment-checklist (use this skill instead when the team needs technical depth on what to actually verify, not just sign-off rows). NOT for Trust Layer feature configuration in isolation (use einstein-trust-layer). NOT for designing the guardrails themselves (use agentforce-guardrails) or the test harness (use agentforce-eval-harness).
agentforce-persona-design
Use when defining or refining the tone, voice, and behavioral personality of an Agentforce agent: system instruction encoding, brand voice alignment, adaptive response formats, multi-persona strategies. NOT for agent topic design (use agent-topic-design) or testing methodology (use agent-testing-and-evaluation).
agentforce-observability
Use when monitoring Agentforce agent sessions, analyzing conversation logs, measuring deflection rates, or diagnosing agent performance issues. Triggers: 'agentforce session analytics', 'how to query agent conversation data', 'monitor agentforce agent effectiveness', 'agent deflection rate', 'utterance analysis agentforce'. NOT for Einstein Trust Layer audit logging (use einstein-trust-layer), NOT for agent topic design or guardrails (use agent-topic-design or agentforce-guardrails), NOT for LLM prompt debugging (this skill covers session metrics and conversation trace, not prompt engineering).
agentforce-multi-turn-patterns
Design Agentforce conversations that span multiple turns without losing context: session variable scoping, conversation memory, clarifying-question patterns, topic-to-topic handoff, and the right abstractions for accumulating state across turns. NOT for single-turn agent actions (use agent-actions). NOT for channel-specific conversation UX (use agent-channel-deployment).
agentforce-in-slack
Use when configuring Slack-specific Agentforce capabilities AFTER the basic Slack deployment is complete: enabling Slack-native actions, managing public vs. private action scope, mapping Salesforce-to-Slack user identity, or troubleshooting Slack action failures. Triggers: 'add General Slack Actions topic to agent', 'configure canvas creation for Agentforce in Slack', 'Agentforce private action requires user identity mapping', 'Slack agent cannot send DMs after deployment', 'Agentforce Look Up User action not working in Slack', 'how to map Slack users to Salesforce identities for private actions'. NOT for core Agentforce setup, NOT for basic Slack OAuth installation or DM vs channel-mention mode — those are covered by agentforce/agent-channel-deployment.