EU AI Act: how Answely uses artificial intelligence

1. Purpose of this page

The AI Act (Regulation (EU) 2024/1689, "AI Act") entered into force on 01.08.2024. Its prohibitions (Art. 5) and the AI literacy obligation (Art. 4) have applied since 02.02.2025, the obligations for providers of general-purpose AI models since 02.08.2025 and the transparency obligations of Art. 50 since 02.08.2026. Under the "Digital Omnibus", postponements of individual deadlines have been proposed, for example for machine-readable marking under Art. 50(2); we already implement this marking today regardless (section 5.2).

At its core, Answely is an AI product: the answer box on our customers' websites answers visitors' questions with texts generated by a language model. This page explains in plain language exactly where AI is involved, what role we and our customers have under the Regulation, which obligations follow from this and where our implementation still has gaps today. It supplements the Privacy Policy, the Terms of Use and the Data Processing Agreement.

The short version. Answely generates answers with freely available language models, which we operate ourselves on own hardware in Łódź, Poland – our own AI, self-hosted. The model may only answer from excerpts of the customer's website and must substantiate every statement with a source; if no evidence is found, no answer is given. Every answer is visibly labelled as AI-generated, and this notice cannot be switched off. We do not train models with customer data. Answely does not make decisions about people and is not a high-risk system. In addition, every answer is marked as AI-generated in a machine-readable form (Art. 50(2), section 5.2).

2. Where Answely uses AI – and where it does not

2.1 Generative AI: the language model

• Answers in the box. For a visitor's question, Answely searches the website's index for matching sections and passes some of them to the language model as numbered excerpts. The model formulates a short answer from them (about 120 words) and must substantiate every statement with the number of its excerpt.
• Follow-up questions and targeted clarifying questions. For a follow-up question, the previous conversation is included as context. Clarifying questions that the customer stores as a rule are fixed texts and are not generated by the model.
• Public website quick check ("Check your website"). Anyone checking a website on answely.com can ask a sample question; for this, the model generates at most one answer from a few public pages. None of this is stored permanently.
• Answer tests, answer diagnostics and answer comparison. In the account, the customer can store its own test questions, trace an individual answer step by step or compare answers from the active index and a new index side by side. Real model answers are generated for this, which only the customer sees.
• Support inbox (currently off). The account software contains a function with which a staff member can have incoming support messages translated or a draft reply generated; nothing would be sent until a human has read and sent the draft. At Answely this function is not switched on: no machine is connected to collect such jobs. Before we switch it on, we will update this page and the privacy policy.

2.2 AI without text generation: embeddings

• Search in the index. Sections of the website and visitors' questions are converted into numerical vectors with an embedding model of our own, likewise self-hosted, so that Answely finds sections that match in content even when different words are used. In addition, a classic full-text search is used. A model for re-ordering the hits (reranker) is prepared but currently switched off.
• Grouping in question analysis. If the customer has switched on question analysis, Answely groups similar, previously masked questions once a night using the same embedding model. The report only shows groups, never individual questions.

Embeddings do not generate any text that a human reads. They determine which excerpts the language model receives and which questions are counted together.

2.3 No AI

• Indexing. Retrieval, splitting into sections, respecting robots.txt, exclusions and noindex.
• Text recognition in PDF files (OCR). For PDF files without extractable text, the customer can switch on text recognition. It runs locally with classic tools (Poppler, Tesseract), generates no new content and is not generative AI.
• Rule adviser, action links, area selection, search glossary, source rules. Fixed rules set by the customer.
• Documentation copilot. It compiles existing text passages with references into a draft (extractively), which appears only in the customer's browser and is not published.
• Public MCP search access. A pure text search in the public index without any model call.
• Figures and reports. Counters, quotas, success measurement and answer test reports are counted, not estimated.

