ANGELA JIAN
LOADING
Back to Portfolio
AI Process Design Human-in-the-Loop

Human-in-the-Loop
AI Accounting Process Design

Role: Product Designer & AI Strategist

Overview

Problem

Traditional manual bookkeeping takes 40+ hours per month, has high error rates, and accountants spend most of their time on repetitive categorization tasks.

Goal

Design a semi-automated AI workflow that dramatically reduces bookkeeping time while maintaining accuracy.

Results

60% reduction in bookkeeping time, 97% classification accuracy, and 75% fewer manual reviews.

-60%
Bookkeeping Time
97%
Accuracy
-75%
Manual Reviews
Design Thinking

Design Thinking

Core Insight

"Not every step needs AI — the key is finding the optimal split point for human-machine collaboration."

Confidence Threshold Mechanism

We route classification results into three processing paths based on the AI model's confidence score, balancing efficiency with accuracy:

> 95%
Auto-Approve

High-confidence results pass automatically, no human intervention needed

70–95%
Human Review

Medium-confidence results are quickly reviewed and confirmed by accountants

< 70%
Manual Processing

Low-confidence results are handed off for full manual processing

Process Architecture

Process Architecture

Step 1
Raw Transaction Data Input
CSV / Excel / API Import
Step 2
AI Auto-Classification + Confidence Score
GPT model analyzes each transaction
Step 3 — Three-Way Routing
> 95%
Auto-Approve
70-95%
Human Review
< 70%
Manual Processing
Step 4
Feedback Loop to Model
Continuous learning, ever-improving accuracy
Key Decisions

Key Design Decisions

Progressive Trust

We started with a high threshold (99%) and gradually lowered it as model accuracy was validated, letting the system grow through earned trust.

Fallback First

The experience when AI makes mistakes matters more than when it's correct. Every automated step has clear fallback mechanisms and human intervention points.

Data Flywheel

Every human review becomes training data for model improvement. The longer the system is used, the less human intervention is needed, creating a positive feedback loop.

Tech Stack

Tech Stack

Python OpenAI API Pandas Streamlit PostgreSQL
Reflection

Reflections & Learnings

"The core of AI product design isn't pursuing 100% automation — it's finding the balance point where humans and machines each deliver maximum value. True wisdom is knowing when to let AI decide, and when to hand the decision back to humans."

This project taught me deeply that the best AI products don't replace humans — they amplify human judgment. When we position AI as an "assistant" rather than a "replacement," user acceptance and trust increase dramatically.

Back to Portfolio