Who is allowed to send what, and from which company?

Picture a family group with three businesses: a hardware shop, a transport company and a small property arm. One office manager holds the logins for all three mailboxes. She is good at her job. One busy afternoon a quote for a transport client goes out from the property company's address, with the property company's signature and bank details at the bottom.

Nothing criminal happened. But the client now has the wrong company's bank details in a thread they will reply to for months, and someone has to explain why.

Anyone who has run shared admin for more than one business knows a version of this story. It gets more likely, not less, once AI tools and assistants can reach every mailbox through the same connection.

Why we wrote this down yesterday

Provecta Group handles communications for more than one company. Each has its own address, its own clients and its own documents. Until this week the separation lived in people's heads. By now our AI agents can reach more than one company's mail and files through the same set of tools, and that makes a wrong-company send a real possibility rather than a theoretical one.

So yesterday, before anything went out wrong, we wrote one rule into the company memory that every AI session and every person reads first:

Before sending or filing anything, name the company first, then use only that company's row in the table. If there is no row, or the row says TBD, stop and ask. Never improvise a sender.

The table itself is plain. One row per company. For each one it says:

1. The address that every human-facing message comes from, whoever it is signed by.
2. How that message is sent, and from which tool.
3. Which address automated mail comes from, if any.
4. Where replies land, and where that company's files live.

A separate standing rule sits next to it: the table decides where a message comes from. It never decides whether the message goes. Sending still needs a yes for each message.

What the table catches

Most wrong-sender mistakes are not decisions. They are defaults. The last mailbox you used is still open. The AI picks the first account it was given. A signature block was copied from an old template. A table that forces the question "which company is this?" before anything else turns a silent default into a visible choice.

It also catches the case nobody planned for. A new brand starts and nobody has set up its mailbox yet. Without the table, someone borrows the nearest address "just this once". With it, the row says TBD and the work stops until someone decides.

For AI agents this matters more than for people. A person might hesitate at the wrong letterhead. An agent will send from whatever account it was handed, with complete confidence, at three in the morning.

What it does not do

The table does not stop a person who ignores it. It does not fix shared passwords, and it does not tell you who should have access to each mailbox in the first place. Those are separate jobs. It also goes stale. If a company changes its sending address and nobody updates the row, the table will send everyone to the old one. Our rule for that is blunt: when a channel changes, the table changes in the same session.

And it says nothing about what the message says. A message can leave from the right company and still be wrong.

Build yours this week

You do not need software for this. A shared document works.

Fill one row per business you run or support. Leave a gap as TBD rather than guessing. Pin it where your team, and any AI assistant you use, will see it before they write anything. Then add the single line that does most of the work: name the company first.

If you would like a second pair of eyes on how mail, files and approvals are split across your businesses, book 20 minutes at pgco.world/book.