AI Fairness and Bias Detection - Agentic Patterns

Learn more about Well-Architected Fairness → Agentic Enterprise Fairness

Patterns

Where to lookWhat good looks like
Agentforce | Agent Configuration✅ Test agents systematically for demographic bias using diverse personas representing protected groups; generate identical requests varying only demographic signals like names, locations, or communication styles
Data 360 | Data Segments✅ Before using Salesforce data for agent training or RAG grounding, audit demographic representation and identify underrepresented groups
Data 360 | Knowledge Base✅ Audit knowledge bases and retrieval sources for demographic balance; ensure documentation quality, completeness, and accuracy are comparable across all customer segments
Agentforce | Agent Monitoring✅ Monitor agent interactions across demographic groups for behavior consistency; analyze response length, detail level, politeness markers, and solution quality stratified by user characteristics
Agentforce | Agent Monitoring✅ Detect feedback loops by monitoring decision distribution changes over time; break loops through human review of edge cases and periodic rebalancing
Data 360 | Training Data✅ Oversample underrepresented demographic groups in training data ensuring balanced representation; synthesize additional data using techniques like SMOTE when real data is insufficient
Agentforce | Prompt Template✅ Constrain agent behavior through system prompts: “Treat all customer demographics with equal professionalism and detail. Provide explanations of comparable length and quality regardless of customer communication style or background.”
Einstein | Prediction Model✅ Calculate correlation coefficients between all model features and protected characteristics; evaluate whether legitimate predictive value justifies inclusion of highly correlated features
Einstein | Prediction Model✅ Identify and remove proxy features that correlate with protected characteristics (zip code proxies for race, name patterns proxy for gender/ethnicity, communication timing proxies for religion/caregiving, device type proxies for income, territory assignment proxies for demographic composition)

Anti-Patterns

Where to lookWhat bad looks like
Data 360 | Training Data⚠️ Using historical CRM data for agent training without auditing demographic representation, allowing algorithmic bias to compound human bias
Agentforce | Agent Configuration⚠️ Deploying agents without systematic demographic bias testing across diverse personas representing protected groups
Data 360 | Knowledge Base⚠️ Providing detailed knowledge articles for enterprise customers but only generic responses for small business customers, perpetuating service quality disparities
Agentforce | Agent Monitoring⚠️ Allowing agents to adapt behavior based on conversational cues that correlate with protected characteristics (communication style, language complexity, timing patterns) without monitoring
Einstein | Prediction Model⚠️ Including proxy features (zip code, name patterns, device type) in consequential decision models without evaluating correlation with protected characteristics
Agentforce | Agent Configuration⚠️ Allowing feedback loops where agent decisions create the disparity that justifies future adverse decisions (e.g., routing agent reducing opportunities to certain territories, generating less success data, further reducing predicted success)
Data 360 | Data Segments⚠️ Defaulting to majority-culture references in content examples that alienate underrepresented users

Patterns

Where to lookWhat good looks like
Agentforce | Agent Monitoring✅ Log complete inputs and outputs for every agent in an orchestration chain so bias in a final outcome can be traced backward to the stage that introduced it and where amplification occurred
Platform | CRM Analytics✅ Build dashboards showing decision distributions at each orchestration stage broken down by demographics to identify where in the chain disparities emerge or expand
Platform | CRM Analytics✅ Calculate demographic parity, equal opportunity, and disparate impact at each stage of the chain, not just on the final outcome; stage-wise metrics catch compounding that looks acceptable at any single stage but degrades across the chain
Agentforce | Model Configuration✅ Maximize diversity across agents in consensus patterns to reduce correlated errors: different foundation models where feasible (such as through Bring Your Own LLM configurations), different retrieval sources, different prompt approaches, and different confidence threshold calibrations
Agentforce | Agent Configuration✅ Insert human review at orchestration inflection points where a stage significantly narrows downstream options, even when individual agent confidence is high, because orchestration creates emergent effects that per-agent confidence scores do not capture
Agentforce | Agent Configuration✅ Configure orchestration circuit breakers that halt and escalate to a human when any agent’s confidence falls below threshold, output distribution shifts significantly between stages, the chain encounters data unrepresented in training or validation, or the user challenges an intermediate decision

