Domiq

AI hazard assessment in Domiq

Edition 1 of 16 November 2026. Assessed by Firetech Engineering BG EOOD, the company that makes Domiq. The same content as open data: domiqteam.com/ai-assessment.json.

Why this page exists.

The Digital Code of the Kyrgyz Republic (Article 193) requires every artificial intelligence system used in the country to be assessed for the hazards its use may create, and the result and the method to be published in plain words. Here are all the Domiq features where AI is at work.

How we assessed.

For each feature we answer six questions: what it does; what is sent to the model; who makes the decision; what can go wrong and who is harmed — people's privacy, their money, the safety of children, the relationship between employer and employee; how likely and how serious that is; and what has been done to prevent it.

Level: low — a mistake is visible and a person corrects it before it affects anything; medium — a mistake may go unnoticed and lead to a missed task or to information being disclosed; high — a mistake or a leak may seriously harm a person.

A feature is auxiliary (Article 194(3) of the Code) if the AI only suggests and a person decides and acts; feature 10 in working mode is the one exception, and it is assessed separately.

As of the assessment date we found no requirements of the Cabinet of Ministers for the method, so the method is our own.

We review the assessment when we add an AI feature, change the model or the provider, or learn of a mistake that changes the result.

What all features have in common.

Most models run at Groq in the United States: openai/gpt-oss-120b for text, whisper-large-v3-turbo for speech and qwen/qwen3.8-27b for document photos. Our organisation has full zero data retention enabled at Groq: requests and responses are not kept after processing and are not used for training. Replies to messages to support are prepared by claude-opus-5-5 at Anthropic in the United States (feature 10). Every feature except replies to support works only with the consent of the person using it — to text or to voice; residents of countries whose law requires consent for a transfer to the United States, Kyrgyzstan included, give that consent first. Without consent, or when the model is unavailable, a fallback without AI works instead — parsing on the phone or ready-made templates — and a draft made from templates is labelled as such. On the server in Russia Groq is not used: the AI assistant, voice input and photo-based form filling there, once switched on, work through Cloud.ru in Russia, and the data does not leave the country.

1. Task from a phrase.

What it does: turns a typed or spoken phrase into a draft task — title, date, time, assignee, checklist.

What is sent: the text of the phrase; names, property names, addresses, phone numbers and emails known to the app are replaced with placeholders.

Who decides: a person — the draft opens in the form, and they save it.

What can go wrong: the wrong day, time or assignee, so that a task — collecting a child, for example — goes to the wrong person or the wrong day.

What has been done: the model's answer is checked against the format and the calendar, and anything invalid is dropped; the whole draft is visible before saving.

Level: low. The feature is auxiliary.

2. Speech recognition.

What it does: turns a voice recording into text.

What is sent: the audio recording as it is — names cannot be replaced in it.

Who decides: a person sees the result before saving and corrects it.

What can go wrong: a wrong name, number or time; everything said aloud reaches the provider, including other people's names and private matters.

What has been done: a separate consent to voice; the provider does not keep the recording after processing; the privacy policy says plainly that sensitive matters are better not dictated.

Level: low for mistakes, medium for privacy. The feature is auxiliary.

3. Instruction draft.

What it does: suggests a draft instruction on request — «how to look after the pool», for example.

What is sent: the request and the property name as typed, without placeholders.

Who decides: a person edits and saves it; the draft is labelled «This draft was written by artificial intelligence — check it before saving».

What can go wrong: an inaccurate step in a safety instruction — electricity, chemicals, childcare — if nobody checks it.

What has been done: the label on every draft and a check of the answer's format; without the model, ready-made templates labelled «This draft was built from ready-made templates — check it before saving».

Level: low; medium for safety instructions, which is why the «check it» label is always there. The feature is auxiliary.

4. Shift report.

What it does: arranges the report an employee dictated or typed into lines, one task per line.

What is sent: the report text as written, without placeholders.

Who decides: the employee sees the result before sending and can correct it or bring back their own text with «Restore my text».

What can go wrong: the employer reads something other than what the employee reported.

What has been done: the label «AI has tidied up the text — check it and edit if needed»; an answer much longer than the original is rejected; editing by hand stops the AI; without the assistant the report is sent as written.

Level: low. The feature is auxiliary.

5. Day summary.

What it does: writes the summary of the day in one or two sentences.

What is sent: only numbers — how many tasks, how many done, how many shifts — without names or titles.

Who decides: the summary is for reference, and the app decides nothing on its basis.

