Where we use artificial intelligence
A.M.P.E.R.
Last updated: 30.10.2026.
This page describes the AI systems of the A.M.P.E.R. platform: what they do, which models they run on, what data they receive and how to stop them. It also contains our risk assessment (Article 193 of the Digital Code of the Kyrgyz Republic).
Each system is turned on and set up by the shop (the Client). If the shop has not turned a system on, it does not run.
1. AI sales assistant in chats
What it does. Answers the shop's customers in WhatsApp, Instagram, Telegram and the website chat: about products, stock, prices, installments, addresses and opening hours. Records requests and reservations and hands the conversation to staff.
What it does not do. It does not refuse returns or warranty, change prices, give discounts beyond the shop's rules or cancel orders. Staff of the shop decide these.
Models. Google Gemini via Gemini API: Gemini 3.1 Flash-Lite for answers, Gemini 3.8 Flash and Gemini 3.1 Pro for photos and voice messages.
What data it receives. The history of this conversation, a short card of the customer in this shop (name, language, products of interest), data of the relevant products from the shop's catalog and the shop's rules. AI does not receive data of other shops.
What the customer sees. The first answer in every conversation starts with a line saying that the shop's AI assistant is answering. AI replies to Instagram comments are marked. When the conversation returns to AI after a staff member, the customer sees "AI assistant is answering you again".
How to reach a person. Write "operator", "manager" or "human". AI hands the conversation to the shop's staff and does not answer until a staff member returns the conversation to AI. When a staff member writes, AI stays silent.
2. Voice operator
What it does. Answers the shop's incoming calls: about products, stock, prices and addresses. Records a callback request, a purchase request or a reservation.
What it does not do. The same as the sales assistant: it does not refuse returns or warranty, change prices or cancel orders.
Model. Google Gemini 3.8 Live via Vertex AI, region us-central1 (USA).
What data it receives. The audio of the current call, the caller's number, product data and the shop's rules. We do not store call recordings. We keep the text of the assistant's answers; the caller's words are kept as text only in test mode.
What the caller hears. At the start of the call: "You are speaking with an automated assistant."
How to reach a person. The line works in test mode. Calls are not transferred to staff during the call: the assistant records a callback request and tells you when a staff member will call back.
3. Recognizing photos and voice messages
What it does. When a customer sends a photo of a product, a price tag or a screen, AI identifies the product and checks it against the shop's catalog. AI turns voice messages into text to answer them and to show them to staff.
Models. Gemini 3.8 Flash, Gemini 3.1 Pro.
What data it receives. The photo or voice message itself and the list of matching products of the shop. If AI is not sure it recognized the product, it asks the customer instead of naming a price at random.
4. Quality review
Platform staff check conversations where AI may have made a mistake: a wrong price, a promise without stock, an unneeded handoff to staff. AI (Gemini) helps by marking such places; a person decides. We fix mistakes in the AI rules.
5. Risk assessment
Assessment date: 30.10.2026. Version: 1.0. Next review: no later than 30.10.2027, and also whenever the model, the AI rules or the way a system is used changes noticeably, and before the voice line leaves test mode.
Legal basis: Digital Code of the Kyrgyz Republic, Articles 191–197; Cabinet of Ministers Resolution No. 770 of 02.12.2025.
Owner of the AI systems: Individual Entrepreneur Bayakeev Rustam Islambekovich.
Machine-readable version: /legal/ai.json.
Approved: Individual Entrepreneur Bayakeev Rustam Islambekovich, 30.10.2026.
5.1 Summary
| System | Result |
|---|---|
| 1. AI sales assistant in chats | Not high-risk, auxiliary |
| 2. Voice operator (test mode) | Not high-risk, auxiliary |
| 3. Recognizing photos and voice messages | Not high-risk, auxiliary |
| 4. Quality review | Not high-risk, auxiliary |
In our assessment, the platform's AI systems are auxiliary (Article 194(3) of the Code). None of them makes decisions with legal effect for a person on its own: returns, warranty, prices, discounts and cancellations stay with the shop's staff. AI results can be reversed. A person can always bypass the AI and reach the shop.
