AI Design Patterns

AI systems typically fall into these core patterns, which can be combined to create more sophisticated solutions.

1. Static Knowledge Assistance

• AI helps answer questions from a fixed corpus of information
• Can be conversational/interactive with back-and-forth dialogue or simple Q&A
• Example: company policy chatbot, documentation Q&A, knowledge base assistant, troubleshooting guides, tax filing assistance, onboarding wizards

2. Entity-Based Processing

• AI processes data about one or more entities, either interactively or in batch mode
• Interactive mode: User queries specific IDs on-demand, AI responds in real-time (e.g., "summarize claim #12345" or "compare these 3 vendors")
• Batch mode: AI processes many entities at scale, typically scheduled or triggered, results stored for later use (e.g., classify 10,000 documents overnight)
• Can be informational (aggregating/comparing existing data) or predictive (recommending actions)
• Example: claim summaries, patient records, account analysis, personalized product recommendations, vendor comparisons, competitive analysis, document classification pipelines, data enrichment, entity extraction

3. Generative/Creative Production

• Creating new content rather than querying existing data
• Typically requires human review and editing before use
• Example: marketing copy generation, code writing, synthetic data creation, product descriptions, training simulations, language learning content, practice scenarios

4. Real-time Monitoring & Decision Systems

• AI watches data streams and either alerts on anomalies or takes automated actions
• May operate autonomously for low-risk decisions, but typically includes human override capabilities
• Example: fraud detection, quality control, dynamic pricing, system health monitoring, network security

5. Multi-Step Assisted Workflows

• AI helps execute complex tasks with dependencies over multiple interactions with human checkpoints
• AI maintains context and state across sessions
• Simple proven autonomy: AI can execute well-defined, low-risk steps automatically (e.g., sending calendar invites, creating standard documents)
• Complex requiring verification: Human approval required for key decisions and non-routine actions
• Example: research with iterative refinement, code development with reviews, travel planning with approvals, document compilation with feedback loops, RFP responses, meeting scheduling

Implementation Dimensions

These characteristics apply across the core patterns and define how you configure and enhance your AI system:

1. Prompt Engineering

• How you structure requests and frame problems
• Controls: explanation depth, optimization criteria, temporal scope (historical vs predictive vs scenario planning), comparison requirements, cross-domain synthesis
• Determines output style, reasoning approach, and response format

2. Governance & Validation

• Rules, constraints, guardrails, and compliance checks applied to AI behavior
• Includes: security/privacy controls, bias prevention, ethical considerations, quality assurance, exception handling
• Example: "Don't discount a person's disability when filtering candidates", contract review for missing clauses, compliance checking
• Can apply to any pattern as an overlay

3. Feedback Loops

• Interactive refinement capability where AI learns from user responses
• Applicable in interactive modes across patterns 1, 2, 3, and 5
• Not typically applicable to batch processing or real-time monitoring
• In pattern 5, feedback loops are a core requirement

4. Execution Mode

• Proactive vs Reactive: Does AI wait for explicit requests or anticipate needs?
• Interactive vs Batch: Real-time user-driven or scheduled bulk processing?
• Background vs Foreground: Passive observation or active engagement?

5. Scope

• Entity scope: Single entity vs multiple entities vs cross-entity patterns
• Domain scope: Single domain vs cross-domain synthesis
• Temporal scope: Current state, historical analysis, predictive forecasting, or scenario simulation

6. Autonomy Level

• Fully automated execution vs human-in-the-loop checkpoints
• Varies by risk level and complexity of decisions
• Can range from AI as advisor (human decides) to AI as executor (human monitors)

Real-World Systems: Combining Patterns

Most production AI systems combine multiple core patterns with specific implementation dimensions.

Example - Contextual Co-pilot/Augmentation:

• Patterns: 2 (user as entity) + 5 (multi-step assistance)
• Implementation Dimensions: Proactive execution mode, background operation, continuous feedback loops
• Examples: GitHub Copilot, Google Smart Compose, meeting transcription assistants

Example - Smart CRM Assistant:

• Patterns: 2 (account data) + 3 (draft emails) + 5 (follow-up workflows)
• Implementation Dimensions: Interactive mode, governance rules for communication tone, feedback loops for learning preferences

Example - Intelligent Document Processor:

• Patterns: 2 (extract entities in batch) + 4 (flag anomalies) + 3 (generate summaries)
• Implementation Dimensions: Batch execution, validation rules for data quality, multi-entity scope

 Note: "AI Agents" typically map to Pattern 5, though the term is used loosely in marketing.