Five Things to Set Up Before Choosing Tools for Small Team AI Automation

Starting with the question, "What AI tools do you use?" is already too late. Based on actual experiences assisting small teams, here are five things to set up before focusing on tools.

1. Start by Dividing AI into Two Types — Non-Developer vs. Developer

The first question teams new to AI automation face is, "Which one should we use?" However, before answering this, there's something important to know. AI tools are broadly divided into two categories.

One is the non-developer collaboration mode, like Claude Cowork. It's a sandbox structure that operates only within designated folders, minimizing the risk of major accidents even if mishandled. However, connecting to external data or systems requires more effort.

The other is developer tools like Claude Code and Codex. They have few restrictions, allowing for API integration and automatic publishing. However, the risks are higher if mismanaged.

For small teams, the ultimate goal is to move towards developer tools, as they significantly broaden the scope of automation. However, there's no need to start there immediately. If the team is still unfamiliar with AI tools, it's practical to first get accustomed to non-developer tools and transition gradually.

2. Ensure All Tools Only Access the "Same Folder"

This is a commonly overlooked aspect. When using multiple AI tools simultaneously, data can easily become scattered across different tools. This makes it increasingly difficult to determine which tool knows what information.

The principle is simple. Whether it's a local folder or a GitHub repository, consolidate data, skills, and agents in one place and ensure all tools focus on that single location. Even when mixing multiple tools, adhering to this principle significantly reduces confusion across the team.

3. Organizing and Documenting Requirements Determines the Outcome

This is a point surprisingly overlooked by many teams. The outcome of AI automation is much more influenced by "what was instructed" than by the tool's performance.

An effective flow observed in practice is as follows. First, articulate what you want to do. Then, ask the AI to "extract the requirements from what has been said so far." Verify each extracted requirement for accuracy. Finally, document this in a Markdown (MD) file.

Skipping this documentation process results in the AI understanding things differently each time, leading to varied outcomes. Conversely, following this process ensures consistent work even if team members or tools change.

4. Create a Golden Set (Answer Key) and Continuously Compare and Refine

Running AI automation once and improving over time are entirely different stories. The latter is achieved by maintaining an answer key, or a golden set.

The flow is as follows. First, create an answer key. Have the AI perform the same task. Compare the results with the answer key. Identify and train on the incorrect parts. Repeat this loop.

For example, in a Korean speech recognition (STT) task, you might use the results from Naver Clova API as the golden set for comparison. This loop becomes a unique asset for the team over time. While others can use the same tools, the same golden set and loop are exclusive to the team.

5. Validate Locally, Then Share with the Team

The final point is about the spread method. If the entire team starts simultaneously, it can actually slow things down. In practice, dividing into two tracks was effective.

One leading person first builds and validates locally. Once stabilized, they upload it to GitHub for team-wide access. Routine communication is managed through messengers like Slack, while further refining automation is separated into a dedicated agent channel, making management much easier.

Conclusion — The Core is the 'Loop'

While we've covered five things, the overarching principle is one: creating a loop that continuously learns and refines what works well. Tool selection is a subsequent issue.

If you're curious about where your team stands among these five stages, you can lightly check with a 2-minute diagnosis. Feel free to answer as much as you're comfortable with.

STAR-T AI Business Operation Diagnosis →

---

Footnotes

① Reference Materials: This article was written based on the author's (Jungkeun Park) direct AI automation mentoring experiences with small teams (1-5 members), without any external references. All specific company, industry, and personnel information has been anonymized according to the anonymity principle.

② Unused Metrics: Performance metrics like "N% reduction" or "N% reduction in work time" were intentionally omitted due to the lack of verified evidence. This article focuses solely on the methodology (what was set up and in what order).

③ Creation Method: The draft was written with AI (Claude), and then reviewed by a person for methodological accuracy and adherence to anonymity rules.