The detail behind a guarded AI employee.
An AI employee that works from your live data, with about seventy real tools, and asks before it changes anything.
The limits are part of the product.
Gavin is built to work inside the boundaries you set, not around them.
Security practices
Security practices are described in our Data Processing Addendum.
You decide what he may touch.
A settings grid covers 21 areas of the business and six verbs: view, create, modify, delete, send, approve. Set each one to automatic, ask first, or never. A setting can only ever narrow what Gavin may do; nothing can widen it.
He can only see what you can see.
Gavin runs under the asking person's own login. Not an admin key, not a back door. If a member of staff cannot see payroll by hand, Gavin cannot see it for them. And every tool needs two separate permissions: the ordinary one, and a specific AI one. Letting someone do a thing themselves is a different decision from letting Gavin do it for them.
Some things he simply cannot do.
There is no delete tool. No refund tool. No tool to mark an invoice paid by hand, merge records, change a price, or send a campaign. Gavin answers your incoming business calls as your AI receptionist. He cannot run code, open a URL or read a file. These are not rules he is asked to follow: the tools do not exist.
He cannot message everybody.
One customer message per request. That ceiling is enforced in the runtime, and there is no bulk tool to get around it. Anything Gavin sends still passes do-not-contact, SMS STOP and email unsubscribe first. When Gavin runs on a schedule with nobody watching, he cannot message a customer or touch money at all, and there is no setting to turn that off.
Personal details are pseudonymised before they leave.
Emails, phone numbers, card numbers, IBANs and account numbers are swapped for placeholders on the way to the model and put back on the way to you. The mapping lives in memory for one request and is never stored or logged. You can also set your workspace to redacted mode, or turn AI off entirely.
Everything is written down.
Every action Gavin takes, every approval, every refusal, in one auditable record. You choose how long it is kept; the default is a year.
Gavin runs on frontier-class reasoning models and falls back automatically across providers, so an outage at any one of them does not take your assistant down. Answers stream as they are written. Ask by typing or by speaking.