Security
An agent will work with your business data
So here is the whole model in plain words: where the system runs, what it may access, what it records and what it cannot do. Your security officer can ask any question before the first line of code is written.
Four principles of our approach
Not promises of good behaviour, but the way the system is built.
-
01
You decide where it runs
The system works in your infrastructure or in your cloud account. Our standard option uses cloud AI models through a private gateway. A closed network with a local model is an additional option that we assess individually: it needs your own hardware, and local models are usually less capable than the best cloud ones.
-
02
Personal data is hidden before sending
Before a request goes to an AI model, personal data is replaced with placeholders, and after the answer it is restored. The key to the placeholders stays with you and the model never sees it.
-
03
Every decision can be traced
The input data, the records used, the rule applied, the result, the person who approved it and the recorded value are all saved for as long as you decide. When an auditor asks why the system did something in March, there is an answer.
-
04
The system belongs to you
The code, the instructions for the agents, the settings, the test sets and the logs are in your repository from day one. If we stop working together, the system keeps running and another team can take it over. We hand over instructions for operating it, not a dependency.
A person stays in control of important decisions
For companies with strict requirements, an AI agent must never decide alone and unseen. We build it around three rules.
-
The agent prepares, the person decides
The agent collects the data, checks it against your rules and proposes a result with the reasons. Until a responsible person confirms it, nothing is written into your systems.
-
One screen is enough to confirm
The person sees the proposed value, the place in the source it came from and the rule that was applied. If the agent is not sure, it says so and passes the case on.
-
Every step leaves a trace
Each proposal, confirmation and correction is recorded: who decided what, on which data and when. For an audit it is enough to open the record.
What the agent is allowed to do
Access is granted per type of operation, not for a whole system. An agent that reads orders cannot write to payroll, because it was never given a path there. This is visible in the settings, not hidden in an instruction.
| Stage | Reads | Writes |
|---|---|---|
| At the start the agent only suggests |
A copy of the data with limited access | Nothing |
| With confirmation after the accuracy is measured |
A copy of the data with limited access | Only after a person confirms |
| Independently for selected operations |
A copy of the data with limited access | Only for operations you approved |
| Always | Only within the granted access | Through the interface agreed with you |
One switch in the admin panel returns every operation to confirmation by a person, without any new deployment. We consider such a switch a mandatory part of a system like this.
What we promise not to do
The commitments that matter most are the ones about what we will not do.
-
We do not train models on your data
Your documents are only input for a request. We do not use them to train models, ours or anyone else's, and we check that the AI provider's terms exclude it as well.
-
We do not keep a copy of your data
Your data stays in your environment. We keep no copy of it on our side: no mirror, no backup and no analytical extract.
-
We do not give an agent unlimited right to write
The right to write is granted per type of operation, based on measured accuracy, and can be withdrawn in one click.
-
We do not hide how a result was obtained
Every value the agent produces has a reference to the record or page it came from. If it cannot be traced, it is not recorded.
-
We do not tie you to us
Ordinary code, standard ways of deployment, your repository. There is no proprietary component that you would have to keep paying for.
-
We do not touch what was not agreed
Access is given to each system and each operation as agreed during the audit, and your team grants it, not us.
Do you have a specific requirement?
An isolated network, a certain certification, a data processing agreement drafted by your lawyers: tell us before the audit, not after. It is cheaper for everyone.