The starting point: each job began as a message and a photo in a LINE group
Every order started as a brief typed by sales, with a photo, into a group chat, then flowed on to pricing, production and delivery with no single place that said where a job stood. To know the status, the owner asked in the group or scrolled back looking for the photo.
Details
- A small team, each person holding one slice: briefs, prices, invoices, shipping
- Knowing which job was which lived in people's memory, not in a file
- Nobody was at fault; the tools in use were never designed to track work
What the client first asked for: a bot that reads the LINE group and reports status
The first request was a bot sitting in the existing group chat, reading briefs and answering where each job was. We did not start building; we asked to walk the whole chain first, from brief to delivery.
Details
- The need behind the request was right; where the bot should live was still a question
- We spent two sessions walking the full pipeline instead of writing code
What field discovery revealed: the team finds work by picture, not by number
Asked how they retrieve an old job, the answer was: by the photo, not by searching a number or a name. A system without images would lose to scrolling the chat from day one, and the LINE group in use was not a place a bot should be reading.
Details
- The image gallery moved from out of scope to the main view
- Brief intake moved from the group chat to a chat panel inside the system: paste the same text and photo, AI extracts the fields
- LINE stays, but as a one-way channel for the digest, not the place where work arrives
- Customer
- Seabreeze Café
- Product
- Insulated cup · 1-colour print
- Quantity
- 500 pcs
- Budget per piece
- 60 THB
- Delivery date
- missing · please add
The system boundary: AI proposes, a person confirms, at three clear points
AI reads a brief and proposes the job data, but a person confirms before a job is created. The price offered to a customer is always set by a person, and drafted messages to partners are never sent without a human check.
Details
- Job status is computed from the data, such as whether a price exists or a due date has passed; nobody updates a status by hand
- A brief with no delivery date is flagged, never guessed
- Who sees what is a field-level policy; customer details are visible only to those who should see them
- The team adds the system's LINE OA to receive a morning and afternoon digest; the bot does not answer in the group
Dashed lines in the pictures = where a person decides
How it was tested: the real team doing real work on the system, not watching our demo
The acceptance bar was that the real team tries the system before accepting the handover, and that the status on the web always matches the LINE digest. There are no outcome figures on this page because we have not measured them with the client.
Details
- A working trial first, then acceptance; never acceptance from slides
- What was checked: a new brief gets a unique job number, status changes the moment a price is entered, overdue jobs appear on the web and in the digest
- What we do not claim: time saved, sales or job counts, because no agreed measurement exists
What stays outside Phase 1, on purpose
Phase 1 covers briefs, status, the image gallery and the daily digest. Document and finance work is not part of Phase 1, but the database is designed so it can be added later without a rebuild.
Details
- The client decided this boundary after seeing a clear in/out table
- The rest is a next block to take or leave, not unfinished work
The transferable lesson: start from how the team already finds its work
If the team finds work by picture, the system must start with pictures. If work starts in chat, the intake must accept chat. And the best status is one nobody has to type.
Details
- Looking at the real work before agreeing the scope is always cheaper than fixing it later
- AI takes the repetitive part: read, extract, draft, summarise. People take the accountable part: confirm, price, send
- Start at the one point the team repeats many times a day, then expand
Does your work start in chat too?
This case is the AI Agent / internal operating system card. Tell us where the work gets stuck and we look at the real work before proposing a scope.