Designing AI Agents for Small Business Owners

The Forgotten Middle

The AI industry has a blind spot. Billions flow into enterprise solutions (Salesforce Einstein, Microsoft Copilot) and consumer apps (ChatGPT, Perplexity). But small business owners — the restaurant owner, the local shop manager, the freelance contractor — are left with tools designed for someone else.

The problem isn't that AI can't help them. It's that AI products aren't designed for how they work.

Three Design Principles for SMB AI

1. No Dashboards. No Logins.

A restaurant owner works 14-hour days. They don't have time to log into a dashboard, interpret charts, and make data-driven decisions.

Our approach with sajang.ai: deliver insights through the messenger they already use. Every morning, a KakaoTalk message. One glance, one approval tap, done.

2. Agent, Not Tool

Traditional SaaS says: "Here's a powerful tool. Learn to use it."

Our approach: "Here's an AI employee. Tell it what you care about, and it handles the rest."

The difference is profound. A tool requires the user to have a mental model of how it works. An agent requires only that the user knows what they want.

3. The $1/day Price Point

Unit economics matter differently at the SMB level. A $99/month SaaS is a non-starter for a small restaurant. But $1/day — the price of a cup of coffee — is psychologically accessible.

At ₩29,700/month (~$22), sajang.ai is priced at roughly ₩990/day. Our COGS per user is ~₩4,000/month, giving us an 86% unit margin. This is only possible because LLM inference costs have fallen dramatically.

The Agent Architecture

sajang.ai runs on a multi-agent system:

• Revenue Agent (세리): Monitors sales data, detects anomalies, suggests marketing actions
• Review Agent (담당이): Watches customer reviews, generates contextual responses, defends ratings
• Marketing Agent (바이럴): Creates SNS content, proposes promotions, drives customer acquisition
• Supervisor (점장): Orchestrates all agents, prioritizes actions, reports to the owner

The owner interacts only with the Supervisor — one conversation thread, one source of truth.

What We've Learned

Building for SMB has taught us that simplicity is not a feature — it's the entire product. Every screen, every interaction, every notification must pass one test: "Would a busy restaurant owner at 11 PM understand this in 3 seconds?"

If the answer is no, we cut it.

This ruthless simplicity is what makes SMB AI hard to build — and why so few companies attempt it. But for the 6.5 million small businesses in Korea alone, it's exactly what's needed.