What not to paste into an AI chatbot

Your organisation's policy decides this, not us. What follows is the rough sorting we use when people ask where the line sits. Most material falls cleanly into one column. The middle column is where the judgement is.

Usually fine	Check first	Keep out
Published policies, procedures and annual reports	Internal drafts that haven't been approved or announced	Passwords, API keys, access tokens and connection strings
Your own draft text, emails and dot points	Meeting notes that name staff or discuss performance	Tax file numbers, Medicare numbers, bank and card details
Templates and made-up example data	Complaints, case files and FOI requests	Health information you don't need for the task
Public tender documents and legislation	Procurement evaluations and supplier pricing	Material classified above what your organisation allows in K9 AI
Aggregated figures with no individuals in them	Rosters, payroll extracts and leave records	Legal advice you have been told not to circulate

A few notes on the table.

Secrets go in the third column without exception. If you need help with a configuration file or a script, strip the credentials out before you paste it, and replace them with something like {api key}. The model doesn't need the real value to fix the problem.

The middle column is mostly personal and commercial information. It isn't banned. It's material where you should know why it's going in and what protection it gets on the way.

Before you paste

• Does the task actually need this detail, or would a description do?
• Would I be comfortable if this prompt were read out at a team meeting?
• Does it name or identify a person who isn't me?
• Is it commercially sensitive to a supplier, a client or us?
• Have I chosen the right classification label for it?
• If it's a spreadsheet, have I removed the columns the task doesn't use?

The last one catches more than you'd expect. A roster pasted in to fix a formula usually carries names, phone numbers and leave reasons the formula never touches.

What to do instead

Pick the classification label under the message box before you send. For anything that identifies people, use Personal, or OFFICIAL: Sensitive if your organisation uses government labels. K9 AI redacts personal details in your browser before anything reaches a model, and shows you what it found. Read that list. Redaction looks for names, addresses and identifying numbers, but it can't know that "the only female electrician at the northern depot" points to one person.

Where you can, describe instead of pasting. "A customer is disputing a $1,200 water bill and says the meter was misread" gets you a useful reply letter without the customer's account history.

K9 AI is hosted in Australia and prompts are not used to train models. That settles where your data goes. It doesn't change what your policy says you're allowed to put in, so the columns above still apply.