Architecture
Hybrid Automation Architecture
Learn why pure AI automation fails in production, and how to build reliable systems by combining AI intelligence with deterministic code triggers.
The Salesforce Case Study
In September 2025, Salesforce launched Agentforce - autonomous AI agents that could handle customer interactions, grant discounts, and modify orders without human oversight.
116 days later, they reversed course.
What Went Wrong
- October 2025: Customers discovered they could get 40% discounts by rephrasing requests. The AI interpreted "I'm a loyal customer" as sufficient justification.
- November 2025: A customer used an AI-granted discount as legal proof. Salesforce's lawyers discovered AI decisions are contractually binding.
- December 2025: Mandatory pivot to "Human-in-the-Loop" for all financially impactful actions.
Key Insight
AI decisions are legally binding. When an AI says "I'll give you 30% off", that creates an implicit contract your company must honor.
Understanding this distinction is fundamental to building reliable automations.
Deterministic Automation (Code/Triggers)
- Based on if/then rules written in code
- ALWAYS produces the same result for the same input
- Example:
if cart > $100 AND loyalCustomer then apply 10% discount - Testable, auditable, predictable
Probabilistic AI (LLM)
- Based on next-token prediction with statistical likelihood
- Can produce different outputs for similar inputs
- Affected by context, phrasing, temperature settings
- Creative, flexible, but unpredictable
The Golden Rule
AI to understand intent. Code to execute action. Never the reverse.
The "Hello World" Fallacy
AI demos are impressive. They handle "Where's my order?" beautifully. But demos show the 80% - the routine cases AI handles well.
The 20% edge cases represent 100% of your risk.
Routine vs Edge Cases
| Routine (80%) | Edge Case (20%) |
|---|---|
| "Where's my order?" | Lost order + VIP + wedding gift + overseas |
| "What's the price?" | B2B negotiated + promo stacking + currency conversion |
| "Return this product" | Personalized item + past deadline + emotional reason |
| Standard FAQ | Legal request + cultural context + medical emergency |
Testing Strategy
Test the 20%, not the 80%. Your QA should focus on edge cases, not the scenarios that already work in demos.
Hybrid architecture separates concerns into three distinct layers, each with clear responsibilities.
Layer 1: Intent Layer (AI)
AI excels at understanding what users want.
- Request classification (support, sales, complaint...)
- Entity extraction (product, amount, urgency, sentiment)
- Context understanding (history, tone, cultural nuances)
- Suggestion of possible actions
Example
User: "I've been a customer for 5 years and this is unacceptable!"
AI Output: { intent: "complaint", sentiment: "angry", vip_likely: true, escalation_recommended: true }
Layer 2: Action Layer (Deterministic)
Code executes actions within strict, auditable boundaries.
- Hard-coded business rules
- Non-negotiable limits (max discount: 15%)
- Pre-execution validation
- Automatic escalation when out of scope
Example
Rule: if discount_requested > MAX_DISCOUNT then escalate_to_human()
No AI can override this. It's code.
Layer 3: Override Layer (Human)
Humans monitor, correct, and handle exceptions.
- Real-time dashboard of AI decisions
- Alerts on detected edge cases
- Veto capability for any action
- Feedback loop for continuous improvement
Cart Abandonment Example
Let's compare pure AI vs hybrid architecture for cart abandonment emails.
Pure AI Approach (Risky)
"AI decides if it sends an email, what content, what discount to offer, and when."
Problem: AI might offer 50% to a $20 cart. Or nothing to a VIP.
Hybrid Approach (Reliable)
Step 1: Deterministic Trigger
Cart abandoned for 1 hour → trigger sequence (code-based, predictable)
Step 2: AI Personalization
Analyze customer history → suggest tone and complementary products
Step 3: Business Rules
if cart > $50 AND first_reminder then max_discount = 10%
Step 4: Human Escalation
If VIP detected → notify sales team for personal outreach
The Five Rules
- AI proposes, code disposes. AI can suggest. Only code applies with limits.
- Test the 20%, not the 80%. Focus QA on edge cases.
- Financial impact = validation. Discounts, refunds, credits need rules or humans.
- Escalation by default. When in doubt, hand off to human.
- Complete audit trail. Every AI decision must be traceable.
Ready to Build Reliable Automations?
Get a free audit of your current workflows. We'll identify risks and design a hybrid architecture for your business.
Request Free AuditReady to implement hybrid architecture?
Access all our courses with your 3A Automation client subscription.
Become a client