AI Transparency

Last updated: July 11, 2026

This page is Declogger's AI system card. It explains, in plain language, what the AI inside Declogger does, what it doesn't do, and how we comply with the transparency obligations of the EU Artificial Intelligence Act (Regulation (EU) 2024/1689), in particular Articles 13 and 50.

1. Intended purpose

Declogger is a decision-support system for governance, compliance, and institutional memory. It helps humans capture decisions, link them to documents and external resources, and surface conflicts, opportunities, and patterns. Declogger does not make decisions for the user, and it does not auto-execute actions.

2. Capabilities

Conflict & opportunity detection — when a decision is saved, edited, or a new legal document is uploaded, Declogger searches the Knowledge Base for items that may be in tension with — or amplify — the new content, and flags them as suggestions.
Ask bar — a natural-language search/QA interface over the organization's Knowledge Base (Standard tier and up).
Summarization & extraction — converting transcripts (Teams, WhatsApp) into decision candidates, and summarizing long documents.
Auto-capture triage — classifying short inbound messages from Slack, Jira, Linear, Asana, and inbound webhooks (as well as Teams and WhatsApp) into Knowledge Base proposals for Client-Admin review, and posting an in-channel reply back to the source with related knowledge and any detected conflicts. Off by default per source.
Detection tuning — Client Admins can adjust strictness, add custom anchors, and "teach" Declogger via examples; tuning is applied to subsequent analyses.

3. Known limitations

Detection quality depends on the size and quality of the Knowledge Base. Thin inputs (very short titles, single-line decisions) are intentionally short-circuited to avoid spurious matches.
The system is optimized for English and other major European languages. Less-resourced languages may produce weaker results.
False positives can still occur. Every finding requires a human to accept, dismiss, or overrule it.
The Ask bar answers from the Knowledge Base only — it does not browse the public internet.
Org Units, when configured, sharpen recall and precision; they do not replace human review.
Confidential decisions are excluded from cross-user analysis, including when Org Units are considered.

4. Human oversight

Oversight is built into the workflow, not optional:

Detected conflicts and opportunities are suggestions. Nothing is applied to a decision without a user action.
Decision-makers and Client Admins can overrule any finding with a reason captured in the audit trail.
Contributor saves and overrules go through an approval queue — a human approver is required.
Client Admins can disable AI features per organization on the Privacy screen.

5. Logging & traceability (Art. 12)

Every model call routed through the Lovable AI Gateway is logged in the ai_usage_events table with: model used, operation, token counts, credit cost, latency, and the calling organization. Platform Admins can review usage per organization and per period.

6. Models used

Declogger uses Google Gemini 3 Flash Preview via the Lovable AI Gateway as its default model for detection, summarization, and the Ask bar. The model provider does not train on Declogger customer data; payloads are not retained for training. Some background tasks may use embeddings or text-extraction models; see Privacy policy.

7. Risk classification

Declogger is, in normal use, a limited-risk AI system under the EU AI Act: it supports human decision-making but does not determine outcomes with legal or similarly significant effects on natural persons. Customers must use the system within the bounds set by our Acceptable Use Policy, which excludes the uses listed in Annex III of the Act (hiring/firing, credit scoring, critical infrastructure, law enforcement, biometric categorization, social scoring, etc.).

8. Transparency to end users (Art. 50)

Users always know when they are interacting with AI: the Ask bar, AI-generated findings, and AI-suggested decision candidates are clearly labelled as such in the interface. We do not deploy AI in ways that pretend to be human.

9. Reporting concerns

If you believe Declogger is producing harmful, biased, or otherwise unsafe output, email compliance@declogger.com with a description and, if possible, the affected decision ID. We investigate every report.