Each risk is rated by likelihood and severity, each from 1 (low) to 3 (high). The score is their product: 1–2 acceptable, 3–4 a measure is required, 6–9 a sign of high risk.
5.2 AI sales assistant in chats
| Criterion (Resolution No. 770) | Assessment |
|---|---|
| Purpose | Advising the shop's customers and recording requests on its behalf |
| Foreseeable misuse | Attempts to get a discount or internal data, or to make the AI promise what the shop does not do. Wrong or unlawful terms entered by the shop into the AI rules |
| Autonomy, legally significant decisions | Answers on its own. Makes no legally significant decisions: the shop confirms requests and reservations. Results are reversible; a person can opt out of the AI |
| Personal data | Name, username, phone, conversation, products of interest. No special categories |
| Likelihood of harm | Medium for a wrong price or stock, low otherwise |
| Residual risks | Occasional inaccuracies about specs; an unusually worded request for a person being missed; delays when Google fails |
| Processes | Message → AI answer → handoff to staff if needed → request or reservation in the shop's CRM |
| Period and frequency | 24/7, every incoming message, while the shop keeps the AI on |
| Affected groups | The shop's customers, including older people, Kyrgyz speakers and possibly minors; the shop's staff |
| Risk | Score | Measure |
|---|---|---|
| Wrong price or stock | 2×2 = 4 | Prices and stock only from the shop's catalog. An answer with a price not in the catalog is intercepted and rewritten before sending. The AI does not calculate instalments or guess product models |
| The customer does not realise it is AI | 2×2 = 4 | A line about the AI at the start of every conversation, a mark on comment replies, a line when the conversation returns to the AI |
| A promise the shop will not keep | 1×2 = 2 | Returns, warranty, discounts and cancellations go to staff |
| No way to reach a person | 1×2 = 2 | Words "operator", "manager", "human" and intent detection; the AI goes silent; it does not interfere when staff write |
| Leak of internal text or other people's data | 1×2 = 2 | Only this conversation and this shop's data; internal notes are removed from answers |
| Unequal treatment by language, name or writing style | 1×2 = 2 | Russian and Kyrgyz answers are checked by the same tests; prices and stock do not depend on who asks |
| An error goes unnoticed | 2×1 = 2 | AI call log, incident log, quality review, a test set run before changes |
| The AI keeps writing | 1×1 = 1 | No more than 2 messages in a row without a customer reply, then silence and an incident |
Result: not high-risk.
5.3 Voice operator (test mode)
| Criterion | Assessment |
|---|---|
| Purpose | Answering calls, advising, recording a callback |
| Foreseeable misuse | Getting a discount or promise by phone; taking the AI for staff; expecting a transfer to a person |
| Autonomy, legally significant decisions | Speaks on its own in real time. Makes no legally significant decisions. The only promise it may give is a callback during opening hours |
| Personal data | Voice, phone number, conversation content |
| Likelihood of harm | Medium for a wrongly spoken price, low otherwise |
| Residual risks | A wrong price is caught only after it is spoken; bad lines; no transfer to a person during the call |
| Processes | Call → greeting with AI disclosure → catalog answers → callback task for staff |
| Period and frequency | Test numbers only while in test mode |
| Affected groups | Calling customers, including older people and Kyrgyz speakers; the shop's staff |
| Risk | Score | Measure |
|---|---|---|
| Wrong price spoken | 2×2 = 4 | Prices only from this call's catalog search. After each turn, spoken amounts and instalment terms are checked against the search results; a mismatch is logged as an incident. Correcting it within the same call is planned |
| No way to reach a person during the call | 2×2 = 4 | A callback request with the shop's opening time. The AI does not promise a transfer. Transfer to staff is planned |
| The caller does not realise it is AI | 1×2 = 2 | The first words of the call: "You are speaking with an automated assistant" |
| A promise the shop will not keep | 1×2 = 2 | No returns, discounts or cancellations |
| An error goes unnoticed | 2×1 = 2 | Statistics for every call, incident log, control calls on the test line |
| Excess data | 1×1 = 1 | No call recordings; the caller's words are kept as text only in test mode |
Result: not high-risk. To be reassessed before the line leaves test mode.