Anti-Patterns

Where to lookWhat bad looks like
Agentforce | Agent Monitoring⚠️ Running orchestration chains without logging each agent’s inputs and outputs, making it impossible to trace which stage introduced bias when a discriminatory outcome surfaces
Agentforce | Model Configuration⚠️ Treating multi-agent consensus as bias correction when all agents share the same foundation model or training data, so agreement only validates shared bias rather than correcting it
Platform | CRM Analytics⚠️ Measuring fairness only on the final orchestration outcome, missing stage-wise compounding where each stage appears acceptable in isolation but the chain compounds to an unacceptable disparity
Agentforce | Agent Configuration⚠️ Relying on individual agent confidence scores to gate an orchestration chain without inserting human review at inflection points where a stage narrows downstream options
Agentforce | Agent Configuration⚠️ Allowing an orchestration chain to proceed with no circuit breaker when an upstream agent has low confidence, distributions shift between stages, or the chain encounters data outside its training and validation sets

Patterns

Where to lookWhat good looks like
Business | Deployment Gate✅ Make fairness evaluation a mandatory deployment gate equivalent in authority to security review before deploying any agent making consequential decisions
Einstein | Prediction Model✅ Calculate demographic parity by dividing the population into demographic segments, measuring positive outcome rates per segment, and comparing rates across segments
Einstein | Prediction Model✅ Ensure equal opportunity by verifying agents identify qualified individuals equally well regardless of demographic characteristics; verify true positive rates are comparable across groups
Einstein | Prediction Model✅ Apply the four-fifths rule: if any group receives positive outcomes at less than 80% the rate of the highest-performing group, investigate for bias
Einstein | Prediction Model✅ Calculate disparate impact ratios for all consequential agent decisions and document investigation results when ratios fall below 80%
Business | Documentation✅ Document when demographic parity violations reflect legitimate business factors rather than discrimination
Business | Legal Review✅ Consult legal counsel when implementing equalized odds through group-specific thresholds, noting that in the United States, adjusting scores or cutoffs by protected class for employment-related decisions is prohibited under Title VII (Civil Rights Act of 1991), and that other jurisdictions impose their own constraints

Anti-Patterns

Where to lookWhat bad looks like
Business | Deployment Gate⚠️ Deploying agents making consequential decisions without calculating fairness metrics across demographic groups
Einstein | Prediction Model⚠️ Systematically missing qualified candidates from underrepresented groups by setting higher evidence thresholds for those groups than for majority groups
Einstein | Prediction Model⚠️ Implementing group-specific thresholds without legal justification or validation that thresholds improve fairness without creating other discriminatory effects, or, in the United States, implementing them for employment-related decisions at all, where Title VII prohibits adjusting scores or cutoffs by protected class
Business | Documentation⚠️ Failing to document when fairness metric violations reflect legitimate business differences versus discrimination

Patterns

Where to lookWhat good looks like
Einstein | Trust Layer✅ Use Trust Layer audit data to analyze response distributions across demographic groups; build CRM Analytics dashboards comparing approval rates, recommendation patterns, and risk scores by customer segment
Einstein | Trust Layer✅ Query Trust Layer logs regularly: show all high-value predictions grouped by customer industry, region, and account size to verify prediction distributions are consistent across segments
Agentforce | Session Tracing✅ Sample 100 agent decisions stratified by demographic group and have reviewers assess whether reasoning quality is comparable across groups using Agentforce Session Tracing reasoning traces
Platform | CRM Analytics✅ Configure alerts when agent decision patterns deviate significantly from baseline (e.g., approval rates for a specific segment drop 20% week-over-week)
Platform | CRM Analytics✅ Build custom monitoring to alert on fairness signals: “Notify when average conversation length for any customer segment exceeds 1.5x the overall average for three consecutive days”
Platform | CRM Analytics✅ Build reusable CRM Analytics dashboard templates tracking fairness metrics for common agent use cases including demographic parity, equal opportunity, disparate impact ratios, and trend charts

Anti-Patterns

Where to lookWhat bad looks like
Agentforce | Session Tracing⚠️ Ignoring Agentforce Session Tracing reasoning traces and the Trust Layer audit trail when users report bias, missing which reasoning steps, data sources, or retrieval patterns introduced the bias
Platform | CRM Analytics⚠️ Failing to configure anomaly detection alerts for fairness signals, allowing sudden distribution shifts to go undetected
Platform | CRM Analytics⚠️ Building fairness dashboards ad-hoc per team instead of sharing reusable templates standardizing fairness monitoring

