AI Agent Best Practices
- Easy to feed, hard to delete.
- It summarizes one file and learns all of it.
- One account for work. One for you.
What this episode teaches
- A personal AI account keeps everything you put into it, and it leaves with you when you leave the company
- An agent asked to summarize one document also retains everything else that document contained, including material you never meant to hand over
- Before putting work into an AI tool, check with your security team, and use an account created with your work email rather than a personal one
Two ways to read it
If you are the one watching
Nobody is in trouble for using these tools, and getting good at them is the point. The catch is where the work goes. If you signed up with a personal email, everything you have ever pasted in lives in an account with your name on it: the contract, the thread, the resume, the things in the file you never actually asked about. It stays there, and it goes with you when you leave. The fix is one minute of setup rather than a decision every time you open a chat. Check with whoever handles security or IT, then make a separate account on your work email and keep work in it.
If you run the program
The precise value of the work-email rule is worth stating plainly, because the episode deliberately does not: an account created on a domain you control can later be discovered, claimed and migrated into an enterprise tenant. Anthropic supports domain claiming on Enterprise, and OpenAI has a working path when a verified-domain user accepts an invite. An account on a personal address can never be recovered, whatever you buy afterwards. So this makes the account findable rather than secure, and the data still reaches the provider either way. The real control is provisioning. Permitting AI without buying accounts accumulates custody in places the organisation cannot see, audit or clean out, and that exposure is rarely on anyone's register. Until seats exist, the settings worth telling people to check are training and model improvement, chat history and retention, memory, and connectors or extensions.
Run the whole library
A new episode every month, captioned in 14 languages, ready to drop into your own LMS. Start a free trial and take the bundle with you.
Start a free trial