- Through their business APIs, OpenAI and Anthropic do not train models on your data by default, but data still leaves your systems.
- Send the model only the data each task needs, and mask names, contact details and identifiers where the task does not require them.
- Keep API keys on the server, restrict who and what can call the AI, and log requests without storing sensitive content in plain text.
- Update your privacy policy, contracts and data processing agreements, and check regional rules such as GDPR before going live.
How do you keep customer data safe when using AI APIs?
Use the providers' business APIs, which do not train on your data by default, send each request only the data the task truly needs, mask personal details where you can, keep keys and access tightly controlled, log carefully, and cover the processing in your privacy policy and contracts. AI APIs are powerful, but every request sends data to another company's systems. Treat them like any other data processor: understand their terms, minimise what you share and keep a record of what happens.
These are the practices we build into every AI automation project.
What do the main providers do with API data?
- OpenAI states that data sent to its API is not used to train or improve its models unless you explicitly opt in. Abuse-monitoring logs are kept for up to 30 days by default, and eligible customers can apply for zero data retention (OpenAI: your data).
- Anthropic's commercial terms state that it may not train models on customer content from its services (Anthropic commercial terms).
Read the current terms for every provider you use, including smaller ones and any service that sits between you and the model, because policies differ and change.
Why is data minimisation the most important step?
The safest data is data you never send. For each AI task, list the fields it actually needs:
- Classifying a support email needs the message, not the customer's address or order history.
- Summarising a call needs the transcript, not the customer's payment details.
- Extracting invoice fields needs the invoice, not your whole supplier database.
Build the prompt from those fields only. It reduces risk, lowers token costs and often improves answers, because the model has less irrelevant material to wade through.
When should you mask or redact data?
Where the task does not need someone's identity, replace names, emails, phone numbers, account numbers and similar identifiers with placeholders before sending, and restore them in your own system afterwards if needed. For example, "Customer A asked about invoice 123" is enough for the model to draft a reply template. Pattern-based redaction catches structured data such as emails and card numbers; free-text names need more careful handling and testing.
How should keys and access be controlled?
- Keep API keys on the server, never in browser code, mobile apps or public repositories.
- Use separate keys per project and environment, with spending limits, and rotate them when people leave.
- Restrict which parts of your system can call the AI, and with which data.
- Give AI agents the smallest permissions they need in your CRM, email or databases, and require approval for sensitive actions.
Our cloud and DevOps team sets up secrets management and access controls for these systems.
How should you log AI requests?
Logs help you debug, audit and improve, but they can become a second copy of sensitive data. Log metadata by default: time, user, task, model, token counts and outcome. Store full prompts and responses only where you need them, protect them like production data, and delete them on a schedule. Make sure your own retention matches what your privacy policy promises.
What legal and contractual steps are needed?
- List each AI provider as a processor or sub-processor in your privacy policy and records.
- Sign the provider's data processing agreement where one is offered.
- Check where data is processed if you serve customers under rules such as the GDPR, and whether transfers need safeguards.
- Update client contracts if you process their customers' data with AI on their behalf.
Rules differ by country and industry, so involve your legal or compliance adviser before processing regulated data such as health or financial records.
How should the system itself be designed?
Put a thin service of your own between your applications and the AI provider. It builds minimal prompts, applies redaction, enforces limits, logs consistently and lets you switch providers without touching every application. This is a small piece of custom software that pays for itself in control and peace of mind.
What should you ask an AI vendor or agency before sharing data?
- Which AI providers and models will process our data, and in which regions?
- Is any of our data used for training, by the provider or by the vendor?
- How long are prompts, responses and logs kept, and who can access them?
- How are API keys stored, and what happens to access when the project ends?
Clear written answers to these questions belong in the contract, not only in a sales call.
What is a quick checklist before going live?
- Business API terms reviewed for every provider.
- Each task sends only the fields it needs, with redaction where possible.
- Keys on the server, with limits and rotation.
- Logging and retention agreed and documented.
- Privacy policy, contracts and processing agreements updated.
- People told when they are interacting with an AI system.
Frequently asked questions
Do OpenAI and Anthropic train on data sent through their APIs?
Not by default. OpenAI states that data sent to its API is not used to train or improve its models unless you explicitly opt in. Anthropic's commercial terms state that it may not train models on customer content from its services. Consumer chat apps can have different settings, so use the business APIs for customer data.
How long do AI providers keep API data?
It depends on the provider and your agreement. OpenAI says API abuse-monitoring logs are kept for up to 30 days by default, with zero data retention available to eligible customers after approval. Check each provider's current documentation and your contract, and choose the retention option that matches the sensitivity of your data.
Can I use AI APIs with health or financial data?
Only with extra care and the right agreements. Regulated data often requires specific contracts with the provider, strict access controls, minimisation and sometimes particular regions or deployment options. Talk to your compliance adviser before sending regulated data to any external service, and design the system so it only sends what is essential.
Should we tell customers we use AI?
Yes. Explain in your privacy policy which AI providers process data and for what purpose, and tell people when they are chatting with an AI assistant rather than a person. Transparency builds trust, and in some places it is a legal requirement.
Last updated