Artificial intelligence in EterShield
In short: EterShield uses AI to help you write, analyse and fill in, but AI prepares proposals and the decision is yours. This page explains which AI is used, what for, which data reaches it, who reviews it and how you control it.
Summary in 5 sentences
- EterShield's AI is a language model from an external provider, Claude (Anthropic). EterShield does not train or fine-tune models on customer data.
- It appears in four kinds of feature: the Vera assistant (chat), the Suggest with AI button on forms, analyses and generators (gaps, policies, reports, supplier assessment, assessment filling) and information extraction (meetings, documents, images).
- The rule is AI prepares proposals, the decision is yours: in form suggestions and in Cook only what you accept is saved, and generated content is labelled as AI.
- AI requests are handled by an external provider outside the European Union (United States), under Standard Contractual Clauses, and the contract provides for zero retention and no training on customer data.
- You control who uses AI through role permissions, and each answer carries a transparency indicator (confidence, sources and limits).
Principles
- AI prepares, the decision is yours. In form suggestions and in Cook, an AI answer is a proposal: it becomes part of your records only after you accept, edit or reject it. Other analyses, such as gaps and supplier assessment, are saved identified as AI-generated, for you to review before using them as a decision or as evidence.
- Transparency in every answer. Each answer shows the confidence level, sources and caveats, in the transparency badge next to the result.
- Minimum necessary. Only the context that is needed, and that the user can already see, is sent, with a size limit.
- Everything is logged. Usage, human decisions and AI incidents are in your organisation's audit log.
- Your data does not train models. EterShield does not train models on customer data.
Map of AI features
The table says, for each feature, what it is for, which data it uses and how it is reviewed.
| Feature | What for | Data it uses | Human review |
|---|---|---|---|
| Vera (vCISO assistant) | Answer questions, look up risks and assessments, create risks and tasks, draft policy text, prepare AI agent drafts | The conversation and the summary of the organisation the person already has access to (risks, gaps, policies, assessments, tasks) and excerpts of your documents | Risks created are labelled as AI and enter a review queue; AI agents stay as drafts until a person confirms them |
| Suggest with AI (forms) | Propose values for the fields of a form | The fields already filled in (without personal fields) and context from records of the same entity | Accept, edit or reject field by field; nothing is saved until you submit the form |
| Gap analysis and executive report | Recommendations for gaps and a summary for management | Data of the module concerned (gaps, assessments, policies, tasks) | Text to read and decide on |
| Cook (assisted filling) | Fill in assessments, risks, tasks, policies and suppliers | Framework controls and entity context | Cook only proposes: you accept, edit or reject each item before anything is saved; accepted items are marked as AI-generated |
| Policies, suppliers and risks | Review policies, assess suppliers, estimate values for risk quantification | Data of the record concerned | Human approval and editing |
| Extraction (meetings, knowledge, documents) | Turn transcripts and documents into structured items | The text or document you submit | Items are proposals; you choose what to keep |
| AI governance (Aria, EU AI Act, AI register) | Record and classify your organisation's AI systems | The data you enter | Human review and approval |
Provider and processing region
- Provider: Anthropic PBC (Claude model), listed as a sub-processor in the EterShield DPA.
- Region: AI requests are the only processing that leaves the infrastructure in the European Union: they are processed in the United States, under Standard Contractual Clauses (SCC) and a transfer impact assessment (TIA).
- Retention and training: under the contract, the provider applies zero retention to API data and does not train models on customer data.
- Other providers: by default there is no other model provider.
Controlling AI in your organization
- AI switched off: an administrator can switch AI off for the organization in Settings, Data and AI. With AI off, no AI feature runs and screens show a clear message.
- Quota and usage: every call counts toward the organization's monthly AI call quota. In Settings, AI usage, you see suggestions, calls and estimated cost, by operation and by user.
- Vera: each conversation has a limit on messages and on characters per message.
Which data reaches the AI
- Only what the feature needs reaches the AI: the text the person writes or submits and the context of the organisation and entity they already have access to.
- On forms with Suggest with AI, personal fields (for example names, emails and tax IDs) are not sent.
- In chat and extraction features, the text you write or submit may be sent to the provider. Avoid entering sensitive personal data in free text unless necessary.
- The text of your requests is not stored in clear in the AI logs; usage, the accepted answer and the human decision are stored.
- The content of suggestions is deleted after 12 months; metadata and the final decision remain.
Human review and transparency
- Each answer shows the confidence level (high, medium or low), sources and caveats.
- What the AI creates is labelled as AI and, where applicable, enters a review queue.
- Decisions (accept, edit, reject) are logged, and you can contest a suggestion.
- No decision affecting people is taken by the AI without human intervention.
Risks and how we reduce them
| Risk | How we reduce it |
|---|---|
| Wrong or invented answer | It shows confidence and sources; in suggestions and Cook a person reviews before saving; other analyses are identified as AI-generated, for review |
| Malicious instructions hidden in text (prompt injection) | Texts are treated as data and not as instructions, suspicious content is filtered and the incident is logged |
| Personal data leakage | Personal fields on forms are not sent; context is limited to what the person can already see; the provider contract excludes retention and training |
| Excessive or abusive use | Request limits and plan quotas; usage log |
| Improper access | Each request respects the role permissions and the entity scope of the person |
| Dependence on a provider | EterShield's core features work without AI |
Customer control
- AI is optional: the core EterShield features work without it.
- Who can use it: the administrator defines, by role, who can use and who can manage the AI features.
- How much is used: each plan includes an AI usage quota, and usage is visible in Settings, AI usage.
- Turning it off: an administrator disables AI for the organisation in Settings, Data and AI.
Compliance
- GDPR: the provider is listed as a sub-processor in the DPA, and the transfer is covered by Standard Contractual Clauses and an impact assessment.
- EU AI Act: EterShield includes a dedicated module to record, classify and review your organisation's AI systems, with human oversight.
- Transparency (AI Act Art. 50): AI-generated answers are labelled as such.
Frequently asked questions
Does the AI use my data to learn?
No. EterShield does not train models on customer data, and the provider contract excludes training.
Where are AI requests processed?
At an external provider in the United States, under Standard Contractual Clauses. The rest of EterShield (application and database) stays on a dedicated server in the European Union.
Can the AI change my records on its own?
No. Forms only save what the person accepts; what the AI creates is labelled and goes through review.
Can I use EterShield without AI?
Yes. The AI features are optional and the core of EterShield does not depend on them.