How we keep your data out of model training.

1. The boundary is designed first

Before any model is selected, we map where data moves: what leaves your network, what reaches a third-party API, what is stored and for how long. That map is a deliverable, not an afterthought, and it is what everything else is built against.

2. Zero retention, no training

We deploy against model endpoints configured for zero data retention and covered by contractual terms that prohibit using your inputs or outputs for model training. Where a provider cannot offer both, we say so explicitly and either propose an alternative or recommend self-hosting an open-weight model inside your own environment.

3. Redaction before the boundary

Personal and sensitive identifiers are detected and redacted or tokenised before a request crosses into any external service, and re-hydrated on return where the workflow requires it. For clinical and claims work, this is enforced in the pipeline rather than left to prompt instructions.

4. Isolation and least privilege

Each tenant, and each agent within a system, receives scoped credentials and access to only the tools and indexes its function requires. An agent that summarises documents cannot reach your billing API because it was never given a path to it.

5. Everything is traceable

Every request is traced: the prompt, the retrieved context, the model and version, the tools invoked, the output and the human decision that followed. When a regulator, an auditor or a customer asks what happened in a specific interaction, the answer is a record rather than a reconstruction.

6. Adversarial testing is part of delivery

Prompt injection, data exfiltration through tool use, jailbreak attempts against guardrails and unsafe-content handling are tested before launch and re-tested when models or prompts change.

7. Your cloud, your region, when it matters

Where data residency or procurement rules require it, systems are deployed entirely inside your own cloud account and region. We build there and hand over the keys.

8. Human oversight where the stakes justify it

In clinical, legal and financial workflows, AI output is a draft for a qualified person to approve. We design the review step into the product, and we measure whether people are actually using it rather than clicking through.

What this page is not

This is a description of our engineering practice, not a certification or a legal warranty. The specific controls, retention terms and compliance obligations for your programme are agreed in writing as part of the engagement. For how we handle personal information submitted through this website, see our privacy policy.

Questions about any of the above — including a control we have not listed — contact@conceptboxtechnology.com.