Written by Aviad Faruz
The wrong way to choose a first AI project is to ask which task looks most impressive in a demo. A better question is: which repetitive decision can you inspect quickly when the system gets it wrong?
I run FARUZO, another jewelry brand and three event venues, and I build many of the workflows behind them myself. The projects that have lasted share three traits. They solve a narrow problem, rely on clean source data, and leave a clear path for human review.
Begin with a decision, not a department
“Automate customer service” is too broad. “Classify an incoming venue inquiry and select the right reviewed reply template” is specific enough to test.
For one of my venues, I built an n8n workflow on respond.io that sorts incoming WhatsApp questions. It classifies the inquiry, then selects a reply template from a Google Sheet. That boundary matters. The system does not invent a new answer every time. It chooses among responses we have already reviewed.
Today, the workflow handles 63% of incoming inquiries automatically. That figure measures a real operational result. The remaining inquiries still go to a person, especially when a request is unusual or needs judgment.
Use note mode before send mode
The same workflow did not start by sending messages to customers. It first ran in note mode, where I could see the draft and classification without letting the system act.
That stage exposed the real exceptions. A message could look like a standard availability request but include a special event detail that changed the answer. Reviewing those misses helped me adjust the categories and templates before the workflow could affect a customer.
This pattern works well for many small businesses. Let the automation observe, classify, or draft first. Give it permission to act only after its mistakes become predictable.
Fix the source data before adding intelligence
Another lasting automation began with a spreadsheet problem, not an AI problem. The same jewelry item had one code in a supplier file, another in a shop listing, and a third in my own records. Price updates reached some listings and missed others.
I built a map linking about 1,130 internal product codes to their supplier codes. One automation now pushes price changes from that single list, with no manual editing. The important part was settling product identity first.
If a business has conflicting customer names, product codes, prices, or status labels, automation will repeat the conflict faster. Clean identifiers are often a better first investment than a smarter model.
Keep subjective and brand-visible decisions human
I also use AI to create alternative backgrounds, lighting treatments, and lifestyle versions of jewelry photographs. Generating several options saves time, but none is published automatically. A person compares every image with the real piece, including its engraving and proportions.
That final check stays human because a visually attractive mistake is still a product mistake. The model can generate options. It should not decide what is true about the item.
A five-question test
Before building a workflow, ask:
- Is the input consistent enough to classify?
- Can the output be checked in under a minute?
- What is the cost of one wrong action?
- Can uncertain cases be sent to a person with the full context?
- Can the workflow run in observation or draft mode first?
If those answers are clear, the project is probably narrow enough to begin. If they are vague, reduce the scope until mistakes are easy to see and recover from.
The best first automation is rarely the flashiest one. It removes a repeated decision while making exceptions more visible. That is how a small business gains speed without giving up control.
Author Bio:
Aviad Faruz is the CEO of FARUZO. He builds and operates automation for his e-commerce and hospitality businesses.