Security
How the assistants are kept in bounds
AI agent security for a logistics company means the assistant can only read what it was given, can only do what is on its permission list, cannot take an external action without a person, and leaves a record of everything it did. TYR-X runs every assistant that way, inside its own forwarding business first, and writes the limits into the agreement before the assistant runs.
CAN
- read
- extract
- match
- draft
- organise
- flag
CANNOT
- send
- pay
- file
- price
- commit
- contact
What can the assistant reach?
Only the sources you approve at onboarding: a shared mailbox or folder, exported job files, your rate sheets, your templates. Not your private inbox. Not your CRM or TMS credentials in the first workflow. Not your accounting system. If a source is not on the list, the assistant cannot see it. If a brief contains confidential material, it is not ingested; I ask for the permitted inputs instead.
What can the assistant do on its own?
Read, extract, match, draft, organise and flag. That is the whole list. Sending, paying, filing, pricing, committing and contacting a customer, carrier or vendor are not permissions it has, so they cannot be granted by a clever prompt, a customer request or a mistake. Every external action is a person's action.
How do you know what it did?
Every packet lists what was produced, what is pending, what was blocked and why. Every figure carries its source document. Every draft shows what it was built from. A fact the assistant cannot verify is withheld and flagged, never filled in. Arithmetic, currency, duplicate and credit-note checks are done deterministically outside the model.
What happens when it is wrong?
It is caught in the packet, by your approver, before it reaches anyone outside. The error is logged with its cause, the workflow is fixed, and the fix is part of the next 30-day improvement. An assistant that starts producing wrong output is stopped, not tuned live.
How is it protected from being tricked?
Text inside a document — an enquiry, an invoice, a rate sheet — is treated as data, not instructions. The assistant never follows instructions found in the material it reads. A document containing instructions aimed at the model produces no action and is flagged. Customer requests cannot expand what the assistant is allowed to do; only the agreement can.
Where does the work run?
Inside systems you own or have approved, or on TYR-X infrastructure that holds no credential for your systems. You can see where each piece of work ran. TYR-X does not use your data to train anything, and does not pass it to any third party.
What happens on exit?
Access revoked. Scheduled work stopped. Your materials exported to you or deleted, under the terms we agreed, with confirmation.
What this page is not
It is not a certification and it does not claim one. It is the working rule set every TYR-X assistant runs under, written so you can check it against what you receive.
Questions
Questions
Is AI safe for a freight forwarder's data?
Yes, when the assistant can read only approved sources, cannot act outside a written permission list, and a person approves everything that leaves. That is how TYR-X assistants run.
Does the assistant need my passwords?
No. The first workflow uses no inbox, CRM or TMS credentials. Later workflows are scoped separately and only with a named approver.
Can a customer's email make the assistant do something?
No. Text in documents is data, not instructions; instructions found in material are ignored and flagged.
Is my data used to train AI?
No.
Who is responsible if something goes wrong?
Your approver sees every output before it leaves, and TYR-X is responsible for the assistant's operation under the agreement. Errors are logged and fixed; nothing is hidden in a report.