Start with the job, not the data
“Give ChatGPT access to our data” is too broad to do safely. Start with a specific job instead: answering questions from the staff handbook, summarising a client’s recent correspondence, or checking project status. The job tells you which data is needed, and, just as usefully, which is not.
The main options
1. Uploading files. People upload documents into a chat or project. This needs no set-up but creates copies, goes stale quickly and is hard to govern. It is also how confidential data most often ends up in the wrong place.
2. Built-in connectors. ChatGPT’s business plans may offer connectors to common services such as file storage and email. These are convenient. Check your provider’s current settings for which services are supported, how permissions work and what administrators can control.
3. Custom connectors. You, or a supplier, build a connector, for example an MCP server, that exposes your own systems. This gives more control and more responsibility.
4. A gateway. ChatGPT connects to a single governed layer, which connects to your systems. Permissions, approval and audit are handled once, for every AI tool.
A step-by-step approach
- Choose the use case. One job, one team, one set of systems.
- Classify the data. Mark what is public, internal, confidential and restricted. Exclude restricted data, such as payroll, health information or legal privilege, until you have a specific reason and a specific control.
- Check the account type. Business and enterprise plans generally offer different data-handling terms from personal accounts. Make sure staff are using the account your organisation has approved.
- Connect with the least access. Prefer read-only, prefer delegated access that respects each person’s existing permissions, and limit scope to the relevant sites, folders or records.
- Separate clients. If you are a professional firm, make sure an employee’s AI cannot mix one client’s information into another’s work.
- Keep a record. Know which data was retrieved, for whom and when.
- Tell people the rules. A short, clear AI policy prevents most problems.
- Review. After a few weeks, check what was used, what was missing and what was over-shared.
UK considerations
If the data includes personal information, UK GDPR applies. You should understand what the AI provider does with the data, have appropriate terms in place, and be able to explain the processing. The ICO publishes guidance on AI and data protection that is worth reading. This is general information, not legal advice.
Common mistakes
- Connecting a whole file share because it was quicker than choosing folders.
- Using an administrator’s account for the connection.
- Letting each employee set up their own connections with their own credentials.
- Treating ChatGPT as the only AI the business will ever use.
How Kroy approaches it
Kroy is designed for the fourth option. You connect your business systems to Kroy once. ChatGPT, and any other AI you approve, connects to Kroy. Kroy holds the credentials, checks each request against the person, the agent, the delegation and your policies, and records what happened. Connect your business once. Use it safely with any AI.