Skip to content

AI privacy and data security.

We build AI systems to protect what your clients entrust to you, and we can show you how.

Security by design.
Controls you can verify.

How VaultIQ protects your data

VaultIQ restricts access and uses business AI APIs. Model training is off by default, and data processing agreements govern how personal data is handled.

For stricter confidentiality requirements, we can deploy the language model and an anonymisation proxy on your premises. The proxy can remove agreed identifiers before data reaches the model.

Frequently asked questions

How does VaultIQ handle data and AI providers?

VaultIQ connects to AI models through business APIs. As with a managed business workspace, the provider’s business terms and applicable DPA govern how it handles covered data. When VaultIQ uses the OpenAI API, OpenAI excludes inputs and outputs from training by default. OpenAI API data controls

VaultIQ adds application-level safeguards, including authentication and access controls. We agree what data can be sent to the model provider and put a DPA in place before Etashi processes personal data on your behalf. VaultIQ security and data protection terms

What does a business AI account give me?

Covered business AI plans exclude customer content from model training by default. A data processing agreement (DPA) sets out how the provider must handle personal data on your organisation’s behalf. Workspace controls let your organisation manage user access, with sharing and retention options depending on the plan.

What security safeguards should I look for?
  • Encryption in transit and at rest.
  • Authentication, access controls and least-privilege permissions.
  • Independent security assessments and relevant certifications or audit reports, with their scope checked.
  • Contractual data-processing commitments.
  • Retention policies, monitoring and incident-response procedures.
  • Data minimisation and restrictions on use for model training.
Why does a data processing agreement matter?

A training switch limits whether the provider can use eligible chats to improve its models. That commitment matters on a personal account, too.

A data processing agreement (DPA) governs what the provider does with personal data on behalf of your organisation. It sets out whose instructions the provider follows, how it protects the data, when it reports breaches, how it uses subcontractors and what happens to the data at the end. When GDPR applies and the provider acts as a processor, Article 28 requires binding terms. These can be part of the service agreement and accepted electronically. GDPR Article 28

Take payroll. With training off, the provider can still process the spreadsheet to answer your question. With a DPA in place, your company also has agreed terms for what the provider can do with the employee data and what it must do if there is a breach or the contract ends. If you only need to compare salary bands, send the bands and leave out the employee records.

What are the four levels of AI privacy?

Choose the level that fits the sensitivity of the data and the controls your organisation needs.

  1. Personal cloud account with training switched off

    Use the provider’s consumer privacy settings to exclude ordinary chats from model training. The provider still processes your message to answer it. Storage, human review and feedback follow the provider’s rules. Use this level for public or low-sensitivity material.

  2. Managed business workspace

    Use the provider’s commercial service under your organisation’s agreement. Covered workspace content is excluded from training by default. Your organisation can manage membership and apply the access, sharing and retention controls available on its plan.

  3. Controlled application using a business API

    Use an application that sends only the fields needed for each request. The application can enforce permissions, remove identifiers and limit what it puts in logs. Check the provider’s retention for the exact API features in use, as well as what the application itself keeps.

  4. Isolated or self-hosted deployment

    Run the model inside a defined environment when data must stay within that boundary. Check every route data can take, including connectors, telemetry, logs and backups. The operator takes responsibility for securing and maintaining that environment.

How do ChatGPT, Claude and Gemini compare?

All three have a personal training control. Business protections follow the account, product and agreement in use. Settings and terms checked .

ChatGPT

Personal. In Data controls, switch off Improve the model for everyone. New ordinary chats are excluded from training and can stay in your history. Feedback can share the associated chat for training, and some features have separate rules. Personal data controls

Business. ChatGPT Business and Enterprise exclude workspace content from training by default. The services agreement restricts use of customer content for improvement unless the customer explicitly agrees. Services agreement · Data processing addendum

Claude

Personal. Turn off Model Improvement in privacy settings. Submitted feedback may still be used for training, and flagged chats may be used to develop safety systems. Personal training policy

Business. Claude Team and Enterprise exclude covered inputs and outputs from training by default. Feedback or voluntary sharing can change that. Workspace owners can disable chat ratings. Commercial terms · Data processing addendum · Training and feedback

Gemini

Personal. Turn off use of Chat content for training in Gemini History where that control is available. If your account shows Keep Activity instead, turn it off. Feedback can still be used for training, and Google can still retain and review chats. Gemini Apps Privacy Hub

Business. Google restricts model training and fine-tuning on customer data without prior permission or instruction for covered Gemini services under the Workspace agreement. Check whether the app is a core Workspace service for your edition. Workspace coverage · Service terms · Cloud data processing addendum

Is enterprise AI like online banking?

The comparison is in how you assess trust. With online banking or enterprise AI, look for encryption, restricted access and independent assessment of the controls. No online service can promise zero risk. At Etashi, we manage that risk through appropriate contracts, careful system design, restricted access and data minimisation.

How long can the OpenAI API retain data?

Standard abuse-monitoring logs may include content and are generally retained for up to 30 days. Some features keep application data separately. Zero Data Retention requires approval and applies only to eligible use. The application’s own access, storage and logging need their own controls. OpenAI API data controls

Can my organisation enforce its no-training policy?

All three providers make contractual training commitments for their covered business services. If a provider breaches that commitment, your organisation can point to the agreed clause. The remedies depend on the agreement and applicable law. To enforce your policy across staff, use the approved workspace, configure the controls available on your plan, and review opt-ins, feedback and the terms for any external tools that receive data.