The invoice carries data the cut does not use
An invoice from ChatGPT, Claude, Copilot, or an API shows what you pay. It usually also shows the admin’s name, their email, a tax ID, an address, or an IBAN. A member export adds one row per person.
Estimating whether to drop seats or optimize does not need any of that. It needs the vendor, the period, the plan, the amount, and how many seats are paid against how many are opened.
Keep the figure and drop the person
Copy only those figures into a note. If the same vendor has two workspaces, call them 1 and 2 in the note and keep the real name in a file you do not paste anywhere.
From a member export, count rows: people, people who signed in during the period, and admins. Do not paste the email column. A box drawn on the PDF does not count if selecting the text still reveals the email.
Personal chat and Business are not the same
On ChatGPT for individuals, OpenAI may use content to improve its models. You can turn that off under Settings, Data controls, with Improve the model for everyone, or in the privacy portal with Do not train on my content. Either one is enough. After that, new conversations are not used to improve the models.
If you send a thumbs-up or thumbs-down, OpenAI may use that whole conversation for training even when improvement is off. On ChatGPT Business, Enterprise, Edu, and the API, inputs and outputs are not used to improve the models by default. The cut is still decided from the count, not from the list of people.
What to compare this week
Pull the invoice and the export yourselves. Move the amount, the plan, and the seats into the note. The original file stays in your inbox.
If the same month you also pay for another tool, an auditamia report lines those figures up with the export you generate. We do not log into the account. From €99, in 48 hours, with a verdict per tool: keep, reduce, cancel, or watch. The public sample is Hatchpad, a fictional company.