In Part 1, we covered the basics of adapting Cynefin for to have better conversations about prioritisation of using AI. In this follow-up, we add more information about when to use which approaches: demystifying AI from Magic Fairy Pixie Dust to the specific technologies that you might employ in a situation to get the best results.
Want more like this? Apply to join CPO Circles today.
Technology Mapping Table
Domain | Recommended Technologies | Why This Approach | Implementation Considerations | Examples |
|---|
Clear | • RPA (Robotic Process Automation) • Rule-based automation • Simple API integrations • Workflow automation tools • Traditional scripting | • Process is fully understood • Rules are explicit and stable • Data is structured • No judgment required • Deterministic outcomes | • Focus on reliability over intelligence • Prioritize maintainability • Clear error handling • Simple monitoring • Low technical risk | • Auto-routing support tickets by keyword • Scheduled report generation • Data entry from structured forms • Standard invoice processing |
|---|
Complicated | • Classical ML (Random Forests, XGBoost) • Fine-tuned LLMs (task-specific) • Basic RAG (Retrieval-Augmented Generation) • NLP for classification • Computer vision models • Decision trees with expert rules | • Expertise can be codified • Patterns exist in data • Outcomes are predictable within known parameters • Requires specialist knowledge • Cause-effect relationships are discoverable | • Include expert review/validation • Build confidence scoring • Create exception handling paths • Plan for model retraining • Document decision logic • Human oversight for edge cases | • Credit risk assessment • Medical diagnosis support • Legal document review • Technical troubleshooting guides • Vehicle damage assessment |
|---|
Complex | • LLMs with human-in-the-loop • Advanced RAG (with context management) • Agentic AI (constrained, with approval gates) • Hybrid ML + rules systems • Semantic search + human judgment • Conversational AI with escalation | • Multiple valid approaches exist • Context heavily influences outcomes • Human judgment adds critical value • Unstructured or incomplete data • Dynamic interactions affect downstream decisions | • Design clear escalation paths • Provide rich context to humans • Log all AI suggestions (and the human evaluation of them) for learning • A/B test different approaches • Build feedback loops • Enable human override always | • Customer negotiation support • Content moderation • Strategic research assistance • Personalized healthcare plans • Complex sales guidance |
|---|
Chaotic (Internal Sensemaking) | • LLM analysis tools • Quick data exploration prototypes • Pattern detection algorithms • Sentiment/theme clustering • Analytics experimentation • Internal research chatbots | • System behavior is unpredictable • Rules cannot yet be established • Need to discover what patterns exist • Internal-facing, low external risk • Learning mode to understand the problem | • Minimize investment • Move fast—you're the user • Expect to throw away prototypes • Focus on discovering patterns • Don't worry about polish • Instrument everything • Iterate based on insights | • Analyzing why a new feature confuses users • Clustering support tickets to find themes • Exploring customer segments with unclear needs • Testing feasibility of novel AI approaches • Understanding unexplained data patterns |
|---|
Chaotic (Discovery / Experiments) | • Simple user-facing pilots • Lightweight LLM prototypes (beta-labeled) • No-code/low-code AI tools • Heavily instrumented MVPs • Limited-release experiments • Opt-in exploratory features | • System behavior is unpredictable • External-facing, higher risk • Need to probe real user response • Rapid change expected • Learning what actually works in the wild | NOTE: This area is high risk, and should only be used when appropriate.
• Set clear expectations ("beta," "experimental") • Transparent labeling with users • Easy human escalation/fallback • Limited rollout (opt-in, small %) • Monitor obsessively • Be ready to pull back quickly • Gather explicit user feedback | • Feasibility testing and alignment with internal teams and partners • Supervised or explicitly labelled user testing of concepts • Testing novel interaction patterns |
|---|
Technology Deep Dive
Clear Domain: Focus on Reliability
Primary Approach: RPA and Rule-Based Automation
- Traditional workflow automation
- If-then logic with known conditions
- Structured data processing
- API orchestration
When NOT to use AI: If rules are simple and stable, avoid the overhead of AI/ML. Traditional automation is cheaper, more reliable, and easier to maintain
Complicated Domain: Leverage Expertise
Primary Approaches:
- Classical ML: For pattern recognition where labelled data exists (classification, regression, forecasting)
- Fine-tuned LLMs: When language understanding is needed but domain-specific accuracy is critical
- Basic RAG: To ground responses in verified knowledge bases without extensive customization
Key Success Factor: The ability to trace decisions back to specific inputs and rules. Build systems that explain their reasoning.
