Automation / Service
AI business automation, with you as the approval gate.
Hand over a task your team does by hand every week. You get back a workflow that runs it, a log of every run and a short runbook. The agents do the slow parts; anything that sends, publishes, pays or deletes waits for your OK.
Who it's for
Small teams with one job that eats an afternoon every week.
The best first project is boring, repeated and easy to check: you already know what a good result looks like, and you can say who signs off before it leaves the building.
A good fit
- The work lives behind a login. A portal, dashboard or supplier site with no export, where someone copies data out by hand.
- It moves between tools. A form, a ClickUp board and a spreadsheet that a person keeps in sync, or an n8n flow that needs a human check in the middle.
- There is no API. Selectors keep breaking, so it needs an agent that works the screen and knows when to stop and ask.
- Content goes out on a schedule. Research, drafts, images and publishing across one or more sites, with review rules you can write down.
Not a fit
- Fully hands-off sending or paying. The approval gate is the point of this service, so I won't remove it.
- A job nobody can describe yet. If "done" can't be written in a sentence, the first step is a scoped plan rather than an automation.
- Shared passwords in chat. Logins go through a safe route agreed during planning, never pasted into a message or stored in notes.
Hand over, get back
What you send, and what you can run yourself afterwards.
Every engagement ends with something you can run and check without me: the workflow, its log and a short runbook.
- You hand over
- The task as you do it today, the tools and sites it touches, a safe way to log in, what "done" looks like, and the person who approves anything that leaves your business.
- The workflow
- A browser run, an n8n or ClickUp flow, or a computer-use agent, depending on what the task touches. Retries and error handling are built in, not added later.
- The log
- Every run is recorded, with screenshots for browser and screen work, so you can see what happened without asking.
- The approval step
- A checkpoint that pauses before anything goes out and asks the named approver, with every decision logged.
- The runbook
- A short document: how to start it, what each alert means, and when the agent is meant to stop and ask instead of guessing.
How an engagement runs
Two sign-offs: one on the plan, one on every send.
Nothing is built until you approve a written plan, and nothing the workflow produces leaves your business until the approver says yes.
Brief
Send the task on Telegram, WhatsApp or email, with the tools and logins it touches.
Written plan
I write down each step, what counts as done, and which steps need your approval. You sign off before I build.
Build and test
Agents do the work while every run is logged, with screenshots, so you can compare the output with how you'd do it by hand.
Approval gate
Sends, posts, payments and deletions pause and wait for the approver's OK. Every decision is logged.
Handover
You get the workflow, its log and the runbook, and the gate stays on after I step back.
The slow, reversible parts
Reading, typing, navigating, collecting, drafting and retrying. If any of it goes wrong, nothing has left your business yet.
Anything you can't take back
Saving over records, sending, submitting, publishing, paying, deleting and anything that touches a password. The run pauses, asks, and records your answer.
Worked example
My own publishing pipeline, across seven sites.
No client names here, so this is my own system. It is the same shape I set up for clients at a smaller scale: a weekly manual job, broken into steps that run alone and steps that wait.
It ran on Hermes agents with GPT models until autumn 2026. When I moved everything back to Claude in October, its projects moved to Claude Code.
-
Handed over
Sites, topics and review rules
Which sites, what each one covers, and the rules a draft has to pass before anyone reads it.
-
Runs on its own
Research, drafting and images
Agents gather sources, write the draft and make the images. Nothing is public yet, so this part needs no sign-off.
-
Gate
Checks before anything goes live
On my own sites I'm the client, so I signed off the release rules up front and a release has to pass its checks. On a client's pipeline, going live also waits for the client's OK, like any other send.
-
After publishing
Live checks, backups and rollback
The live page is checked after release, with a backup taken first and a rollback path ready.
-
When a check fails
It drops back to review-only
A release that fails a check stops publishing and waits for a person. The pipeline doesn't try to push through on its own.
- AI agent automation: what to automate first
How to pick the first task, before talking about tools.
- Approval-gated AI automation for small business
How the approval gate works and where the checkpoints go.
- n8n error handling: a small-business runbook
What a runbook covers when a flow fails at 2 a.m.
Contact
Send the task you did by hand this week.
Describe it as you'd explain it to a new colleague: the tools, the login, what done looks like and who signs off. Telegram or WhatsApp is fastest.
- Telegram@nguyenlnp
- WhatsApp / phone+84 914 859 234
- Email[email protected]