Patterns

Where to lookWhat good looks like
Agentforce | Agent Configuration✅ Design agents for high-stakes decisions (employment, credit, healthcare, legal, housing) to analyze, recommend, and prepare decisions while requiring human approval before execution
Platform | Approval Process✅ Configure Agentforce to generate recommendations with supporting evidence, route above-threshold decisions to human reviewers via Approval Process, and surface agent reasoning traces in approval interface
Agentforce | Agent Configuration✅ Implement confidence-based escalation: high confidence (>90%) proceeds autonomously, moderate (70-90%) requires human review, low (<70%) defers entirely to human with context summary
Agentforce | Agent Configuration✅ Calibrate confidence thresholds using production data; sample agent decisions at each confidence band and calculate actual success rates to validate calibration
Agentforce | Agent Configuration✅ Set time-bounded escalation triggers: after 5-7 turns without resolution, after 10 minutes elapsed, when user expresses frustration, or when agent repeats same response
Platform | Omni-Channel✅ Configure Omni-Channel to route escalated cases to appropriate skill-based queues with full conversation context; train human agents to handle escalations without requiring users to repeat information
Agentforce | UX✅ Provide a “Connect with human agent” option in agent chat interfaces at all times, not gated behind a minimum number of failed interactions
Agentforce | Agent Configuration✅ Enable authorized humans to override any agent decision at any point with documented rationale; add “Override Agent” button to all agent decision interfaces requiring multi-line text explanation
Platform | Custom Object✅ Log overrides to custom object with decision ID, override timestamp, rationale, and overrider; create CRM Analytics dashboard tracking override patterns by agent, decision type, and customer demographics
Business | Documentation✅ Require documented rationale for approvals or denials diverging from agent recommendation; track override patterns to identify systematic agent deficiencies

Anti-Patterns

Where to lookWhat bad looks like
Agentforce | Agent Configuration⚠️ Deploying agents as autonomous decision-makers for high-stakes decisions (employment, credit, healthcare, legal, housing) without human approval gates
Agentforce | Agent Configuration⚠️ Using miscalibrated confidence scores where “high confidence” decisions succeed only 75% of the time, undermining trust in escalation mechanisms
Agentforce | Agent Configuration⚠️ Trapping customers in unhelpful automated loops without time-bounded or turn-bounded escalation to humans
Platform | Omni-Channel⚠️ Routing escalated cases without full conversation context, forcing users to repeat information already provided to the agent
Business | Governance⚠️ Allowing diffusion of responsibility where everyone assumes the AI is accountable and no one takes ownership of decisions
Business | Governance⚠️ Assigning accountability after incidents occur rather than before deployment with named individuals documented in architecture decision records

Patterns

Where to lookWhat good looks like
Business | Documentation✅ Assign accountability before deployment: model owner (fairness/accuracy), decision owner (use case/impact), appeal authority (fairness appeals), audit authority (periodic audits)
Business | Documentation✅ Document all accountability roles with names, not just titles, ensuring accountability persists through organizational changes
Business | Governance✅ Empower appeal authority (ethics review board or designated team) to handle appeals from users who believe agent decisions were unfair with accessible, timely processes that can reverse agent decisions
Business | Governance✅ Authorize audit authority (compliance team) to conduct periodic audits and suspend agents failing fairness standards pending remediation
Business | Governance✅ Schedule monthly review of high-frequency override patterns with agent development team

Anti-Patterns

Where to lookWhat bad looks like
Business | Governance⚠️ Documenting accountability roles with titles only (e.g., “Data Science Lead”) without named individuals, allowing accountability to disappear during organizational changes
Business | Governance⚠️ Allowing decision owners to delegate accountability to AI systems rather than retaining personal accountability for real-world impact
Business | Governance⚠️ Making appeal processes inaccessible, untimely, or lacking authority to reverse agent decisions

Patterns

Where to lookWhat good looks like
Agentforce | Agent Configuration✅ Architect multiple explanation layers: user-facing (plain language), business (key decision factors), technical (model details/feature weights), and audit (complete decision traces)
Agentforce | Agent Configuration✅ Display real-time summary reasoning when making decisions: “I recommended Product A based on your purchase history (3 similar purchases), current promotion (20% off), and inventory availability”
Agentforce | Agent Configuration✅ Provide “Why did you recommend this?” link enabling users to view complete reasoning including all knowledge articles retrieved, records consulted, and decision logic
Agentforce | Session Tracing✅ Retrieve reasoning traces from Agentforce Session Tracing for historical reconstruction when users challenge past decisions weeks or months later
Agentforce | Session Tracing✅ Sample Agentforce Session Tracing reasoning traces stratified by demographic groups to analyze whether decision quality is consistent; review 100 traces from each customer segment
Agentforce | Agent Configuration✅ Communicate confidence in user-appropriate terms: “High confidence” / “Moderate confidence” / “Low confidence” with plain language reliability context instead of raw probability scores
Agentforce | Agent Configuration✅ For moderate or low confidence decisions, explain what additional review will occur and expected timeline
Agentforce | Agent Configuration✅ Provide counterfactual explanations showing users what would change the outcome: “Your lead score would increase with verified employment information (+8 points)”