3. Which models we use

Model	Purpose	Origin and licence	Where it runs
Language model (our own AI, self-hosted)	Formulating answers	freely available model with open weights, which we operate ourselves on our own hardware; we use a version reduced by a third party at a pinned revision	two own computers in Łódź, Poland; if one is busy or fails, the other takes over
Embedding model (an embedding model of our own, likewise self-hosted)	Embeddings for search and grouping	freely available model with open weights	own servers in the same internal network
Reranker	Re-ordering of hits	prepared	currently switched off

The models run on our own computers; we also use these computers for other services of our own. A further computer of our own does not formulate answers; it handles background work such as the embeddings during index runs and the nightly grouping. Answely passes them text only: the image component that comes with the model software is switched on, but Answely sends no images. Questions, excerpts and answers are not sent to any external AI provider. The main server never calls the AI computers itself: programs on the AI computers fetch jobs via a signed connection and return the result. The service only starts if every configured AI endpoint is on an allowlist of servers in the EU.

We will only use a further AI provider if it processes in the EU, contractually neither trains nor stores beyond the processing, is listed among the sub-processors and you have been informed at least 30 days in advance. Such a fallback provider is technically prepared but not active. Providers outside the EU are excluded.

4. Our role under the Regulation

4.1 Answely is the provider of an AI system

The answer box together with the answer service is an AI system within the meaning of Art. 3(1): it infers outputs (an answer) from inputs (question and excerpts) that influence the visitor's environment. We develop this system and place it on the market under the name Answely. We are therefore the provider (Art. 3(3)), and the obligations of Art. 50(1) and (2) are addressed to us.

4.2 Our customers are deployers

Anyone who integrates Answely on their website uses the AI system under their own authority and is therefore a deployer (Art. 3(4)). Art. 50 provides for separate obligations for deployers, above all for deep fakes and for texts published with the purpose of informing the public on matters of public interest (Art. 50(4)). Answely generates neither images nor audio nor video. Whether answers in the box count as such published texts depends on the website; for the usual cases (product questions, documentation, support), we consider this remote. The labelling that cannot be switched off (section 5) covers disclosure in any case.

If a customer changes the purpose in such a way that the box is used in a high-risk area (section 4.4), it may itself become the provider of a high-risk system under Art. 25. Such use is not covered by our Terms of Use.

4.3 The original publisher is the provider of the model

The language model we use is a general-purpose AI model (Art. 3(63)). The obligations for the model itself (Art. 53, for example technical documentation and a summary of the training data) lie with its original provider; for models released under a free and open-source licence, the exemptions under Art. 53(2) apply. We do not modify, train or fine-tune the model and therefore do not ourselves become a model provider. The reduced (quantised) version we use was not created by Answely but published by a third party. In our understanding, the original publisher remains the provider of the model in that case as well; Answely does not become the provider by using this version.

4.4 No high-risk system, no prohibited practices

Risk classification. Answely does not fall within any area of Annex III (biometrics, critical infrastructure, education and vocational training, employment, access to essential services such as credit or insurance, law enforcement, migration, administration of justice and democratic processes) and is not a safety component of a product under Annex I. The box answers questions about a website's public content. It does not assess persons, does not grant benefits and does not decide on access to anything. We therefore do not classify Answely as a high-risk system.

Application block (since 16.10.2026). So that the box cannot become a tool for selecting applicants after all (Annex III point 4), it is technically blocked: questions by which someone submits a job application, CV or cover letter, wants it reviewed or assessed, or asks about their suitability for a position are detected by Answely before any model call. In that case no model is asked, nothing is cached and nothing is taken into question analysis; the box replies with a fixed text and refers to the company's own application channel. Questions about open positions and how to apply remain allowed. In addition, a fixed rule in the model's system text prohibits any acceptance or assessment of applications. The block applies to every path that generates an AI answer (box, answer tests, diagnosis, answer comparison, sample question in the website check) and can be switched off neither by the customer nor by the operator. Detection is pattern-based and cannot catch every wording.

Prohibited practices (Art. 5). Answely does not perform emotion recognition, biometric categorisation, social scoring or prediction of criminal offences, and does not use subliminal or manipulative techniques. The answers are guided by the content of the website, not by characteristics of the visitor; Answely knows neither the visitor's name nor any profile of the visitor.

5. Transparency (Art. 50)

5.1 Recognisable that an AI is answering (para. 1)

Directly below the title, every answer box bears the notice "AI-generated answer – may be incomplete; the sources are authoritative." (correspondingly in German and Polish). The notice appears even before the answer becomes visible.

