Support automation
Repetitive questions are answered automatically, while sensitive requests reach a person sooner.
This is a Noafarin concept and is not presented as completed client work.
- Format
- Original Noafarin concept
- Business scenario
- Retail
- Year
- 2026
- Capabilities
- Process analysis، AI integration، Development، Measurement
Niloofar A. · VERIFIED
Invoice issue
Intent
Risk
Route
DECISION ENGINE
POLICY SET 7.2 · GDPR SAFE
The tax amount does not match our contract. Please review it before payment.
AUDIT LOG ON•PII REDACTED• POLICY 7.2
Interactive concept
Watch the system make a decision
Choose a request and follow intent detection, risk checking and routing. The demo deliberately shows when a person should take over.
My invoice is incorrect
Intent
Risk
Route
System decision
Ready to analyse
Only repetitive, low-risk work is automated. Sensitive requests always reach a person.
This is a usable concept prototype. Its data is illustrative and does not represent a real client or shipped product.
Challenge
Support teams usually answer the same five questions a hundred times a day: where is my order, how do I return it, what does shipping cost.
The experience is frustrating for customers and the support team. Its cost also rises in step with sales.
Discovery
Before any implementation, three months of conversation history is categorised. It nearly always turns out that a small number of topics make up the majority of volume.
The same analysis shows which answers belong inside the product so the customer never needs to ask for them.
Strategy
The goal is not to replace the support team. It is to let them spend time on the cases that genuinely need human judgement.
The boundary is defined explicitly: the automated agent answers informational questions only. Anything involving money, returns or complaints goes straight to a person.
Approach
The agent is connected to real order data, not to a generic knowledge base. "Your order shipped yesterday" is worth something; "it usually takes three to five days" is not.
Design
The conversation is designed in Persian, in a short tone without unnecessary formality. The route to a human is available at every step and is never hidden.
Technology
Answers are constrained to real data, and when the model is not confident it says so and escalates. That decision is the most important part of the project.
Every conversation is logged and reviewed so that errors surface and the boundaries can be tightened.
Launch
It goes live outside business hours first. The risk is low and real data accumulates.
Growth
Two numbers are tracked: the share of conversations resolved without a human, and customer satisfaction within those same conversations. The first is meaningless without the second.
Results
Proposed success measure
Because this is not a live client project, we do not invent results. On a real engagement, success measures are agreed before design and validated with live data after launch.