5.4 Recognizing photos and voice messages
| Criterion | Assessment |
|---|---|
| Purpose | Understanding which product a photo shows; turning a voice message into text |
| Foreseeable misuse | A photo of a look-alike product, a document or a person; a voice message with third parties' speech |
| Autonomy, legally significant decisions | Only passes the result to the sales assistant and staff. Makes no decisions |
| Personal data | Voice in voice messages. A person or document may appear in a photo by accident; the system does not analyse them |
| Likelihood of harm | Low |
| Residual risks | A misread name on a blurred photo; a misheard word in a voice message |
| Processes | Photo or voice message → recognition → sales assistant answers or asks |
| Period and frequency | For every photo without a model name in the text and every voice message |
| Affected groups | Customers who send photos and voice messages |
| Risk | Score | Measure |
|---|---|---|
| Wrong product, wrong price | 1×2 = 2 | Answers only from the catalog's model list. A model is named only when its name is readable in the photo; otherwise the customer is asked |
| A misheard voice message | 1×2 = 2 | Staff see the text of the voice message. If something is unclear, the AI asks again |
| Processing a face or document | 1×1 = 1 | People are not recognised; such a photo gets the answer "not a product" |
Result: not high-risk.
5.5 Quality review
| Criterion | Assessment |
|---|---|
| Purpose | Finding AI mistakes in conversations to fix the AI rules |
| Foreseeable misuse | Using the marks to evaluate customers or staff |
| Autonomy, legally significant decisions | Only marks places; a person decides. No decisions about customers |
| Personal data | Conversation content |
| Likelihood of harm | Low |
| Residual risks | A missed mistake |
| Processes | Conversation with a suspected error → AI mark → review by a person → fix to the rules |
| Period and frequency | For conversations with signs of an error |
| Affected groups | Customers whose conversations are reviewed; the shop's staff |
| Risk | Score | Measure |
|---|---|---|
| Use for evaluating people | 1×2 = 2 | Marks concern AI answers only, not customers or staff; a person decides |
| A mistake missed | 2×1 = 2 | Rule-based checks in addition to the AI marks |
Result: not high-risk.
5.6 Method
- When. Before launch; when the model, the rules or the way of use change; after any unplanned change that may affect the result; at least once a year.
- What is protected (Article 192): life and health, human rights and freedoms (including personal data and reliable product information), the environment, defence, national security, public order.
- Criteria. For each system: purpose; foreseeable misuse; autonomy and legally significant decisions without a person, reversibility, the ability to opt out; personal data and protected secrets; likelihood of harm; residual risks; processes; period and frequency; affected groups.
- Steps. Describe the purposes and misuse → list the risks → rate likelihood and severity from 1 to 3 → assign a measure and its owner → decide on high risk.
- Severity. 1 — inconvenience, fixed at once. 2 — wrong information or lost time, fixed by staff. 3 — legal effect or damage that is hard to undo.
- Auxiliary nature (all at once): no decisions with legal effect without a person; the result is reversible; the person can opt out and reach staff; decisions on sales, returns, prices and cancellations stay with staff.
- Sources. Platform code and settings, AI call and incident logs, test results. We take into account the OECD AI Principles, the NIST AI RMF 1.0 and ISO/IEC 23894.
- Who. The platform development team prepares the assessment. The owner of the systems approves it and is responsible for the measures. Approved: Individual Entrepreneur Bayakeev Rustam Islambekovich, 30.10.2026.
- Method review. At least once a year. The outcome is a record "no grounds for revision" or a new version on this page.
6. How to complain
If AI answered wrongly or you believe a decision about you was made automatically, write to the shop you contacted or to us: support@kpi.kg. You may ask for the decision to be reviewed by a person (Article 57). We reply within 7 working days.
You can also contact the State Agency for Personal Data Protection under the Cabinet of Ministers of the Kyrgyz Republic: dpa.gov.kg, info@dpa.gov.kg.
7. Contact
Individual Entrepreneur Bayakeev Rustam Islambekovich, 7 mkr., Oktyabrsky district, Bishkek 720000, Kyrgyzstan. E-mail: support@kpi.kg. Phone: +996 556 308 106.