Authors
Moving a team of 10 to 100 people from shared AI logins to a key per person is a one-week job if you do it in order. This is the order that works. Each day has one goal. Nothing on Friday depends on anything you have not done by Thursday.
Monday: list what you have
Write down every AI account the company pays for: provider, plan type, who pays, who knows the login. Include the ones on personal cards that get expensed. Then list the tools people actually use: Claude Code, Codex, Cursor, Cline, chat apps, internal scripts. Expect to find an account or two nobody remembered. You now know the size of the job.
Tuesday: open the right accounts
For each provider you use, make sure the company holds an account licensed for team use: an API account and, if people use the chat apps, a Team or Enterprise plan. These are the accounts every token will run on from now. Keep the keys with one admin. Nobody else needs them.
Wednesday: connect and issue keys
Add the provider keys to Proxen. They stay server-side. Create teams that match how people actually work (engineering, content, ops, contractors). Give each team a model list and a default per-person budget. Issue one key per person and one per service. Send each person their own key through your password manager, never through a chat.
Thursday: point the tools
Each person changes one base URL and one key in each tool they use. For Claude Code it is two environment variables; for Cursor and Cline it is a settings panel; for scripts it is an environment variable on the service. The per-tool steps are in Pointing Claude Code, Codex and Cursor at one gateway. Ask people to make one request and check it appears under their name in the spend view. That check catches every mistake on the day it happens.
Friday: retire the shared logins
Once everyone's requests are arriving under their own key, change the password on every shared consumer plan and do not share the new one. Cancel the plans at the next renewal. Export any chat history people need first. From this point the company has one key per person and zero shared passwords.
The following month
This is where the bill starts to drop. Look at a month of spend by person, team and model. Trim history and context, fix caching, write routing rules for the three most common request types and check quality on real work. Move non-interactive jobs to batch. Set budgets from the real numbers. The levers and the figures behind them are in Cutting the AI bill without cutting usage.
When we do this with you, we run Wednesday and Thursday together on a call. If you want the week to start next Monday, get a free AI bill review first so you know what the rollout should save.