Anti-Patterns

Where to lookWhat bad looks like
Agentforce | Agent Configuration⚠️ Displaying raw probability scores (e.g., “73% confidence”) that users misinterpret instead of user-appropriate confidence language
Agentforce | Agent Configuration⚠️ Providing counterfactuals that encourage gaming the system or inadvertently reveal protected characteristics as decision factors (e.g., “Your application would score higher if you were 10 years younger”)
Agentforce | Agent Configuration⚠️ Providing only one level of explanation that serves neither end users (too technical) nor auditors (too simplified)
Agentforce | Agent Configuration⚠️ Failing to store historical reasoning traces, making it impossible to explain past decisions when challenged

Patterns

Where to lookWhat good looks like
Agentforce | Agent Configuration✅ Clearly indicate when users interact with Agentforce agents with prominent, continuous disclosure (not buried in terms of service or shown only once at interaction start)
Agentforce | Agent Configuration✅ Display persistent visual indicators: agent avatar marked “AI Assistant,” header showing “You’re chatting with an Agentforce service agent,” and visible “Connect with human agent” option throughout conversation
Agentforce | Agent Configuration✅ Ensure users who prefer human interaction can switch without friction or service quality penalties

Anti-Patterns

Where to lookWhat bad looks like
Agentforce | Agent Configuration⚠️ Burying AI disclosure in terms of service or showing it only once at interaction start rather than maintaining persistent visual indicators
Agentforce | Agent Configuration⚠️ Penalizing users who choose human interaction with degraded service or increased wait times
Agentforce | Agent Configuration⚠️ Hiding the “Connect with human agent” option or requiring users to complete agent interactions before accessing human support

Patterns

Where to lookWhat good looks like
Platform | Event Monitoring✅ Configure Event Monitoring to track: API event logs showing agent system interactions, login events for agent service accounts, and report exports when agents access bulk data
Platform | Event Monitoring✅ Stream Event Monitoring logs to external SIEM for long-term retention and tamper evidence meeting regulatory requirements
Platform | Field Audit Trail✅ Enable Field Audit Trail (indefinite retention of archived history) on the objects that store or are influenced by agent decisions, choosing from the standard objects it supports plus custom objects with field history tracking enabled; critical for credit, employment, and other high-stakes use cases
Platform | Shield✅ Use Shield Event Monitoring for highest-stakes agent decisions where audit trail integrity is critical (enhanced capabilities, structured event fields, compliance reporting integration)
Platform | Custom Object✅ Create AgentDecision__c custom object storing: Decision_ID, Timestamp, Agent_Type, Model_Version, Confidence_Score, Outcome, User_ID, Record_ID; with junction object DecisionInput__c linking decisions to input feature snapshots
Platform | Custom Object✅ Store exact input feature values at decision time (not current values which may change) to enable testing whether decisions would differ with updated data
Business | Documentation✅ Record model version, configuration state (thresholds, business rules, parameters), and environmental context (user identity, timestamp, conversation history) for every consequential decision
Platform | LWC✅ Build LWC component embedding in records to display decision history with “View Reasoning” links

Anti-Patterns

Where to lookWhat bad looks like
Platform | Field Audit Trail⚠️ Relying on standard field history (18-month retention) for high-stakes decisions requiring multi-year regulatory retention
Platform | Custom Object⚠️ Storing only current record values instead of decision-time input snapshots, making it impossible to reconstruct historical decisions accurately
Business | Documentation⚠️ Failing to record model version and configuration state active at decision time, making root cause analysis impossible when bias is discovered
Platform | Event Monitoring⚠️ Keeping audit logs only within Salesforce without external SIEM streaming, lacking tamper evidence for regulatory compliance