What can go wrong: a wrong figure in the text.

What has been done: an answer with a number that is not in the input is rejected, and the summary is then built from the numbers without AI; an «AI» mark is shown next to it.

Level: low. The feature is auxiliary.

6. Support bot.

What it does: answers questions about using Domiq.

What is sent: the question and up to eight earlier messages of the conversation; names and addresses known to the app, as well as detected phone numbers and emails, are replaced with placeholders.

Who decides: a person; before the chat it says «You are talking to artificial intelligence, not a person».

What can go wrong: wrong advice about the app, the subscription or data.

What has been done: the disclosure before the chat and in the bot's first message; a human answer by email at support@domiqteam.com.

Level: low. The feature is auxiliary.

7. Translation.

What it does: shows tasks, comments, instructions, requests, messages and expenses in the reader's language.

What is sent: the text as written, without placeholders.

Who decides: the reader — the translation is labelled «Translated by AI», the original opens with one tap, and the text itself is not changed.

What can go wrong: an inaccurate translation of a time, a dose, an allergy or an instruction step.

What has been done: a translation whose numbers do not match the original is not shown; the original is always at hand; translation can be switched off in settings, and anyone can object to their own texts being translated by writing to support@domiqteam.com.

Level: low; medium for health and childcare information, which is why the original is always one tap away. The feature is auxiliary.

8. Anonymised aggregated data.

What it does: at night, marks up a workspace's texts with codes — which tasks and how often — so that anonymised aggregated data can be built from them.

What is sent: only texts by people who consented to AI, and only in workspaces whose owner agreed to share anonymised data; names and places are replaced with placeholders; records about children are not used.

Who decides: nobody — the feature decides nothing about people.

What can go wrong: a specific workspace being recognised from the aggregated data.

What has been done: only codes from a fixed list are accepted from the model and stored.

Level: low. The feature is auxiliary.

9. Fill in from photo.

What it does: fills in a property document's fields from its photo — an insurance policy, an equipment passport, a warranty: expiry date, number, model.

What is sent: up to three photos as they are, without placeholders; a photo may show names and addresses.

Who decides: a person — the fields appear in the form, and they check and save them.

What can go wrong: a wrong date or number, so that, for example, an insurance renewal is missed; unneeded details from the photo reach the provider.

What has been done: before the first use, a window says that the photo will go to Groq; the model's answer is checked, and anything invalid is dropped; the provider does not keep the photo after processing; without consent to AI there is no button.

Level: low for mistakes, medium for privacy. The feature is auxiliary.

10. Replies to messages to support.

What it does: reads an email or a message from the app to support and prepares a reply.

What is sent: the text and subject of the message; for emails — the sender's name and email domain, without the address; for messages from the app — role, platform, app version and interface language.

Who decides: in working mode, a reply to a simple question about using Domiq is sent on its own, without a person; the code decides what counts as simple, by fixed rules, not the model. Money, data, account access, complaints, errors, legal questions and requests for a person always go to a person, and the AI prepares a draft for them. While the feature is in trial mode, all replies are sent by a person. Every reply prepared by AI says so: «This reply was prepared by the Domiq AI assistant. Need a person? Reply with the word “human”.»

What can go wrong: wrong advice about the app reaches a person without being checked; a reply misses the point of the question.

What has been done: the model has no tools, and the reply address is taken from the message itself; it is forbidden to promise refunds, discounts or deadlines; a reply is sent on its own only on topics from the answer base, in English, Russian or Spanish, no longer than 2,500 characters and with links only to our websites and app stores; one person gets no more than one such reply a day, and there are no more than 40 a day in total; people who refused the transfer to the United States get a reply from a person; Anthropic does not use the data for training and usually deletes it within 30 days.

Level: low. The feature is not auxiliary: in working mode the reply is sent without a person; it makes no decisions about people, their money or their data.

What AI in Domiq does not do.

It decides nothing about people: it does not hire, does not assess people and does not calculate pay; it does not recognise faces or voices to identify people; it does not track where people are. Money calculations, shift totals and reminders are ordinary code without AI.

Result.

None of the features is an AI system of increased hazard. In features 1–9 the AI only suggests and a person decides (Article 194(3) of the Code). Feature 10 in working mode replies on its own, but only to simple questions about using Domiq: it makes no decisions about people, their money or their data, and such messages are handled by a person.

Questions and complaints about AI in Domiq: support@domiqteam.com. Privacy policy: domiqteam.com/privacy.html.