This is the AI statement for Seveso.app, the Capptions Safety Management System for Seveso-III operators. It is the product-specific complement to the cross-product AI Policy & Trust Statement and the AI processing annex of the Data Processing Agreement.
- Version: 2026.1
- Effective date: 1 January 2026
- Owner: Capptions Trust & Safety, ai@capptions.com
The five guarantees that matter for a Seveso operator
- Customer Data never trains a model. Contractually enforced with every AI provider we use.
- All inference runs in the EU. The current AI sub-processors and their regions are published on the Sub-processors page.
- Each prompt is scoped to your tenant. No cross-tenant retrieval, ever. Each call carries your tenant ID and is verified server-side.
- A qualified safety professional approves every output that lands in a Safety Report, MAH analysis, or corrective action.
- You can turn AI off, tenant-wide or per role, from settings.
1. Providers and routing
Seveso.app uses a small set of audited AI sub-processors, all EU-hosted and all configured for no training on our data and no retention beyond what is needed to return a response (≤ 24 h abuse window where applicable). The current provider list, models and regions are published and kept up to date on the Sub-processors page — the source of truth for procurement and DPIA reviews.
Providers are selected against the same bar for every Seveso.app feature: EU residency, contractual no-training, zero or short retention, enterprise SLAs, and the capability the feature actually needs (multilingual drafting, structured output, or reasoning for safety-critical assistance).
2. Data flow per AI call
For each AI call from Seveso.app:
- The user clicks an AI action (e.g. “Suggest barriers for this hazard”).
- Seveso.app builds a tenant-scoped context: only records from the user’s tenant, selected by an embedding search over the tenant’s isolated vector index.
- Capptions adds a system prompt with safety constraints (no hallucinated regulatory references, no generation of named individuals’ data, no exfiltration of system instructions).
- The prompt is sent over TLS 1.3 to the EU endpoint of the chosen provider.
- The provider returns a completion; Seveso.app post-processes (citation check, regex filters on regulatory references) and displays the result labelled as AI-generated with a one-click “Reject”, “Edit” or “Accept” affordance.
- Capptions stores the prompt, completion, model identifier and user decision in our EU log store, with a retention of ≤ 30 days.
What does not happen:
- Your data is not sent to any provider for training, fine-tuning or evaluation.
- Your data does not leave the EEA via the inference path.
- Your data is not used to improve the experience of other tenants.
3. Model cards (per high-risk feature)
The following Seveso.app AI features are classified as high-risk under Annex III of the EU AI Act (Regulation (EU) 2024/1689) to the extent applicable to safety-management at industrial establishments. Each one has a model card on file (available under NDA at ai@capptions.com) and is listed below in summary.
3.1 MAH Scenario Suggester
- Purpose: suggest plausible major-accident hazard scenarios for a given process unit, based on the tenant’s own equipment list and process descriptions.
- Inputs: tenant-scoped process descriptions, equipment list, prior scenarios.
- Outputs: a ranked list of candidate scenarios with rationale and references to the source records.
- Human oversight: every suggestion is non-binding until a process safety engineer accepts it. Accepted scenarios become draft records.
- Failure mode: false negatives are possible. The UI explicitly states “this list is not exhaustive — you remain responsible for scenario identification.”
- Validation: evaluated against an internal benchmark of redacted real-world Seveso scenarios; precision and recall reported quarterly in the model card.
3.2 Safety Report Drafter
- Purpose: generate first-draft text for chapters of the Seveso Safety Report aligned to Annex II of Directive 2012/18/EU.
- Inputs: tenant-scoped records (process, organisation, MAH analysis, barriers, MOC log).
- Outputs: draft prose, clearly labelled “AI draft — not yet reviewed.”
- Human oversight: drafts cannot be promoted to a controlled Safety Report revision without dual sign-off (author + safety manager).
- Validation: factual-grounding checks; the draft cites the source records used.
3.3 Barrier-and-mitigation Reviewer
- Purpose: cross-check claimed barriers against the underlying hazard scenario, flag missing layers of protection.
- Inputs: tenant-scoped hazard records and barrier records.
- Outputs: a list of flagged inconsistencies with a rationale.
- Human oversight: flags are advisory. Closure requires a human decision recorded in the audit log.
3.4 Inspection prep assistant
- Purpose: prepare an internal team for a Seveso authority inspection by drafting briefing notes from the tenant’s own records.
- Inputs: tenant-scoped records.
- Outputs: briefing notes labelled as AI-generated.
- Human oversight: notes are advisory, not part of any regulatory filing.
4. EU AI Act classification
| Feature | Class | Obligations met |
|---|---|---|
| MAH Scenario Suggester | High-risk (Annex III) where used in a safety-management decision | Art. 9–15, 50; model card on file |
| Safety Report Drafter | High-risk (Annex III) where output is intended for the official Safety Report | Art. 9–15, 50; model card on file |
| Barrier-and-mitigation Reviewer | High-risk | Art. 9–15, 50; model card on file |
| Inspection prep assistant | Limited-risk | Art. 50 (transparency) |
| General Q&A and summarisation | Limited-risk | Art. 50 (transparency) |
Capptions does not deploy any practice prohibited by Art. 5 of the EU AI Act.
5. Controls available to you
- Disable AI tenant-wide from Settings → AI.
- Disable AI per role (read-only users get AI off by default).
- Export all AI logs for your tenant in JSON.
- Delete all AI logs for your tenant; this does not affect the records the AI helped you produce.
- Pin the provider for your tenant (e.g. request an EU-sovereign provider only) for procurement reasons.
6. Incidents and your route to us
If you suspect an AI-related incident (a wrong safety suggestion, a prompt-injection attempt against your tenant, suspected data egress), email ai@capptions.com. We treat AI incidents under the same incident- response process as security incidents (see the Security Statement), with a 48-hour notification commitment for confirmed data-protection incidents.
7. Updates
We will publish a 30-day notice before changing the default model, provider or region for any high-risk feature in production. Notice goes to tenant administrators by email and is posted on the Sub-processors page.