• Cannot be switched off. No setting and no plan hides the notice. It is protected against the website's own CSS by fixed style declarations, and the box shows no answer if the notice is nevertheless not visible in the browser.
• Own wording only from Pro and only if it still labels. A custom text must contain an AI keyword (for example "AI", "KI" or "artificial intelligence") and must not negate it. Answely rejects unsuitable texts and shows the standard text.
• Customer obligation. The customer must not integrate the box in such a way that visitors can no longer recognise its AI nature (clause 17.2 of the Terms of Use).
• Accessibility. The notice is marked up as a text paragraph with a role for assistive technology and is read out by screen readers. Acceptance testing with a real screen reader is still pending.

5.2 Machine-readable marking of generated texts (para. 2)

Art. 50(2) requires providers to ensure that generated texts are marked in a machine-readable format and detectable as artificially generated, insofar as technically feasible. In addition to the visible notice, Answely marks every answer in a machine-readable form in three ways: the element of the box that contains the answer carries the attribute data-ai-generated="true"; the final fertig event of the application programming interface contains the field ki_generiert: true; and every answer of the application programming interface carries the header X-Answely-AI-Generated: true, which a script on the customer's website can also read (exposed via CORS). This also applies to editorial answers, because the box labels them in the same way. Anyone who retrieves answers via the application programming interface and displays them themselves receives the field and the header with every answer. There is no watermark in the answer text itself; none is provided for plain text.

It is doubtful whether the exception for systems that do not substantially alter the input data provided by the deployer applies: the answers summarise the website's content but reformulate it. We therefore assume that para. 2 applies to Answely. We will review this classification and the form of the marking as soon as the European Commission publishes guidelines or a code of practice on it.

5.3 What Answely does not generate

Answely does not generate images, audio or video and therefore no deep fakes. Answely does not pretend to be a human and does not conduct a conversation that simulates a human interlocutor.

6. No answer without evidence

The most important safeguard against fabricated answers is the architecture itself:

• Evidence check before the model. If the best section found is below a fixed threshold, the language model is not called at all. The visitor then receives the notice that the website contains no substantiated answer on this.
• Only from the excerpts. The instruction to the model prohibits its own knowledge. If the fact sought is missing from the excerpts, the model must decline to answer.
• Citation requirement. Every statement must bear the number of its excerpt. Answers without a valid source marker are discarded. The cited pages are listed as links below every answer.
• Content is data, not commands. Text on the website and in the question is expressly passed to the model as data, not as an instruction. This reduces the risk that someone redirects the model's behaviour via a question or a manipulated page; it cannot be ruled out entirely.

The limits, stated plainly. The source marker proves that a statement refers to an existing excerpt, not that it reproduces it correctly. A language model may omit a condition, overlook a negation or combine two excerpts incorrectly. The answer appears while it is being written; the final check of the source markers takes place at the end. A second, stricter check is prepared but not active in the default setting: in it, the same language model would check in a further call whether every statement is supported by its cited excerpt, and the answer would only appear afterwards. This also applies in live operation. That is why the notice in every box says that the sources are authoritative, and that is why, under clauses 17.4 and 21.2 of the Terms of Use, we are not liable for decisions based solely on an answer.

7. Human oversight

Answely provides information that a human reads and can verify. Oversight of the content lies with the customer, who knows its website; we provide the customer with tools for this:

• Answer tests. Up to ten own test questions with expected keywords, sources or an expected refusal are run again after every change to the index or configuration. The report shows which cases have passed.
• Answer diagnostics. For an individual question, Answely shows which sections were found, which were selected and passed to the model, and why an answer was declined.
• Editorial answers. The customer can store its own answer with a source for specific questions, block topics and prioritise sources.
• Index approval. On request, a new index only becomes active once the answer tests have been passed; the previous state is retained until the next approval and can be restored.
• Exclusions and source rules. Pages and paths can be excluded, limited in time or weighted higher or lower.
• Ratings and question analysis. If the customer has switched them on, it sees where visitors receive no answer or rate an answer as unhelpful.
• Switching off. The customer can switch off the box at any time; its normal search continues to work. We ourselves can suspend individual websites or all boxes at once.

