Fundamentals
AI & Intelligence
Apply reasoning, extraction, and generation to key workflow moments with Agents and lightweight AI actions
AI in Elementum provides programmatic reasoning, data extraction, and content generation within your workflows. Configure reusable Agents for multi-step processes or lightweight AI actions for specific tasks like classification, summarization, and field extraction.
Example Workflow:
Two Ways to Add Intelligence
Agents
Reusable AI components with tools, memory, and policy controls. Ideal for multi-step reasoning and actions within a workflow.
AI Actions
Lightweight actions powered by LLMs for one-off tasks like categorization, extraction, and transformation.
Common Use Cases
Support Triage Agent
Conversational agent that answers FAQs, gathers context, and triages requests via chat or phone. Creates/updates records, kicks off workflows, and escalates to L2 with a structured handoff when needed.
AP Vendor Outreach
Accounts Payable agent emails vendors for missing documents (W-9, PO, invoice details), validates responses, updates Element fields, and advances the approval workflow automatically.
Triage & Routing
Classify incoming requests, detect intent, and assign to the right team with confidence scores.
Information Extraction
Pull structured fields from unstructured content (emails, PDFs, log files) into Elements.
Summarization & Drafting
Generate summaries, replies, or knowledge base entries with human approval steps.
Search & Reasoning
Retrieve relevant context and reason over it to propose next steps or detect anomalies.
Enterprise AI Orchestration
The key to successful AI implementation is embedding non-deterministic AI capabilities within deterministic workflow structures. This approach enables reliable, auditable, and scalable AI deployment.Core principle: Use deterministic workflows to contain AI uncertainty, ensuring predictable business outcomes regardless of AI model variability.
Governance & Controls
Workflow Boundaries
Define clear input/output contracts and validation rules that AI actions must respect.
Security & Permissions
Respect roles and data access controls when reading or writing records.
Transparency
Log prompts, context, and outputs for auditability and improvement.
Human Oversight
Require approvals for high-impact actions; build review queues into your flow.
Provider Choice
Use different providers and models (OpenAI, Gemini, Cortex, etc.) based on data sensitivity, cost, and latency.
Compliance Assurance
Built-in audit trails, approval workflows, and policy enforcement for regulatory requirements.
Orchestration Patterns
Structured AI Workflows
Structured AI Workflows
Embed AI actions within predefined workflow paths with known outcomes and fallback procedures.Pattern: Input validation → AI processing → Confidence evaluation → Route based on score → Human review if needed → Execute actionBenefits: Predictable outcomes, risk mitigation, compliance assurance
Multi-Agent Coordination
Multi-Agent Coordination
Coordinate multiple AI agents with different specializations within a single business process.Example: Document processing agent extracts data → Classification agent categorizes → Approval agent routes to appropriate reviewerImplementation: Use workflow variables to pass context between agents, maintain audit trail of all agent interactions
Human-AI Collaboration
Human-AI Collaboration
Design workflows where AI handles routine processing while humans focus on exceptions and strategic decisions.Approach: AI processes high-confidence cases automatically, routes uncertain cases to human review with full context and recommendationsKey: Provide humans with AI reasoning and confidence scores to make informed decisions
Confidence Scoring & Error Handling
AI actions return confidence scores that determine workflow routing and human oversight requirements.Confidence Thresholds
Confidence Thresholds
Configure score ranges to automatically route AI outputs based on reliability.High confidence (90-100%): Auto-approve and execute actions
Medium confidence (70-89%): Route to human review queue
Low confidence (0-69%): Escalate to exception handling or manual processing
Medium confidence (70-89%): Route to human review queue
Low confidence (0-69%): Escalate to exception handling or manual processing
Set thresholds based on business impact—use higher thresholds for financial or compliance-critical actions.
