Sachi AI
Get updates

Your data stays where you choose, and every agent action is logged

We design our own products to the same rules we follow in client work. This page sets out what we do with your data and your systems when we build an AI agent for you, and what goes into the contract. If it does not answer a point in your security review, send us the question and we answer it in writing.

Only the people and tools that need your data can reach it

Where data stays
Where you choose: your servers, your cloud region, or our private GPUs. The agents, their records and any AI models we host run there. Where your rules say the data must stay in place, none of it leaves.
Who can reach it
Only the people on your project who need it for the task. They are named in the project plan and use accounts that you give and can take back. Access ends when the work ends.
Permissions for each tool
Every connection an agent uses has set limits: which system, which records, and whether it can only read or can also change them. An agent can only do what its role allows. Giving it more access is a change you approve.
Your data, your use
We do not use your data to train AI models, for you or for anyone else, and we do not reuse it outside your project.

A person approves any action that matters

Every agent starts in approval mode. It prepares the action, shows the evidence on the approval screen and waits. Nothing leaves the agent until a named approver says yes.

  • Before the agent goes live, we name, by role, who approves each kind of action.
  • Which actions need approval is agreed with you and written down. Relaxing this later is your decision, made in writing.
  • Every approval and rejection is recorded, with who made it and when.

Where your data may go decides which model we use

We are not tied to one model provider. There are three options, and the sensitivity of your data picks between them.

Your dataIts sensitivity decides where it may be processed.

  • Models from large providers

    Runs on: The provider’s business service

    Commercial models such as GPT, Claude and Gemini, used through the provider’s business service, on terms that do not allow training on your data.

    Fits when: Your rules allow this data to be processed outside your own systems.

  • Open models on our private GPUs

    Runs on: Private GPU servers we operate

    Open weight models, such as Llama, Qwen and Mistral, running on GPU servers we operate for you, behind one gateway with keys and quotas. Your requests and documents go to no model provider.

    Fits when: Data must not go to a model provider, but you do not want to run AI hardware yourself.

  • Open models on your own hardware

    Runs on: Servers you own, inside your network

    The same open weight models, installed on servers you own, inside your network or in your own cloud region.

    Fits when: Data must not leave your own network at all.

Naming a model family does not mean we have a partnership with its maker.

Every run is recorded, and we build to the rules that apply to you

What is recorded
Each time an agent runs: the request it received, the tools it used, the data it read, the action it proposed, who approved it and what happened. You can see why an agent did what it did.
How long it is kept
How long records are kept is agreed for each project and written into the contract, to match your own records policy. The records stay with the agent, where you chose to keep your data.
Personal data
Where an agent handles personal data, we process it only on your instructions, under a written agreement, and within the data protection law that applies to you. We help you meet your duties as the data controller.
Your rules and standards
At the start we list the laws, sector rules and security standards that apply to your system, and build to them. We hold no certifications and claim none. We build so your auditors can check the controls themselves.

Report a security issue to us by email

Write to us about a suspected incident, a security weakness or a security question. How quickly we tell you about an incident is set in each contract.

Other ways to reach us

Two kinds of client project put these controls into practice

Send us your security questionnaire

Tell us which rules apply to you and which systems an agent would touch. We reply to each point in writing.

Send your questions