AI work agents for small businesses · Built with Claude
Hand off the work.
Trust the result.
AXXOLAB is an early-stage Korean AI company that helps small businesses put AI agents into everyday work. Claude does the work, a reviewer agent checks the output, and a person approves any send, delete or payment action.
Problem
The hard part of delegating to AI is not getting an answer. It is knowing whether the answer can be trusted.
A chat assistant replies, but someone still has to check that the task followed the rules, what was read and changed, and whether anything left the building. A small business has no one to spare for that. We turn the checking into structure: rules kept in files, work and review split between different agents, and actions graded by risk.
The agents we build.
Work that follows the same criteria every time, yet takes hours because a person has to read and judge each item.
Email triage and reply drafts
Sorts incoming mail by importance and prepares drafts where a reply is needed. A person does the sending.
Reports and slide decks
Gathers material into a set format and drafts it. Anything that needs checking is left blank and flagged.
Meeting-note summaries
Takes a transcript and a template, organises decisions, action items and owners, and runs the result through review.
How it works.
One request runs through drafting, review and a verdict. The last decision stays with a person.
- 01 / Rules
Read the rules
Each run starts from an instruction file: role, forbidden actions, output format and completion checks.
- 02 / Work
Claude does the work
The main worker produces drafts and outputs. Unknown facts are left blank and flagged, not invented.
- 03 / Review
A reviewer agent checks
The reviewer is read-only. It can point out problems but cannot change the result.
- 04 / Decide
A person decides
Whether to apply review comments, and whether anything is sent out, needs human approval.
Built with Claude.
Claude is at the centre of the product. Here is exactly where it is used.
| Where | How |
|---|---|
| Main worker | Claude Code run non-interactively handles each request. The model can be chosen on screen. |
| Judge | Claude reads the review comments and decides whether a fix is needed. |
| Work rules | Role, forbidden actions and output format live in an instruction file read on every run. |
| Extensions | Features such as work memory are connected as Claude skills and invoked in plain language. |
| Development | This site was built with Claude Code, and we use Claude Code to develop the product. |
In the current build the reviewer runs on a different vendor's model, on the reasoning that a different model is less likely to pass the same mistake the worker made. How to evaluate reviewer agents properly is one of the open questions we want to work through, along with Korean-language output quality, the safety of human-approval steps, and latency and cost before scaling.
Alongside the agents.
Two things a company needs when it brings agents into its work.
Hands-on Claude workshops
Participants bring their own tasks and work through them with Claude, so the team can handle the tool before an agent is introduced. Details (Korean)
Testing AI-generated ad copy
We produce several versions of ad copy, images and video with AI and compare them. AI-generated material is labelled as such. Details (Korean)
Stage and roadmap.
Pre-launch. We separate what works from what does not yet.
| Phase | Scope | Status |
|---|---|---|
| 1 | Chat screen, two models connected, routing, automatic review, minutes desk | Working. 28 automated tests pass |
| 2 | Permission policy, approval screen, run and audit logs, failover | Designed, not yet built |
| 3 | File tools, email integration, shared skills | Planned |
| 4 | Remote access, authentication, usage and cost tracking | Planned |
We use the product on our own work while building it, and we are preparing our first client agents. There are no external customers or revenue yet. Our goal is to turn Claude workflows that have proven themselves into repeatable services for Korean small and mid-sized businesses.
Company.
AXXOLAB was founded in September 2026.
| Name | AXXOLAB |
| Founder | Hyunil Kim (김현일) |
| Founded | September 2026 |
| Legal form | Sole proprietorship, Republic of Korea |
| Location | Hanam, Gyeonggi-do, South Korea |
| What we do | Work agents built on Claude; Claude workshops; AI ad creative testing |
| Website | https://axxolab.com |
| hello@axxolab.com | |
| YouTube | @AXXOLAB |
The rest of this site is in Korean: product, case studies, notes, about. The case studies cover what we have built and run on our own work.
Get in touch.
We are preparing our first client agents. If you have a repetitive task you would like to hand to an agent, tell us about it and we will look at a small trial together.