Letting AI Trigger Actions Without Human-in-the-Loop Approval
“AI should suggest. Humans should approve. Not the other way around.”
What Happened
Set up an AI-powered Flow that analyzed Case descriptions and auto-escalated to legal if it detected lawsuit-related language. No human review step. A customer wrote 'this product is so good it should be criminal' and the AI flagged it as legal risk. Auto-created a Legal Case record, emailed the legal team, froze the customer's account, and sent a 'we take your concerns seriously' template email. To a happy customer. Who then actually did contact a lawyer — about our bizarre response to their compliment.
The Wrong Way
// Flow: AI Case Escalation (fully automated, no review)
Record-Triggered Flow: Case - After Save
→ Action: Prompt Template "Analyze Case Sentiment"
→ Prompt: "Does this case mention legal action? Reply YES or NO."
→ Input: {!$Record.Description}
→ Decision: If AI_Response contains "YES"
→ Create Record: Legal_Case__c (auto-created, no review)
→ Email Alert: Legal team notified
→ Update: Account.Status__c = "Under Review" (account frozen)
→ Email Alert: Customer "we take your concerns seriously"
// All automated. No human checks if the AI was right.The Right Way
// Flow: AI Case Escalation (human-in-the-loop)
Record-Triggered Flow: Case - After Save
→ Action: Prompt Template "Analyze Case Sentiment"
→ Input: {!$Record.Description}
→ Output: Risk score (1-10) + reasoning
→ Decision: If Risk_Score >= 7
→ Create Task: "Review AI Escalation Flag"
→ Assigned To: Case Owner's Manager
→ Description: "AI flagged this case for potential legal risk.
AI Reasoning: {!AI_Reasoning}
Score: {!Risk_Score}/10
REVIEW and decide: Escalate to Legal (Yes/No)"
→ Due Date: TODAY
→ Priority: High
// Manager reviews → manually escalates if warranted
// Screen Flow for manager: "Confirm Escalation" with all details
// ONLY manager action creates Legal_Case__c and notificationsThe Lesson
AI should flag, recommend, and draft. Humans should review, approve, and execute. Never let AI take irreversible actions without human approval — especially anything customer-facing.
Enjoyed this? Get more like it.
Glen's Musings — AI, investing, and building things. Occasional. Free.
More AI & Agents Mistakes
Turning On Einstein Prediction Builder with Dirty Data
AI trained on garbage data gives you garbage predictions with confidence scores.
Read moreCareer-EndingDeploying Einstein Copilot Without Testing Its Guardrails
An AI assistant with full org access and no guardrails is a data breach waiting to happen.
Read moreAnnoyingBuilding Prompt Templates Without Grounding in Record Data
An ungrounded prompt hallucinates. A grounded prompt uses your actual Salesforce data.
Read more