Short version
Kyrgyzstan's Digital Code has applied since 6 February 2026. It has a chapter on AI systems and a new set of personal data rules. For a company deploying a chatbot, an AI agent or document search for Kyrgyz customers or staff, it creates five pre-launch tasks:
- Assess the hazard of the AI system before use and publish both the result and the methodology on your website (Article 193). This applies to every AI system, not only risky ones.
- Decide whether the system is high-risk (Article 194). Purely assistive systems that do not raise the risk of harm are excluded. High-risk systems need risk management, technical documentation, logs and a signed declaration (Articles 194–196).
- Pick a legal basis for transfers abroad (Article 89). As of 29 October 2026 the regulator has not published its list of countries with adequate protection, so sending data to OpenAI, Anthropic or Google needs consent or another basis from Part 4.
- Sign a processor agreement with mandatory terms, and give written permission for each sub-processor such as the model provider (Article 86).
- Appoint a data protection lead and keep a processing log if you have more than ten employees (Article 88).
This summary works from the Code's text as published by the State Agency for Personal Data Protection. It is not legal advice; have a Kyrgyz-qualified lawyer review your plan.
Scope and dates
The Digital Code (No. 178 of 31 July 2025) was brought into force by Law No. 179 six months after that law's publication on 5 August 2025, which gives 6 February 2026. The parts that matter for AI deployments are the personal data chapter (Articles 80–90) and Chapter 23 on AI systems (Articles 191–196).
The Code defines an AI system as a digital system that, from data and human-set goals, can with some autonomy produce predictions, recommendations or decisions that affect its environment (Article 1, item 63). Any LLM-based assistant that composes its own replies fits. A button-menu bot without a model probably does not.
Obligations sit with the system's owner, the business that deploys it. If a vendor builds the agent and you run it under your name, you are the owner.
Article 193: hazard assessment for every AI system
The owner assesses the hazard the system could pose to protected interests at three points: during design or when planning changes; after development and before use; and after any unplanned change in the system or its environment that could affect the result. The methodology is the owner's own, following requirements set by the Cabinet of Ministers. The result and the methodology go on the owner's website in plain language and as open data.
The third trigger is the one teams forget. Switching to a new model version or adding a data source is a change that calls for a fresh assessment. It pairs naturally with a regression eval suite: the same test run that clears a model upgrade documents the reassessment.
On the Cabinet requirements: economist.kg reported that Resolution No. 770 of 2 December 2025 adopted them, effective 15 days after the Code. We could not read the official text in this review because the national legal database, cbd.minjust.gov.kg, was not reachable. Check your methodology against it before publishing.
Article 194: what counts as high-risk
A high-risk AI system is one whose use, compared with alternatives, raises the risk of harm to protected interests (life and health, rights and freedoms, the environment, defence, national security, public order) to a level that needs risk management. Part 3 is the practical exemption: a system whose role in decisions is purely assistive and that does not raise the risk of harm cannot be classed as high-risk.
So a support bot answering delivery questions, an agent drafting replies that a person sends, or internal policy search will usually be assistive. A model that rejects job applicants without review, or gives medical advice without a clinician, needs a serious assessment.
If a system is high-risk, the owner must meet the Cabinet's requirements, run a risk management system across the lifecycle, maintain technical documentation and logs, and confirm conformity before use (Article 194, Part 5) through a declaration signed with a qualified digital signature and posted on its website (Article 195). Deployers of high-risk systems that serve consumers must also publish how the system works (Article 196).
The design lesson for a first deployment: keep humans on decisions about money, hiring and health, and let the AI answer and prepare.
Article 89: sending data to model providers abroad
Prompts sent to OpenAI, Anthropic or Google servers are cross-border transfers. Transfers to countries on the regulator's adequacy list cannot be restricted, and the owner must check the list before transferring (Parts 1–3). As of 29 October 2026, no such list appears among the regulator's published acts. Until it does, every transfer relies on Part 4, where three bases fit commercial use:
- the data subject's consent to cross-border transfer (item 1);
- performance of a contract to which the data subject is a party or beneficiary (item 4), such as handling a customer's own order;
- a controller–processor contract with adequacy terms set by the regulator (item 6). We found no published terms.
In practice: put transfer consent in the bot's first message and the privacy policy, name the model provider, and strip what the model does not need (passport numbers, full addresses, card data) before the request. Special categories such as health data and biometrics (Article 80) need separate legal review.
Articles 86 and 88: vendors and internal roles
A vendor that builds and supports the agent and sees customer conversations is a processor. Article 86 requires a written agreement covering duration, purpose, data types and categories of people, processing only on written instructions, access limits, correction and deletion procedures, and a log of operations. A processor may not bring in another processor without the owner's prior written permission, which means the model provider and cloud host must be named and approved in writing.
Article 88 requires security measures, a processing log kept to the regulator's recommendations, staff training, and, above ten employees, a designated person responsible for personal data. A separate data protection impact assessment (Article 82) applies to automated decisions with legal effects and to special-category processing for more than a thousand people.
Pre-launch checklist
| Step | Article | Action |
|---|---|---|
| Describe the system | 1, 191 | What the agent does, what data it sees, where a human decides |
| Assess hazard | 193 | Own methodology to Cabinet requirements, result before launch |
| Publish | 193 | Plain-language web page plus an open-data file |
| Rule on high-risk | 194 | Record why the system is assistive, or prepare the declaration |
| Transfer basis | 89 | Consent in the bot and policy, minimal data to the model |
| Processor agreement | 86 | Mandatory terms, written approval of model provider and cloud |
| Lead and log | 88 | Named owner, processing log |
| Reassess on change | 193 | New model or data source means a new assessment |
How we handle it
We deliver from Kazakhstan, and OpenAI, Anthropic and Google all list Kyrgyzstan as a supported country. For a Kyrgyz client, an AI agent pilot includes the consent text, a data map of what reaches the model and what is masked, an Article 86 processor agreement, and a draft Article 193 assessment. The draft draws on the same evaluation set we use for quality: real questions, failure types, and the points where the agent hands over to a person (see What are AI evals and why projects need them). The owner approves and publishes the final version after legal review.
For Kyrgyz-language support we test models on your conversations first; the published evidence is summarised in How well do GPT, Claude and Gemini handle Uzbek, Kyrgyz, Azerbaijani, Armenian and Georgian. The full engagement is on the AI implementation page.
FAQ
On 6 February 2026, six months after Law No. 179 was published on 5 August 2025.
Yes, if a model composes its replies. Article 193 covers every AI system used in Kyrgyzstan, and the result and methodology must be published on the owner's website.
Usually not. Article 194, Part 3 excludes systems whose role is purely assistive and that do not raise the risk of harm. Record that reasoning in your assessment.
Yes, on a Part 4 basis of Article 89, typically consent to transfer abroad or performance of the customer's contract, because no adequacy list has been published yet.
Sources
- Kyrgyz Republic. Digital Code, No. 178 of 31 July 2025. Text published by the State Agency for Personal Data Protection. Articles 1, 80, 82, 86, 88, 89, 191–196.
- Kyrgyz Republic. Law No. 179 on bringing the Digital Code into force. Published 5 August 2025, in force 6 February 2026.
- Economist.kg. Kyrgyzstan sets requirements for high-risk AI systems. 5 December 2025. Cabinet Resolution No. 770 on hazard assessment requirements.