GDPR and AI in small businesses: customer data, ChatGPT and your ERP
Your team already uses AI with customer data. A practical, non-legal guide to the GDPR questions that matter: access, credentials and records.
Your team already uses AI. The GDPR question is how
In most small and medium-sized businesses, AI arrived without a project plan. Someone pasted a customer complaint into ChatGPT to draft a reply. Someone uploaded an order export to get a quick chart. Someone connected an AI app to the shop.
None of this is automatically illegal. But the GDPR asks questions that an individual chat account cannot answer well: whose data is this, who can see it, why, and can you show it?
This article explains practical considerations. It is not legal advice. Talk to your data protection officer or lawyer about your specific situation.
Five questions to ask about any AI tool
1. Whose personal data goes in?
Customer names, addresses, order histories, invoices and, in hotels, notes about allergies or accessibility are personal data. Health information is a special category with stricter rules.
2. Who can see it afterwards?
When an employee uploads a customer list to a personal AI account, that list lives in that person’s chat history. Can a colleague open it? Does it leave when the employee leaves? Was it shared with a link?
3. Is access limited to what each person needs?
The GDPR principle of data minimization applies inside your company, too. Your warehouse team does not need customer payment histories. A pasted API token gives everyone who has it the full rights of the token’s owner.
4. What happens to credentials?
API tokens and logins to your ERP or shop are the keys to all of your customer data. If they sit in chat histories, scripts and spreadsheets, you cannot say who has access.
5. Can you document it?
If a customer asks what happened to their data, or an authority asks how you control access, you need records: who accessed which system, which changes were made and who approved them.
Personal AI accounts versus a team workspace
| Question | Personal AI accounts | Team workspace such as Steward |
|---|---|---|
| Where data ends up | Each employee’s chat history | Shared company projects; processed by Steward and the AI model providers it uses |
| Access per person | Whatever each person uploads or connects | Granted per person and team, per system |
| Credentials | Often pasted or stored by individuals | Entered once, encrypted, never displayed |
| Changes to customer records | Often direct | Proposals reviewed before they run |
| When someone leaves | Their chats leave with them, or stay with them | Remove their access in one place |
| Records | Scattered | Questions, proposals and approvals in one place |
How Steward supports good data protection practice
Steward is built to make the sensible approach the default:
- Connections belong to the company. An administrator connects weclapp, Shopify or ASA HOTEL once. Credentials are encrypted, never shown again and never given to the AI.
- Project roles. Viewers follow results, Contributors use configured connections and prepare changes, and Managers approve. Invited project members need no separate connection grant.
- Four-eyes principle for changes. The default Ask for approval mode and approval rules put a person between the AI and your records.
- Traceability. Proposals show who decided and the execution results. See Who approved that?
- Clean offboarding. Removing a person or team member removes their route to the data without new passwords. See Offboarding and AI access.
- Careful sharing. Public share links copy only selected visible content, and Steward strips common credential patterns. You still review business text before sharing. See Share a result.
Questions to clarify with your DPO
Even with good tooling, a few points belong in your own documentation:
- Which AI tools are approved for which data, and which are not.
- Your processing agreements with the providers involved, including the AI model providers. For Steward, ask us for the data processing terms.
- Whether special categories, such as health data in hotels, may be processed, and by whom.
- How long chats and files should be kept in your workspace.
- Who is responsible for granting and reviewing access each quarter.
A simple policy to start with
- No credentials in chat tools. Business systems are connected only through the company workspace.
- No customer exports in personal AI accounts.
- Read-only by default. Changes go through proposals and approval.
- Access by role. Review project and connection access quarterly.
- Offboarding includes AI. Remove workspace, team and connection access on the last day.
Get started
Project roles. Viewers follow results, Contributors use configured connections and prepare changes, and Managers approve. Invited project members need no separate connection grant.