---
title: "How we use AI"
description: "How Conphora uses AI — which agents produce what, who reviews it, which providers process data, and what our tools never decide"
type: "legal"
url: "https://www.conphora.com/ai-transparency/"
markdown: "https://www.conphora.com/ai-transparency.md"
locale: "en"
language: "en"
updated: "2026-09-18"
author: "Conphora"
publisher: "Conphora"
license: "Free to quote with attribution and a link to the page"
---

# How we use AI

> How Conphora uses AI — which agents produce what, who reviews it, which providers process data, and what our tools never decide

<!-- LEGAL DRAFT HEADER
Purpose: The public AI transparency notice required by SPEC-78 R11 / D4 and SPEC-57 R11.
  It is the proof page behind the footer's "EU AI Act" line: what AI does at Conphora, who
  reviews it before anyone sees it, which providers process which data, and what AI never
  decides on its own. The footer link (AI_TRANSPARENCY_PATH in src/components/Footer.astro)
  is switched on by Claus's team only after this page is approved.
Jurisdiction: Danish law (default, SPEC-57 D6). EU acts apply directly; national overlays
  named where they exist (Danish Marketing Practices Act). No deviation intended.
Language/master: EN master · DK mirror queued (Lína, Mon 2026-09-21) — Claus approves both
  before the footer links them.
Audience: public (visitors, Academy learners, platform customers, partners)
Assumptions (verify before approval):
  1. None of Conphora's AI systems is high-risk under Reg. (EU) 2024/1689 Annex III, and none
     falls under the prohibited practices of art. 5. Conphora acts as a *deployer* of
     general-purpose models supplied by Anthropic, not as a provider of an AI system placed
     on the market — this changes if NOMOS (SPEC-58) or a tool is offered to third parties
     under Conphora's own name (art. 3(3), art. 25).
  2. Every published regulation guide, article and press release is read and signed off by a
     named human before publication (guide frontmatter `reviewer` / `reviewedAt`; press
     releases through Claus).
  3. Our model provider does not train on the text we send it, and retention is limited to
     what abuse monitoring requires. The contractual wording is still to be confirmed in
     writing (sub-processors page, Anthropic row: "zero-retention API terms to confirm").
  4. No AI system at Conphora makes a decision that produces legal effects concerning a
     person or similarly significantly affects them (GDPR art. 22).
  5. The agents listed below are the complete set that can touch published content or
     customer-visible output as of 2026-09-18 (admin/data/agents.builtin.json +
     agents.custom.json).
Open points for counsel:
  1. Does our published compliance content count as "text published with the purpose of
     informing the public on matters of public interest" under art. 50(4), and is our human
     review + named editorial responsibility enough to rely on the exemption in the second
     sentence of art. 50(4)? If not, every guide and article needs a visible AI-generated
     label, not only this notice.
  2. If Conphora offers NOMOS or a free tool to third parties under its own name, does it
     become a provider under art. 3(3) and owe art. 50(1) (inform the user) and art. 50(2)
     (machine-readable marking of synthetic output)? What is the technical marking we would
     have to implement, and by when?
  3. Art. 4 (AI literacy) applies to deployers since 2 February 2025 — is our internal
     instruction set (docs/agentinstruction/) sufficient documentation, and does it need a
     named responsible person and a record?
  4. Confirm the no-training and retention terms with Anthropic and Voyage AI in writing,
     and whether the transfer mechanism named on the sub-processors page still holds.
  5. The greenwashing and recall tools analyse third parties' public pages and name them in the
     result. Exposure under the Danish Marketing Practices Act §§ 5, 20 and under defamation
     law: is the wording of the tool disclaimer (docs/legal/templates/tool-disclaimer.md)
     enough, and should results naming a company be non-public by default?
  6. Whether this notice must also state a contact point for AI-specific complaints separate
     from the privacy contact, and whether the market surveillance authority for the AI Act
     in Denmark should be named once it is designated.
Sources:
  - Regulation (EU) 2024/1689 (AI Act), CELEX 32024R1689 — art. 3(3)–(4), art. 4, art. 5,
    art. 50(1), (2), (4), (5), art. 113 (art. 50 applies from 2 August 2026)
  - Regulation (EU) 2016/679 (GDPR), CELEX 32016R0679 — art. 5, 6, 13, 22, 28, 32
  - Directive 2002/58/EC (ePrivacy) art. 5(3) and the Danish Cookie Order (bekendtgørelse
    nr. 1148 af 09/12/2011) — see the cookie policy; this notice adds nothing to it
  - Markedsføringsloven (lov nr. 426 af 03/05/2017) §§ 5, 13, 20 — misleading practices and
    the duty to document factual claims
  - Commission Decision 2011/833/EU, CELEX 32011D0833 — reuse of Commission documents;
    EUR-Lex reuse notice. Harmonised standards texts are not reproduced (copyright).
  - Internal: SPEC-57 (Sigrid), SPEC-58 (NOMOS), SPEC-78 R11/D4, SPEC-83 (translation
    pipeline), ADR-020 (confirmation guardrail for production writes),
    content/legal/sub-processors.md, content/legal/privacy-policy.md
Prepared by: Sigrid (AI) on 2026-09-18 — AI-prepared, not legal advice; review by counsel required
-->

**Applicable to:** conphora.com, Conphora Academy and the Conphora platform

Our content and our tools are AI-assisted and reviewed by a named person before you see them. Every regulation page cites the official text on EUR-Lex. The output of our tools is labelled and is never legal advice.

