Contracts forbid it
Drawings and specifications held on behalf of a customer carry clauses banning disclosure to third parties. Sending them to a cloud AI is a breach on its own.
On-premise AI agents
Generative AI that never leaves your network.
Drawings, cost tables, supplier records, patient data. For organisations that have kept generative AI on hold because the data legally cannot go to a cloud service, we build AI agents that run entirely inside your network — architecture, access control and ongoing monitoring in one engagement.
Tell us your security requirements and we will tell you what is buildable.
Inside your closed network
File servers / core systems Drawings, costs, orders, quality records
AI agent — runs entirely inside the network Reading, verification, workflow execution
Firewall
✕ Nothing sent to an outside cloud✕ No calls to outside APIs✕ Nothing sent out as training data
“We can see the value. That data still cannot leave.”
For organisations that evaluated a cloud generative AI service and were stopped at the last step by their IT, security or audit function.
There is no outside connection
The plant network is physically separated from the internet. SaaS is not an option that exists.
Audit and internal approval will not clear it
“Your data is not used for training” does not clear internal audit in financial services, healthcare or the public sector. What is asked is where the data physically sits.
You do not have to start on-premise.
Pick by how sensitive the data is. It is the same product either way, so moving later is a configuration change, not a new project.
| Comparison | Ichiyomi Sheet | On-premise |
|---|---|---|
| Where the data sits | Our cloud, hosted in Japan | Inside your own network |
| Outbound communication | Yes | None |
| Internet connection | Required | Not required |
| Core system integration | Limited | Yes |
| Initial cost | None | Quoted per project |
| Running cost | JPY 3,000 per user / month | Annual support contract |
| Suits | Teams that want to try it first | Data that cannot leave the building |
Build and annual support for the on-premise configuration are quoted per project. Three weeks of verification fixes the total.
A closed network is not the same as a safe one.
Misreads happen inside a closed network too
Closing the network does not stop an AI from reading a value wrongly. We trace every extracted value back to where it came from in the source document, and values without that evidence never reach the next step.
Who can see what is a design decision
Keeping everything in house is exactly why access design matters. We scope which documents each department and grade can reach, and keep an operation log.
The operating model is part of the build
On-premise systems tend to be installed and then left alone. We design the accuracy monitoring and the update cycle that keep it working after go-live.
What you keep in house stays in house
Rights to data processed inside your network remain yours, and we never use it for training. You get the operating documentation, so day-to-day running sits with your own team.
Find out whether on-premise is realistic for your requirements
Tell us the security requirements and the work you want covered. We will come back with an architecture and an indicative cost — or tell you it is not worth doing.
Book a 30-minute call →Three weeks to a fixed scope, price and schedule.
- Day 0
Free 30-minute call, online
We go through your security requirements, the work in scope and your existing setup. If a cloud or PC-resident option would be enough, we will say so.
- Week 1–3
Verification
We verify accuracy on your real documents and settle the architecture, network design, access model and operating model. You receive a firm quotation for the build. If you proceed, this fee is credited against it. Everything runs online; on-site is available if you prefer it.
- Month 2–4
Environment setup and pilot
Environment setup, then a pilot on one workflow. Integration with your existing core systems is confirmed at this stage.
- Month 5 –
Production and accuracy monitoring
We widen the scope of work covered while monitoring accuracy and uptime. Annual support includes keeping pace with updates.
What buyers ask before committing
Is it less accurate than a cloud AI?
It depends on the task. For open-ended conversation and creative writing, cloud services are ahead. For business tasks — reading forms, searching internal documents, running routine processes — a configuration tuned to your documents and your workflow often does better. During verification we show you the comparison on your own data.
Does it work with no internet connection?
Yes. All processing happens inside your network, so day-to-day use needs no outside connection. Software maintenance and updates need either a temporary connection or delivery on physical media. We can design the procedure for a fully air-gapped environment.
Can we use our existing infrastructure?
It depends on what you have. During the free call and verification we review your existing setup and tell you what carries over and what has to be added. Starting small, with the scope limited to one workflow, is a normal way in.
Start from “our data cannot leave our network”.
We will tell you what can be built against your requirements, and where the limits are. If on-premise is not necessary, we will say that too.
A 30-minute call
Bring your security requirements and the work in scope; we will sketch what is buildable on the call.
Talk to us →Not sure on-premise is needed
If the work can start with data that is allowed to leave your network, the cloud product is the lighter way in. It is the same product, so moving later is a configuration change.
Ichiyomi Sheet →You need something purpose-built
If it involves integration with existing systems or a workflow specific to your organisation, talk to us about custom development.
Custom AI agent development →