Patterns

Where to lookWhat good looks like
Platform | CRM Analytics✅ Build dashboards showing: decision distribution by demographic segment (bar charts), fairness metrics over time (trend lines with alert thresholds), disparate impact ratio with four-fifths rule indicator, top predictors contributing to decisions, override rate by demographic group
Platform | CRM Analytics✅ Alert when input distributions change significantly from training data demographics; monitor for distribution shift indicating potential fairness degradation
Agentforce | Agent Configuration✅ Add “Report Concern” option to Agentforce chat interface; route reports to ethics review board with complete conversation context; track in custom object with required investigation workflow
Business | Governance✅ Monitor fairness dashboards weekly, investigate threshold breaches within 24 hours, conduct monthly override pattern analysis, perform quarterly comprehensive fairness audits
Business | Governance✅ Enable users to report perceived bias through accessible mechanisms; user reports surface qualitative issues that quantitative metrics miss

Anti-Patterns

Where to lookWhat bad looks like
Business | Governance⚠️ Treating bias detection as one-time pre-deployment validation rather than continuous monitoring as agents evolve through prompt updates, model version changes, and shifting data distributions
Platform | CRM Analytics⚠️ Failing to configure alerts when fairness metrics breach defined thresholds, allowing gradual drift to go undetected
Agentforce | Agent Configuration⚠️ Not providing users with accessible mechanisms to report perceived bias, missing qualitative issues that quantitative metrics cannot surface

Patterns

Where to lookWhat good looks like
Platform | Integration✅ Build export functionality enabling compliance teams to extract complete decision datasets with demographic stratification for external auditor review while respecting data privacy regulations
Business | Documentation✅ Maintain current documentation covering: agent purpose and intended use cases, training data demographics and known limitations, fairness metrics (pre-deployment and production), bias mitigation strategies, human oversight configurations, monitoring approach, and accountability assignments
Business | Governance✅ Before deploying agents making consequential decisions, conduct algorithmic impact assessments evaluating potential positive and negative effects across stakeholder groups
Business | Governance✅ Schedule semi-annual external fairness audits by third parties; maintain regulatory horizon scanning adapting to EU AI Act, state laws, and sector regulations

Anti-Patterns

Where to lookWhat bad looks like
Platform | Integration⚠️ Designing agent architectures without anticipating external audits by regulators, civil rights organizations, or customers; lacking export capabilities for auditor review
Business | Documentation⚠️ Allowing documentation to become stale as agents evolve through prompt updates or model changes without updating fairness metrics and mitigation strategies
Business | Governance⚠️ Deploying agents making consequential decisions without conducting impact assessments evaluating effects across stakeholder groups

Patterns

Where to lookWhat good looks like
Agentforce | Agent Configuration✅ Require explicit opt-in for consequential agents (credit decision, employment screening, service eligibility) rather than activating by default
Platform | Custom Object✅ Create AI_Consent__c with fields: User_ID, Use_Case, Grant_Date, Grant_Method, Withdrawal_Date, Status; enable Field Audit Trail for long-term consent record retention
Agentforce | Agent Configuration✅ Add consent check to agent invocation logic: query AI_Consent__c for active consent before processing; support use-case-level consent (service agent vs sales agent)
Platform | Community/Experience Site✅ Build self-service consent management interface accessible from user profile; schedule consent renewal reminders every 12 months
Agentforce | Agent Configuration✅ When deploying new capabilities affecting existing users, proactively seek consent for new use cases rather than relying on original consent that did not contemplate specific applications
Agentforce | Agent Configuration✅ Enable consent withdrawal at any time with immediate cessation of agent processing; make withdrawal as easy as granting without requiring support contacts
Agentforce | Agent Configuration✅ Enable user preference management: communication style (concise vs detailed), proactiveness level, escalation preference; store in User record custom fields and reference in agent system prompts
Platform | Omni-Channel✅ Provide persistent “Connect with human agent” option throughout interactions; when selected, route to Omni-Channel with full conversation context and priority flag with no degraded service or wait time penalties

Anti-Patterns

Where to lookWhat bad looks like
Agentforce | Agent Configuration⚠️ Activating consequential agents by default without explicit user opt-in understanding how agents will affect them
Agentforce | Agent Configuration⚠️ Requiring blanket AI consent rather than enabling granular, use-case-level consent (a customer might consent to service agents while declining sales agents)
Agentforce | Agent Configuration⚠️ Making consent withdrawal require support contacts or bureaucratic processes instead of self-service with immediate effect
Agentforce | Agent Configuration⚠️ Relying on original consent when deploying materially new agent capabilities that were not contemplated in the original consent
Agentforce | Agent Configuration⚠️ Requiring users to complete agent interactions or explain their preference before connecting them with a human agent

