You decide what we can see and do.
Whatever we install, whether connected software, automations or AI, works inside limits you set, and a person stays in charge of the decisions that matter.
We ask only for the access the job needs.
Every connection gets least privilege, which means it can reach only what the work requires. A connection can only see information unless you’ve approved it to act. Access is granted from your own accounts and written down with the reason for it, and you can take it back at any time. Credentials stay in the platform’s secret storage and never go into an AI’s instructions.
Every action has one of three levels.
We set them with you.
-
Runs on its own
Routine tasks you’ve approved.
- Answer questions from your approved information
- Look up a quote or an order
- Draft a message for review
-
Asks first
Anything that reaches a customer, changes a record or involves money waits for a person to approve the exact action.
- Send a message to a customer
- Reschedule an appointment
- Change a customer record
-
Off limits
Anything you rule out stays ruled out.
- Issue a refund
- Delete a record
- Anything you mark off limits
The AI can suggest anything. Your permissions decide what actually happens, because a well-worded request is not the same as an authorized action.
Built to stop safely.
- Fixed rules where the answer is known. We use deterministic rules for anything with a known answer, AI where language and judgment help, and a person where the stakes are high.
- Information is not instruction. Text from an email or a website can inform the AI, but it is treated as information and never as permission.
- It stops when something is unclear. If a result can’t be confirmed, the work pauses and a named person is notified instead of the system guessing. This is called failing closed.
- The same step doesn’t run twice. Steps are built to be idempotent, so a retry never sends a second message or creates a second record.
- Any automation can be paused.
You can see what happened.
Where a system keeps a record, we show it to you: what happened, when, and who approved it. We also check the result in your own system, not just that a message was sent.
- Confirmation sent
- Approved by Dana
- Confirmed in your system
Before anything goes live we test the failure paths as well as the normal one: access denied, a step repeated, a connection dropped.
Bring your security questions to the first conversation.
We’ll answer them plainly, and we’ll write down what we agree.