Entrusting an AI project to a team you never run into in the hallway raises a legitimate question, and our clients almost always ask it in the very first conversation: “concretely, what guarantees that our data stays protected?”. We work remotely, as an extension of our clients’ teams. This setup has real advantages, but it is only worth something if trust is solid.
Rather than answering with general promises, this article describes the rules we actually apply on every engagement. They do not depend on individual goodwill: they are part of how we work, from contract signature to the last day of the engagement.
Before any data is shared: the contractual framework
No data moves before the framework is in place. In practice, three commitments cover every engagement:
- A non-disclosure agreement (NDA), signed before any sensitive information is shared, including during scoping.
- A GDPR data processing agreement (DPA) whenever the engagement involves personal data. It sets out what we may do with that data, and above all what we may not do.
- An individual confidentiality commitment: every team member is personally bound by a confidentiality clause. Protection does not rest only on the company, but on every person involved.
Access: named, protected, strictly necessary
Working remotely means security cannot rely on a badge or a locked office. It relies on controlling access.
- Multi-factor authentication everywhere: email, cloud tools, code repositories, client tools. A stolen password is not enough to get in.
- Named accounts: one person, one account. Never a shared account, so you always know who has access to what, and can remove access individually.
- Least privilege: everyone has the rights their task requires, no more.
- Client-provided access: whenever possible, we go through the VPN or environments provided by the client. The client keeps control of its own security rules and of connection traceability.
- A vault for secrets: passwords, API keys and credentials are stored in a dedicated password manager, never in a file, a message or a code repository.
Working remotely does not lower the bar. It simply forces every rule to be made explicit.
During the project: as little real data as possible
The best way to protect data is not to handle it when you do not need to. That principle guides how we build.
- Synthetic data first: design, development and most testing are done on synthetic or anonymized datasets built to mirror the structure of the real data.
- Real data stays with the client: when we need to work on production data, we do it inside the client’s infrastructure, not on our machines.
- Pseudonymization before anything is sent to a model: names, email addresses and case numbers are replaced with identifiers before text is sent to an LLM, then reinserted into the response if needed.
- Strict separation per client: environments, code repositories and API keys are separate for each client. Nothing moves from one project to another.
AI models: business access, chosen by sensitivity
We work with several families of models: the OpenAI and Anthropic APIs, Mistral models, and open models that can be self-hosted. We only use business API access, never consumer apps, and the choice depends on how sensitive the data is.
For low-sensitivity content, a contracted API with no reuse of data for training is generally appropriate. For more sensitive data, European hosting or a model deployed inside the client’s infrastructure ensures no data leaves its perimeter. This choice is made with the client, documented, and revisited if the project scope changes.
At the end of the engagement: nothing stays behind
A well-finished engagement is also one that leaves no trace. At the end of every project:
- Client data is deleted from our machines and tools.
- Access is revoked: accounts, VPN, rights on the client’s tools.
- API keys and secrets are rotated or handed over to the client, so no credential we knew remains active.
- A written confirmation tells the client these steps have been completed.
What you can ask us
Trust is not declared, it is verified. Before starting a project with us, you can rightly ask to see the NDA and DPA, find out which models will be used and where data will be processed, know who will have access to your systems, and agree together on how the engagement will end. We encourage these questions, because they are part of a well-run AI project.
In short
Working with a remote team does not mean compromising on security. It takes a clear framework, explicit rules and discipline applied on every engagement. That framework is what lets us work close to your data without ever losing control of it, and hand it back intact at the end of the project.