Error Recovery
Error Recovery
Handle AI failures with fallback actions and retry policies.Timeout handling: Set maximum execution time for AI actions
Retry logic: Configure retry attempts for transient failures
Fallback routing: Define manual processes when AI is unavailable
Retry logic: Configure retry attempts for transient failures
Fallback routing: Define manual processes when AI is unavailable
Implementation Patterns
Agent Handoffs
Agent Handoffs
Configure structured data transfer between AI and human steps in workflows.Context preservation: Store reasoning steps, input data, and confidence scores
Escalation triggers: Define when to route to human review or alternative processing
Handoff format: Standardize data structure for consistent processingImplementation: Use workflow variables to pass AI outputs and metadata to subsequent steps
Escalation triggers: Define when to route to human review or alternative processing
Handoff format: Standardize data structure for consistent processingImplementation: Use workflow variables to pass AI outputs and metadata to subsequent steps
Embedded AI Actions
Embedded AI Actions
Integrate lightweight AI processing within automation rules for real-time decisions.Event-driven: Trigger AI actions on field changes, record creation, or scheduled intervals
Bounded execution: Set timeouts and resource limits to prevent runaway processes
Idempotent design: Ensure repeated executions produce consistent results
Bounded execution: Set timeouts and resource limits to prevent runaway processes
Idempotent design: Ensure repeated executions produce consistent results
AI actions within automations should be fast-executing (under 30 seconds) to avoid workflow delays.
Async Request-Response
Async Request-Response
Handle multi-step AI processes that require external data or human input.State management: Persist workflow state between request and response cycles
Timeout handling: Set deadlines for responses with escalation paths
Validation: Verify response format and content before proceedingUse case: Agent requests missing information via email, resumes processing when received
Timeout handling: Set deadlines for responses with escalation paths
Validation: Verify response format and content before proceedingUse case: Agent requests missing information via email, resumes processing when received
Tool Integration
Tool Integration
Allow AI agents to interact with external systems and platform APIs.Permission controls: Restrict tool access based on agent role and data sensitivity
Audit logging: Track all tool usage for compliance and debugging
Rate limiting: Prevent excessive API calls and resource consumptionAvailable tools: Database queries, HTTP requests, file operations, notification sending
Audit logging: Track all tool usage for compliance and debugging
Rate limiting: Prevent excessive API calls and resource consumptionAvailable tools: Database queries, HTTP requests, file operations, notification sending
Context Retrieval (RAG)
Context Retrieval (RAG)
Enhance AI responses by retrieving relevant data before generation.Data sources: Search across records, documents, and external knowledge bases
Context filtering: Apply permissions and relevance scoring to retrieved data
Caching strategy: Store frequently accessed context to improve response timesConfiguration: Define search scope, relevance thresholds, and context window limits
Context filtering: Apply permissions and relevance scoring to retrieved data
Caching strategy: Store frequently accessed context to improve response timesConfiguration: Define search scope, relevance thresholds, and context window limits
Technical Implementation
Prompt Engineering
Prompt Engineering
Design effective prompts for consistent AI behavior and output formatting.Structure: Use system prompts for behavior, user prompts for specific tasks
Output format: Specify JSON schemas or structured formats for parsing
Examples: Include few-shot examples for complex classification tasksBest practices: Keep prompts concise, test with edge cases, version control changes
Output format: Specify JSON schemas or structured formats for parsing
Examples: Include few-shot examples for complex classification tasksBest practices: Keep prompts concise, test with edge cases, version control changes
Model Selection
Model Selection
Choose appropriate AI models based on task requirements and constraints.Factors: Latency requirements, data sensitivity, cost constraints, accuracy needs
Provider options: OpenAI (general purpose), Gemini (multimodal), Snowflake Cortex (data residency)
Model types: Classification models for routing, generation models for content creation
Provider options: OpenAI (general purpose), Gemini (multimodal), Snowflake Cortex (data residency)
Model types: Classification models for routing, generation models for content creation
Start with general-purpose models, then optimize for specific use cases based on performance metrics.
Performance Monitoring
Performance Monitoring
Track AI action performance and reliability over time.Metrics: Response time, confidence scores, success/failure rates, user feedback
Alerting: Set thresholds for performance degradation or high failure rates
Optimization: A/B test different prompts and models based on outcomes
Alerting: Set thresholds for performance degradation or high failure rates
Optimization: A/B test different prompts and models based on outcomes
Next Steps
Agents Architecture
Learn how Agents are configured and powered.
AI Actions
Explore lightweight AI actions inside automations.
AI Providers
Configure models and data residency options.
Automations
Trigger AI actions inside event-driven rules and workflows.
Core Concepts
See where AI fits among Apps, Elements, Automations, and Collaboration.