Control and safety
The AI assistant speaks to your customers on your behalf. That is why you control it: rules, limits, a log and a switch.
Why it matters
AI can be inaccurate. “Watch for mistakes” is not enough when a message goes to a real customer. So between the assistant and the outside world sits a control layer called Agent OS. The assistant prepares a reply, and the layer enforces the rules before anything is sent.
Rules the system enforces
- Personal data
- Before text reaches the AI model we mask e-mail addresses, phone numbers, national ID numbers, bank account numbers, addresses and full names. If anything still looks like personal data, the model call is blocked.
- Prices and promises
- A message containing a price, a quote or a promise of a subsidy does not go to the customer. The customer gets a safe, general reply, and you see the reason in the panel.
- Customer consent
- A message goes to the customer only if their consent is recorded and they have not opted out.
- Sending hours
- Messages to customers go out only within a set window (by default 8:00 to 20:00). Outside it they wait until morning.
- AI costs
- Daily and monthly limits in zloty, separately for each agent and each customer. An alert at 80% of the limit, a block at 100%.
You decide how much the assistant may do
- Observation
- The system only records in the log what it would do. It blocks nothing. We start here.
- Approval
- The assistant prepares a message and you approve it with one click: “Approve and send” or “Reject”.
- Automatic
- For selected kinds of messages you can turn on sending without approval. Rules, limits and switches still apply.
When a human steps in
The AI assistant carries the conversation with the customer, and every message it sends is marked as AI. The note “from now on an employee is writing to you” goes out only when someone at your company really takes over: they click “Taking over” in the notification e-mail and start writing to the customer instead of the assistant. It is sent once, in the customer’s language.
Log and emergency switch
- Every action of the assistant is recorded in a log where entries are linked by a hash chain. Changing or deleting an entry is detectable.
- Emergency switch in the panel: turn off everything, one agent, or all agents of one customer. It works immediately. New inquiries wait and are handled once it is turned back on.
- The panel is password-protected. You see what awaits approval, what was spent and what the system did.
What the panel looks like (illustration)
Awaiting approval
Reply to a customer in German. Priority: hot.
Switches
Illustration with no real data. The panel runs in the browser, after login.
Where this stands
The control layer has passed automated tests. In a pilot we start in observation mode and check it with you before turning on approval or automatic sending. Rules about consent and personal data are worth discussing with a lawyer.