8. No automated decisions

Answely does not make any decision that produces legal effects concerning a visitor or similarly significantly affects them (Art. 22 GDPR). An answer is information from the website, not a commitment, an assessment or a refusal. Action links and support links only open when the visitor clicks. If a customer wishes to use Answely for questions whose answers may have significant consequences for visitors, for example on contractual rights, health or deadlines, it should maintain the sources with particular care and design the answer tests accordingly.

9. Data and no training

• No training, no fine-tuning. Neither website content nor questions nor answers are used to train or fine-tune a model, neither by us nor by third parties. The weights of the models are fixed.
• What the model receives. The question (at most 300 characters), for follow-up questions the previous conversation, and excerpts from the website's index. No IP address, no visitor identifier, no cookies.
• The index is not training. The vectors in the index are a search aid for exactly this website. They are deleted together with the website.
• What is stored. In the default setting, questions and answers are not stored. What is stored, and for how long, when "Allow answer cache" or question analysis is switched on is set out in the Privacy Policy.
• Protected knowledge areas. For non-public content, a connection to the customer's sign-in system is prepared but not yet enabled with any provider. The visitor's credentials are not passed to the model.

10. Quality and measurement

We measure answer quality with fixed test sets on our own public reference website, without caching and with the real model. The test is automated: it evaluates whether the expected source is cited and expected keywords are mentioned, and whether questions with no answer on the website are declined. It is an indication, not a substantive review by a human.

Since the model change of 16.10.2026, we have measured a common test set of 178 acceptance questions with two models. With the model in use today, 91.6 % passed at first and 93.3 % after an adjustment of the answer rules; with the model kept as a reserve until 18.10.2026, 87.1 % and then 89.9 %. Questions to which the website contains no answer were still all declined with both models.

Earlier measurement, still with the model formerly used alone:

Test set (as of 09.10.2026, historical)	Source and keyword matched	Questions without an answer correctly declined
German, 50 known and 10 unknown questions	43 of 50 (86 %)	10 of 10
English, 27 known and 3 unknown questions	22 of 27 (81.5 %)	3 of 3
Polish, 27 known and 3 unknown questions	22 of 27 (81.5 %)	3 of 3

What these figures do not say. Our own target is a hit rate of at least 85 % per language; in the earlier measurement, English and Polish did not reach it, and we do not report the new measurement separately by language. The English and Polish test sets cover the same topics as the German one and are not independent data bases. A measurement with websites of pilot customers and a substantive evaluation by humans are still pending; we will only commit to a rate of 95 % after this pilot acceptance. Under a peak load of 480 requests per minute over five minutes, a large proportion of requests were rejected with a notice of overload; the necessary capacity has not yet been demonstrated. If the service is overloaded, no answer appears, rather than a worse one.

11. Logging

The answer service logs only technical codes and figures, never questions, excerpts or answers. The programs on the AI computers log only the type, duration and number of characters or vectors of a job, and error codes. Answer diagnostics are held only in working memory and are removed no later than 15 minutes after completion. Answer test reports contain case numbers and results, no question or answer texts. The record-keeping obligations of Art. 12 apply only to high-risk systems; we nevertheless adhere to this principle, because logs without content cannot harm anyone.

12. AI literacy (Art. 4)

Answely is developed and operated by a sole proprietorship. Anyone working on the service knows how the models work, the limits of the evidence check and the obligations under this Regulation; the measurements in section 10 are part of this work. There is no separate written documentation of AI literacy; what we know about the models, their limits and our measurements is set out on this page.

The developer documentation, the answer tests and the diagnostics help our customers to use the box competently. As deployers, they are themselves responsible for the AI literacy of their own staff.

13. Changes and contact

We update this page as soon as an AI function is added, changed or removed, a different model is used or the legal situation changes. We announce a further AI provider at least 30 days in advance. Please direct questions and concerns about our use of AI to support@answely.com. Complaints follow the procedure under Complaints; please report illegal content as described under Report abuse and illegal content.

Provider: LW IT Solutions Company Lukas Wójcik, al. Tadeusza Kościuszki 80/82, lok. 301, 90-437 Łódź, Poland.

Version 1.0 – as of 18.10.2026