Your IT department sets the rules. Every employee's agent follows them.

Matej and Nina

1. Matej · IT

   Creates the company and invites people

   In his coding agent he says create my company, then invite nina@firma.si. Nina gets one mail: a join code, the install line for her tool, or the connector address if she works in claude.ai or ChatGPT, and the first thing to type.

   create my company · invite nina@firma.si
2. Matej · IT

   Writes the rules, once

   Which AI tools are allowed and which are forbidden. Which MCP servers. Which data may go into an AI: customer data, financial data, source code, internal documents, credentials, public. Which kinds of work need his approval first: research, a new feature, sharing outside, automation, customer data. One page, on the Teams site; the rules reach every member's agent.

   Rules v1: customer data → needs approval · credentials → forbidden · research, new features → ask first
3. Nina · marketing

   Joins, and her agent knows the rules

   She says join the company with code K7QD-92MX. From then on her agent reads the company rules at the start of every session. She never configures anything.

   join the company with code K7QD-92MX
4. Nina · marketing

   Has an idea

   "I have an idea for a feature: a loyalty badge for customers who ordered three times. I want to research it." Research and a new feature are on the approval list, so before any work her agent sends Matej a request: what, with which tools, with which data, why, and tells her it is waiting.

   Sent to IT: "Loyalty badge research" · tools: web search, the CRM · data: customer orders · waiting for approval.
5. Matej · IT

   Gets the request and decides

   A mail, and an item in his inbox (in his agent: inbox; or on the Teams page). He approves it with a condition: without customer names, counts only. Nina gets the decision in her agent and by mail.

   approve with conditions: no customer names, counts only
6. agentgov → dataready

   Is the idea worth it?

   Approved, the answer names the next step: test the idea on the company app's synthetic community with dataready, installed next to agentgov. Simulated customers try the badge; Nina sees at once who would use it, who would not, and why. The report link is attached to the request, so Matej and the management see the approval and the estimate together.

   Next: recommendation_test on dataready → the report → team_attach_report.
7. Nina · marketing

   Does the work, within the conditions

   Her agent follows the conditions: counts, no names. Everything is on record: who asked, who approved, what the condition was, what dataready estimated.

8. Matej · IT

   Watches, and gets a weekly mail

   Every Monday: new requests, decisions, members, and the MCP servers the watch has seen on people's computers that the rules do not allow. The watch sends names, transport and a hash, never arguments or secrets, and only when the member agreed or the rules require it.

   This week: 3 requests, 2 approved, 1 denied · 1 new member · 1 unapproved MCP server seen (github, on 1 computer).