Security and models
Security and governance
Nine control layers that together decide what a user and AI can see and do.
Governance in Slimaco isn't a single panel. It's a set of layers that together limit access to content, models, tools and actions.
Governance layers
| Layer | Control question | Mechanism |
|---|---|---|
| Identity | Who is the user? | Login, session, user identifier and tenant |
| Role and permissions | Which features can they see and run? | Built-in roles, custom roles and permission codes |
| Plan and entitlement | Does the organization have this feature? | Plan catalog, limits and feature availability |
| Knowledge access | Which sources and documents can they see? | Rules at the collection, source and document level |
| Assistant scope | What knowledge and tools does this assistant have? | Knowledge scope, tool selection and visibility |
| Skill scope | What procedure and tools can be applied? | The instruction plus a narrowing tool list |
| Guardrails | Is the content or action safe? | Pre- and post-model checks and the fail-closed principle |
| Action consent | Can AI perform a write? | Per-call approval, thread consent or time-boxed consent |
| Audit | What ran, and with what result? | Safe metadata, correlation and access events |
| External AI (MCP, coming soon) | Which AI tools may use company knowledge? | The admin enables the connection and picks allowed tools; the asking person's permissions apply |
The fail-closed principle
If the platform cannot confirm a plan, permission, scope, safety check or the availability of a key control mechanism, it should refuse or limit execution rather than default to granting access.
Content protection and telemetry
- Raw document content never reaches event queues, logs or metrics.
- Events carry references and safe statuses instead of full content.
- Identity passed downstream is reduced to the necessary context, without raw tokens or full identity-provider data.
- The vector index is used to retrieve approved knowledge, not as a source of truth for documents or audit.
- Guardrails are designed to fail closed; disabling them is intended only for local or test environments.
External AI tools via MCP
Soon ChatGPT, Claude, Gemini and Microsoft Copilot will be able to use Slimaco knowledge via MCP. Slimaco then passes only the knowledge fragments needed for the answer, which the asking person is allowed to see, with personal data masked. From there, the organization's agreement with that tool's vendor applies, including its processing and data-location terms; the EU processing commitment covers the Slimaco assistant's default model, not external tools. The organization's admin decides whether the connection is enabled and which tools may use it. We recommend connecting company accounts, not private ones. Details: Your AI via MCP.
Retention and deletion
Content that a company enters into the platform (conversations, documents, sources) belongs to the company. The company decides how long it is kept and when it is deleted. Deleting a thread is permanent in the platform; archiving is reversible. Withdrawing a document is reversible; a permanent deletion (purge) is not. Technical copies (logs, diagnostic traces, backups and, for content from connected accounts, data held by the integration intermediary) expire according to the periods in the table below, not immediately on deletion.
Periods we set:
| Type of data | Retention period |
|---|---|
| Conversations, documents, AI agent memory | Set by the company (Customer). After the agreement ends, the company can download its data within 30 days; within the same period we delete it from the platform databases. |
| Usage analytics archive (company, model, tokens, cost, time; user identifier pseudonymised) | 13 months |
| Backups | Database: 7 days (point-in-time recovery). Disk snapshots: 14 days. Database dumps are sometimes kept longer (up to about 60 days); we are shortening them to 7 days. |
| Security and technical logs | 30 days; sign-in events 90 days |
| User account data | Until the account is deleted or the agreement ends |
Retention periods for data we control are described in the Privacy Policy.
Personal data and the data processing agreement
We process personal data that a company enters into the platform as a processor, under a data processing agreement (DPA). The company remains the controller of that data. We publish the list of providers we entrust with processing (sub-processors) on the Sub-processors page and notify customers of changes. Access to one person’s data (export) and deletion of their data at the company’s request are handled within the periods set in the agreement.
Feature in preparation. Until it is released, we handle this only for a User’s account and their threads. We cannot yet fully handle requests concerning third parties mentioned in content who have no account of their own, because we don’t have a tool to find every mention of such a person.
Deployment model
Slimaco AI runs primarily as a SaaS service, but the platform can also be deployed on-premise or in dedicated client infrastructure — both modes are available, and the exact model, processing region and way of integrating with AI models are agreed individually as part of the deployment conversation and contract.