Put AI inside the processnot in a chat window
A message arrives. It checks the roster, the stock, the drawing — then writes the reply and leaves it for you to read. You press send. Only then does it send.
Runs in our own facility in Taiwan, on your own servers, or on a GPU that serves nobody but you. Your call.
- It can look, not touch
- It asks before it changes anything
- Every step leaves a record
What OrbanAI is, in one paragraph
OrbanAI is an AI agent platform built for organisations. It is not another chatbot — it connects to the systems you already run, starts read-only, and does the work: looking things up, cross-checking, compiling, drafting, and reading drawings and photographs. Every answer points back to its source, every action leaves an auditable record, and nothing is written back to your systems without a person approving it first. You decide where it runs: our Taiwan nodes, a dedicated GPU, your own server room, or a hybrid. Data is isolated per organisation and is never used to train a shared model.
Three steps into your process
Connect
It connects to what you already run: ERP, MES, rosters, booking back ends, file servers. To begin with it can only look — not one character of your systems changes.
Hand it the work
It looks things up, cross-checks them, and writes the reply or the quote for you. Drawings, spec tables and shop-floor photos included.
It moves when you say so
Every action leaves a record of who did what and when. You can look it up afterwards, and you can undo it.
The reply waits here until you send it
Two real situations. The agent reads the systems and writes the answer — but it cannot send. That part is yours.
Ms. Wang
Official messaging account
You decide where it runs
Four options, one agent. Who can see what and who has to approve what is identical across all four. Changing machines does not mean changing product.
Our Taiwan nodes
Runs in our own facility in Taiwan. Start today, own no hardware.
A dedicated GPU
One GPU, yours alone. Buy it or let us host it.
Your own server room
Installed on your machines. Data never leaves your network.
Hybrid
Sensitive data stays on site; borrow our nodes at peak.
The shape of one real engagement
A service business that runs on bookings, changes and rostering. This is the order things actually happened in — not the timeline from a proposal.
First
Process review and assessment
We watched how bookings arrived, how changes were handled, how the roster got built. We found the step costing the most hours and left everything else alone.
Connect
Connected to their existing systems
The agent went into the app they already had, carrying the signed-in user's own identity — what it can see is decided by their existing permissions, not by us.
Live
Live in production
Three systems and their real-time message flow, connected. Deliberately no write tools: anything that changes something sends the person back into their own app to confirm it.
Since
Ongoing tuning and operations
After launch we added an input guard and an intent check for the edge cases you only meet once it is really running. A rollout does not end at handover.
Quoted by scope and by where it runs
No price list, because no two customers connect the same systems or run on the same machines. You get a firm number after we have talked.
One process
Connect the single process that costs you the most hours.
- One process, one department
- One existing system, read-only
- Runs on our Taiwan nodes
- Full audit record
- An engineer of ours stays with it through rollout
For teams that want proof before scope.
Talk about thisOperations team
Booking, rostering, dispatch and customer messages, connected end to end.
- Several processes, shared across departments
- Connects booking, roster and dispatch back ends
- Our Taiwan nodes or a dedicated GPU
- Set who can see what, and who has to approve what
- Document and image recognition
Where service businesses, chains, clinics and fleets usually start.
Talk about thisEnterprise
Data never leaves the plant; the model runs on hardware you choose.
- Connects ERP, MES, drawings and specification libraries
- Reads drawings, compares specs, checks QA records
- Dedicated GPU, your own server room, or hybrid
- SSO, audit export, and a retention policy you set
- Named contact and a service level agreement
- A process walkthrough on site before anything is built
- NDA and data processing agreement available
Precision manufacturers, job shops, and anyone whose data cannot leave the building.
Talk about thisThe five questions we hear in every conversation
They come up in the first ten minutes, every time. The answers are written down here.
- Our IT team is one or two people. Who does the work?
- Our engineers do. We walk the process, connect the systems and stay with it through rollout; your IT grants access and checks that what we connected is right. It is why we start with a single process — we can finish that, so we can promise it.
- Our data cannot leave the plant. Does this still work?
- Yes. The software installs on your own servers and neither your data nor the model leaves your network. Hybrid works too: sensitive work stays on site and peak load borrows our nodes. The agent, the permissions and the audit trail are identical across all four deployment options.
- Could it change something in our systems by mistake?
- It cannot. In the first phase it holds read access only, so it has no write capability at all. Write access is granted later one action at a time, each behind an approval rule you set — it writes what it intends to do and leaves it there until you say yes.
- We tried AI before and it failed. What is different?
- Most failures come from handing staff a second system to learn, or from something that answers fluently but is not connected to the real data. Our agent lives inside the interface they already use, every answer links back to the screen it came from, and when it cannot answer it says so rather than inventing one.
- How long before we see anything?
- One process first, usually running within weeks. Prove one works, then widen — every rollout we have seen succeed went in that order.
Thirty minutes, three answersWhich process first, what it connects to, how soon it runs
Online or at your site. Nothing to prepare.