Set a spend boundary
Set a monthly budget or payment limit where the provider offers one, and use usage alerts or a usage dashboard.
Safety before keys
Do the small safety work before real data.
Before keys, connectors, source folders, or business data, set the practical limits that stop a demo becoming uncontrolled work.
Before using real business material
Start with limits, not trust.
Check training and data-sharing controls
Know whether conversations, tasks, or feedback can be used to improve models before you add real material.
Use fake or safe data first
A live demo should use public, synthetic, or harmless information. Do not use customer, member, payment, medical, HR, or confidential data.
Read first
The first useful agent usually reads selected sources and drafts an output. It should not send, update, delete, pay, or approve.
Leave evidence
Require source notes, assumptions, gaps, changed files, and review points so the human can see what happened.
OpenAI notes
Checked against official OpenAI pages on June 10, 2026.
OpenAI says API users can set a monthly budget in billing settings, with a possible delay in enforcement, and can also configure email notification thresholds. Check OpenAI API pricing and budget notes.
For ChatGPT, OpenAI's Data Controls FAQ says signed-in users can go to Settings, then Data Controls, and turn off Improve the model for everyone
so chats still appear in history but are not used to train ChatGPT. Read the Data Controls FAQ.
OpenAI's data-use FAQ also says business products, including ChatGPT Business, ChatGPT Enterprise, and the API Platform, are not used for training by default unless an organisation explicitly opts in. Read OpenAI's data-use FAQ.
Permission boundary
The first useful agent usually lives at read and draft.
May read
Selected public or internal-approved context for this workflow.
May draft
Briefs, checklists, summaries, questions, review packs, and first versions.
May not change
CRM, finance, production, contracts, official records, messages, permissions, or payments.
Must stop
When the source is missing, the instruction conflicts, the data is sensitive, or the next action would affect a real person or system.