Human-in-the-loop (HITL) is a model for AI to work, while humans supervise and authorize important steps. AI processes most quickly, making decisions when there are risks, sensitivities or strange situations. Both fast and safe. Zenify recorded 98.4% AI accuracy and 34% cost reduction.
TL;DR
- HITL: AI doer, custodian and licensor.
- People who decide on risky, sensitive, strange situations.
- Neither "free AI" nor "useless AI".
- Balance speed and safety.
- Especially important in refunds and contracts.
What is Human-in-the-loop?
Human-in-the-loop brings humans into the AI process right at the decision point. Participant levels:
Level | AI does | Maker
Automatic | Handled all | Periodic monitoring
Reviewer | Recommended | Browse/decide
Transfer | Recognize | Direct processing
Freedom | Shouldn't | No
When do you need decision-makers in customer care?
Situation | Why do we need people
Big Cashback | Financial risks
Contracts, legal | Responsibility
Customers are extremely upset | Empathize, soothe (see Complaint)
VIP guests | Personalized, long-term relationships
Situations AI doesn't understand | Judgment outside the rules
Sensitive information | Security, compliance (see AI Security)
Why is HITL needed instead of free AI?
1. Reduce risk
Wrong refund AI, wrong contract = loss of money, loss of customers.
The reviewer is at a risky point, safe but still fast.
2. Increase customer acceptance
Customers feel more secure when important matters are decided by someone (see AI Agent vs human).
3. AI learns from humans
People fix, AI learns, HITL is the source of improving AI (see What to do when AI goes wrong).
4. Comply with regulations
Many industries require people to be ultimately responsible (insurance, finance, healthcare).
How to design HITL properly
List risk points: How much refund needs to be approved, what type of contract.
Threshold configuration: AI self-determines below the threshold, transfers approval above the threshold.
Push jobs to the right people: AI Routing pushes jobs to the right person (see AI Routing).
Include context: reviewer sees summary + reason for recommendation (see AI summary).
Measuring browsing time: the bottleneck is often waiting for approval, which needs to be optimized.
Sample HITL model in customer care
Jobs | AI | People
Answers to frequently asked questions | Self-treatment | No
Change address | Self-updating | Monitoring
Refund < 1 million | Recommended | Browse
Big Cashback | Summary | Processing
Contract | Draft | Sign