Patterns

Where to lookWhat good looks like
Agentforce | Agent Configuration✅ Review all data sources used in agent reasoning for protected characteristic content; remove or mask protected characteristics from agent inputs when not legally justified for the decision type
Agentforce | Prompt Template✅ Include explicit non-discrimination instructions in system prompts: “Treat all customers with equal professionalism regardless of name, location, communication style, or any other characteristic. Never make assumptions about customers based on demographic characteristics.”
Agentforce | Agent Configuration✅ Test with diverse personas before production: generate identical requests varying only demographic signals (names suggesting different ethnicities, locations, communication styles); compare responses for quality, length, professionalism, and outcomes
Agentforce | Agent Monitoring✅ Calculate first-contact resolution rates stratified by customer segment; investigate when significant disparities exist between segments
Agentforce | Agent Monitoring✅ Sample agent conversations across segments and have reviewers assess response quality on consistent dimensions (accuracy, completeness, professionalism, helpfulness) with inter-rater reliability testing
Agentforce | Agent Monitoring✅ Track escalation rates stratified by customer demographics; higher rates for certain segments indicate agents are less effective for those users
Business | Surveys✅ Survey users from all demographic groups and compare satisfaction scores; overall high satisfaction may mask poor experiences for minority groups

Anti-Patterns

Where to lookWhat bad looks like
Agentforce | Agent Configuration⚠️ Including protected characteristic data (disability status, religion, national origin) in agent inputs for decisions where those characteristics are not legally justified
Agentforce | Prompt Template⚠️ Deploying agents without explicit non-discrimination instructions in system prompts, relying on foundation model behavior alone
Agentforce | Agent Configuration⚠️ Testing agents only with majority-demographic personas, missing bias against underrepresented groups
Business | Surveys⚠️ Relying on overall satisfaction scores that mask poor experiences for minority groups drowned out by majority satisfaction

Patterns

Where to lookWhat good looks like
Platform | LWC✅ For custom chat implementations, ensure: semantic HTML structure with proper landmarks, ARIA live regions announcing new messages, keyboard navigation through message history, clear focus indicators on interactive elements
Platform | LWC✅ Test with JAWS, NVDA, and VoiceOver throughout development, not just automated scanning
Agentforce | Agent Configuration✅ Configure agent responses to use plain language for general audiences; avoid jargon and provide glossaries for unavoidable technical terms; structure long responses with headings and bullet lists
Agentforce | Agent Configuration✅ Validate response quality is comparable across languages through native speaker evaluation, not just automated metrics that miss cultural nuance
Agentforce | Agent Configuration✅ For business-critical use cases serving diverse language populations, configure language-specific handling (language variables, localized prompts, knowledge article filtering by language) and validate quality per language rather than relying solely on foundation model cross-lingual transfer

Anti-Patterns

Where to lookWhat bad looks like
Platform | LWC⚠️ Building custom chat implementations without manual accessibility testing (relying only on automated scanning)
Agentforce | Agent Configuration⚠️ Using jargon and complex language without glossaries, creating cognitive barriers for users
Agentforce | Agent Configuration⚠️ Relying solely on foundation model cross-lingual transfer for lower-resource languages, providing degraded quality for non-majority language users
Agentforce | Agent Configuration⚠️ Validating multilingual response quality only through automated metrics that miss cultural nuance

Patterns

Where to lookWhat good looks like
Business | Governance✅ Establish ethics review board with enforcement authority before deploying any production agents
Business | Governance✅ Define mandatory deployment gates: data audit, fairness metrics calculation, and algorithmic impact assessment
Platform | CRM Analytics✅ Create baseline fairness monitoring dashboards in CRM Analytics before first agent deployment
Platform | Event Monitoring✅ Enable Einstein Trust Layer audit capture and configure Event Monitoring for agent actions as foundation before production
Business | Governance✅ Implement automated fairness regression testing in CI/CD pipelines for scaling
Business | Governance✅ Standardize consent management patterns across all new agents as deployments scale
Platform | Integration✅ Build agent decision reconstruction tooling enabling compliance team self-service

Anti-Patterns

Where to lookWhat bad looks like
Business | Governance⚠️ Deploying production agents before establishing an ethics review board with enforcement authority
Business | Governance⚠️ Treating fairness as an afterthought rather than a mandatory deployment gate equivalent to security review
Business | Governance⚠️ Skipping algorithmic impact assessments before deploying agents making consequential decisions
Platform | CRM Analytics⚠️ Deploying first production agent without baseline fairness monitoring dashboards in place