This notice explains what that means in practice: where we use AI, what a human checks, which providers process which data, and what AI at Conphora never decides on its own. It is the AI transparency information required of us as a deployer of AI systems under the EU AI Act (Regulation (EU) 2024/1689, art. 50, applicable since 2 August 2026) and it supplements — it does not replace — our [Privacy Policy](/privacy-policy), [Cookie Policy](/cookie-policy) and [Sub-processors](/sub-processors) page.

### 1. Where we use AI

We use large language models as assistants. Each assistant has a written instruction, a defined area and a named human who is responsible for what it produces.

**Published content on conphora.com**

- **Regulation guides.** Drafted from the official act on EUR-Lex and the Commission's guidance, with a source for every fact, then checked and signed off by our compliance editors. Each guide names its reviewer and the date of review, and links the CELEX record of the act it describes. We do not reproduce the text of harmonised standards.
- **Knowledge articles and the Academy.** Drafted with AI support from the same sources and edited by a person before publication.
- **Press releases and news.** Drafted by our compliance journalist assistant from an outline, then fact-checked and released by a named person at Conphora. Nothing is published automatically.
- **Translations.** Our pages are translated by an AI localisation pipeline into the languages we publish, with a per-language glossary, a proofreading pass and an expert review before a translated page goes live. The English text is the master; where a translation and the English master differ, the English master governs.
- **Compliance Radar.** Public sources (EUR-Lex, the Official Journal, authority pages, the EU Safety Gate) are monitored and triaged by AI. A radar item is reviewed by a person before it is published or sent in an update e-mail.
- **Interest updates (newsletter).** The weekly digest is assembled from reviewed items. It is not personalised profiling: it contains what you chose to follow.

**Tools and the platform**

- **Free tools** (for example the greenwashing check and the recall check) analyse publicly available web pages and the text you enter, and return an assessment with its sources. The result is an indication for your own review, not a verdict about a company or a product.
- **The Conphora platform** uses AI to classify products and documents, to summarise them and to search them. Classifications and summaries are suggestions that the user in your organisation confirms or corrects; they are not a determination that a product is compliant.
- **NOMOS**, our compliance specialist agent, answers questions with citations to the underlying act and an indication of how confident the answer is, and declines questions outside its scope. Every answer carries the same boundary: information, not legal advice.

**Internal use.** We also use AI assistants internally for work you never see — code and security review, quality and link checks, market monitoring, planning and the preparation of legal drafts such as this one. They produce no published content and make no decisions about people.

### 2. What a human always does

- **Publication.** No AI assistant publishes to conphora.com by itself. Work is handed to a person, who reviews it and merges it. Regulation guides additionally record the reviewer and the review date on the page.
- **Customer-visible output.** AI output in the platform is presented as a suggestion with its sources; a user confirms it.
- **Production changes.** Any action of an assistant against our production systems requires an explicit human confirmation; without it the action is refused.
- **Editorial responsibility.** Conphora ApS is responsible for everything published on our pages, whether or not an AI assistant helped write it.

### 3. What AI never does here

- It never makes a decision about you that has legal effects or similarly significantly affects you. We do not use automated decision-making or profiling in the sense of GDPR art. 22.
- It never gives legal advice. Our pages, tools and answers are information about the law; whether a rule applies to your product is a judgement you — or your adviser — make.
- It never decides that a product is compliant, or issues, signs or replaces a declaration of conformity.
- It never scores, ranks or profiles individuals, and it does not infer emotions or categorise people.
- It never publishes or sends anything to you without a human having released it.

### 4. Which providers process data, and what we do not allow

The models we use are supplied by **Anthropic** (Claude) and, for search on the platform, embeddings from **Voyage AI**. They act as our data processors under a data processing agreement and process only what we send them for the task at hand. The full list, with entity, location and transfer mechanism, is on our [Sub-processors](/sub-processors) page.

- We do not allow our model providers to use the text we send them to train their models.
- We keep what we send to a model to what the task needs. Prompts and answers of our own agent runs are logged in our own systems in the EU so that a run can be audited; the logs are kept as described in section 7 of the [Privacy Policy](/privacy-policy).
- Customer documents on the platform are processed for classification, summarisation and search on your organisation's instructions and under the data processing agreement, and are deleted according to it.
- What you type into a free tool is processed to produce the result. We do not use it to build a profile of you.
- Publishing an AI-assisted page involves no personal data of visitors at all.

The legal basis, the retention periods and your rights are set out in the [Privacy Policy](/privacy-policy).

### 5. Limitations — and how to tell us we got it wrong

AI-assisted text can be wrong, out of date or incomplete, and a model can state something confidently that is not correct. That is why every factual claim on our regulation pages carries its source, why a person reviews the page, and why we date it.

- Read the source. Every regulation page links the act on EUR-Lex; the consolidated text there is the authority, not our summary.
- Tool results are indications. They tell you where to look; they do not conclude anything about a company or a product.
- The law changes. A page is accurate as of the review date shown on it.
- Denmark is our default jurisdiction where a national rule matters; other Member States may apply the same EU act differently.

Found an error, or want to know how a specific result was produced? Write to [info@conphora.com](mailto:info@conphora.com) with the page or the result. We correct errors on the page and note the correction. If your question is about personal data, section 9 of the [Privacy Policy](/privacy-policy) applies.

### 6. Contact

Conphora ApS
Copenhagen, Denmark
CVR: 43489968
E-mail: [info@conphora.com](mailto:info@conphora.com)
Phone: +45 71 74 11 12

This notice is reviewed when we add an AI assistant, change a model provider, or when the rules change. It was last updated on the date shown at the top of the page.

<!-- AI-forberedt — ikke juridisk rådgivning; gennemgang af advokat påkrævet. -->