Complex Domain: Augment Human Judgment
Primary Approaches:
- LLMs with Human-in-the-Loop: AI suggests, humans decide with full context
- Advanced RAG: Sophisticated context retrieval to support nuanced decisions
- Constrained Agentic AI: Agents that can take preliminary actions but escalate for approval
Key Success Factor: Design for human-AI collaboration. AI reduces cognitive load and provides options; humans apply judgment, ethics, and contextual wisdom
Chaotic Domain: Probe and Sense
Two Distinct Scenarios:
Internal Sensemaking:
- Fast prototyping with LLMs: Quick experiments to understand the problem space
- Pattern discovery tools: Analytics and clustering to find signal in noise
- Research-grade AI: Using AI as your research assistant, not your product
Key Success Factor: Speed of learning. You're the user, so iterate freely without external risk.
Discovery/Experiments:
- Minimal viable AI: Simple implementations to gather real-world data under controlled circumstances
- Design sprints: Build and test concepts to generate alignment, understand viability, and test feasibility
- Transparent pilots: Beta programs with explicit experimental labeling
- Heavy instrumentation: Focus on measurement and learning, not optimization
Key Success Factor: Balance learning speed with user safety. Set expectations, provide fallbacks, monitor obsessively.
Simplified Decision Flow
Start with the Discovery Questions from Part 1
↓
1. Is the process stable and rule-based?
YES → Clear → Use RPA
NO → Continue
↓
2. Can experts codify their decision-making?
YES → Complicated → Use ML/Fine-tuned LLMs + RAG
NO → Continue
↓
3. Is human judgment essential at key points?
YES → Complex → Use LLMs with Human-in-the-Loop + Advanced RAG
NO → Continue
↓
4. Is the system poorly understood/rapidly changing?
YES → Chaotic (Internal) → Use Pattern Discovery Tools
OR → Chaotic (External) → Use for Discovery/Alignment purposes, or experiment (carefully) in a Transparent Beta with appropriate Safety Rails
Anti-Patterns to Avoid
Don't Use | When | Use Instead |
Agentic AI | Process is Clear or well-mapped | RPA or rule-based automation |
Heavy ML infrastructure | System is Chaotic | Quick LLM prototypes |
Rule-based systems | Process is Complex with context-dependency | LLM + human-in-the-loop |
Fully autonomous AI | Human judgment is legally/ethically required | Augmentation, not automation |
Enterprise LLM deployment | Still in Chaotic discovery | Lightweight experimentation |
Movement between domains
As you learn and stabilize processes, opportunities may shift domains:
- Chaotic → Complex: After experimentation reveals patterns
- Complex → Complicated: When you've documented decision rules
- Complicated → Clear: After all exceptions are handled
Implication: Your technology choice should evolve as the domain shifts. What starts as agentic AI exploration may become fine-tuned ML, then eventually become simple RPA.
Technology maturity considerations
Technology | Maturity | Best For | Risk Level |
RPA | Mature | Clear / Complicated | Low |
Classical ML | Mature | Complicated | Medium |
Fine-tuned LLMs | Maturing | Complicated | Medium |
Basic RAG | Maturing | Complicated | Medium |
LLMs (general) | Emerging | Complex/Chaotic | Medium-High |
Advanced RAG | Emerging | Complex | Medium-High |
Agentic AI | Early | Complex/Chaotic | High |
Questions to Guide Technology Selection
- Stability: How often do the rules change? (Daily → Chaotic; Never → Clear)
- Expertise: Can the best person in your company explain every decision in the system? (Yes → Clear / Complicated; No → Complex; No system in place to make decisions → Chaotic)
- Data: Is it structured and complete? (Yes → Clear/Complicated; No → Complex/Chaotic)
- Risk: What's the cost of a wrong decision? (High → Keep humans involved)
- Scale: How many decisions per day? (High volume + low complexity → Automate fully)
