AI-Agent-Driven Recommendations for Small Businesses

When I worked as a bartender at a neighborhood bar this past year, I learned firsthand how running a small business often means making tough decisions on the fly. Some decisions are simple, others feel like juggling knives. That’s where AI might be able to play a part—not to replace human judgment, but to surface observations that may have otherwise slipped by in the hustle and bustle.

In this project, I built a prototype AI-powered decision system to help small business owners turn daily operational metrics into actionable, prioritized recommendations. I tested it on a bakery example—my brother’s business, Hayes Baked Goods—but the approach could extend to restaurants, cafés, or other food and beverage operations.

GitHub repository: https://github.com/k80hys/ai-recommendations-smallbusiness

Business Context

Hayes Baked Goods is a small bakery that sells several different kinds of cookies. The owner tracks daily sales, staffing, inventory, and other factors—but making decisions from these numbers can be overwhelming.

Challenges for small businesses include:

• Tracking and interpreting daily sales, inventory, and staffing numbers

• Prioritizing actions amid fluctuating demand

• Balancing risk, staffing, and customer satisfaction

The goal of this project was to demonstrate a structured, AI-native workflow to turn these signals into daily recommendations that are actionable, interpretable, and confidence-weighted.

Project Overview

This system uses a three-agent architecture to separate business reasoning into clear stages:

1. Analyst Agent – Converts raw metrics into structured observations and signals

2. Decision Agent – Transforms observations into actionable recommendations

3. Reviewer Agent – Evaluates recommendations, providing confidence scores and contextual notes

At the end of the workflow, the system produces a Daily Decision Brief that summarizes prioritized actions for the business. The aim is to make the output digestible for non-technical staff while preserving reliability and interpretability.

How It Works

The system runs on daily operational metrics stored in data/daily_snapshot.json. Example inputs include:

• Sales, traffic, and inventory counts

• Business configuration like staffing, operating hours, and risk preferences

• Industry-specific assumptions and priorities

From there:

1. Analyst Agent converts numbers into structured signals, like "low inventory risk" or "high revenue growth."

2. Decision Agent generates actions, such as restocking certain ingredients or adjusting production focus.

3. Reviewer Agent assigns confidence levels and adds contextual notes—for example, reminding staff to consider weather or seasonal trends.

The output is both human-readable and machine-consumable, allowing for easy review, automation, or integration with business systems.

Example output:

{
  "date": "2026-02-08",
  "items": [
    {
	  "action": "Adjust inventory levels for high-demand items",
      "rationale": "Sales data shows inventory depletion faster than historical average",
      "confidence": "high",
      "notes": "Consider weather impact on foot traffic"
    }
  ]
}

Console output is similarly concise, with clear reasoning, confidence, and notes for each recommendation.

Highlights

• Stepwise reasoning: Observations → Actions → Confidence

• Guardrails: Schema validation ensures input and signal reliability

• Human-in-the-loop ready: Recommendations are interpretable and reversible

• Customizable by industry: Industry profiles allow tailoring signals, priorities, and assumptions, because we all know different food and beverage industries vary vastly

• LLM integration-ready: Swap the mock model for OpenAI, Anthropic, or local LLMs

What’s Next

This project is a prototype, but the roadmap is clear:

• Connect to real POS systems like Square for live metrics

• Integrate a full LLM for more nuanced recommendations

• Extend to other industries and business types

• Add event-driven automation to run the agent daily

Final Thoughts

This recommendation engine was created with real small businesses in mind. Whether it's an East Side bar like the one I worked at, or a cookie bakery like my brother's, I wanted to bring my real-life intuition about business priority and put it up against AI agent capability.