AI implementation for your business in Uzbekistan

Implementing AI in a business means working on one process, not buying a neural network. A model takes over one step — processing incoming invoices, say, or the first reply to a customer in Telegram — while a person keeps every decision that changes money or the customer’s rights. The step has an input, an output and a metric fixed in advance. The audit and pilot run online, so whether you are in Tashkent, Samarkand or elsewhere in Uzbekistan affects neither the timeline nor the price.

What AI implementation is — and why it is not a ChatGPT subscription for staff

AI implementation means building a model, usually a language model, into a company process: it takes data from your systems, performs its step and returns the result to where the work continues. Technically, it is AI integration for business processes: your CRM, 1C, Google Sheets, Telegram and telephony stay, and the model sits between them.

A ChatGPT subscription for staff works differently: each person decides what to ask and moves the answer by hand. According to the 2025 MIT NANDA report, 40% of companies bought an official LLM subscription, while workers at over 90% of companies regularly use personal AI tools for work. In the authors’ view, such tools mainly raise individual productivity rather than company P&L (our breakdown).

Turnkey implementation covers what a subscription will not:

• one process and the step that goes to the model;
• a connection to your data, so results land in your systems without copying;
• rules for what the model does itself and what it hands to a person;
• a metric fixed before the start;
• handover of the code, access and documentation.

What AI can automate: tasks where a model pays off

AI pays off on repetitive steps with free text or accumulated data as input, where an error can be caught before it reaches the customer. Eight typical tasks:

The model prepares; a person decides wherever money or the customer’s rights are at stake. For replies in messengers see chatbots, for leads and the pipeline sales automation; a model that picks its own tools and their order is an AI agent. All these AI solutions for business in Uzbekistan are one step inside wider business automation.

Where you do not need AI

You do not need AI where a rule or a ready-made integration does the job, or where an error costs more than the savings.

A rule will do. If the decision fits “if this, then that” — a website lead goes into the CRM, the manager is notified, an overdue invoice triggers a reminder — it is an n8n workflow, and a model would only add cost and random errors. A model earns its place where the input is free text: an email, a voice message, a scan.

Something ready-made will do. If your CRM’s own integration covers it, or an off-the-shelf model solves it in two days, there is no point paying us for development, and we will say so.

The error is too costly. Uzbek law does not allow legally significant decisions to rely solely on AI output (see the section on the law); on our reading, that includes refusing a refund, instalments or credit. The model can gather data and suggest an option; the final word stays with a person.

The process is rare, changing or unrecorded. A step run less than weekly will not repay its support, a process about to be rewritten would be automated twice, and data kept in someone’s head must first be written down. Sometimes a step is better removed than automated — see “Don’t Automate, Obliterate”.

We do not take on tasks where success cannot be measured, and we do not sell “AI implementation” without a named process and a metric.

Examples of AI in business: our own systems

Our examples are two working systems and three Lab products: we have no client case studies with numbers yet, since the company was registered in August 2026.

• Valli. The bot handles conversations with customers in Telegram and passes everything non-standard to a person. It is currently in pilot.
• CorpVisor. Oversees the process at a manufacturing company in Samarkand: tasks by role, reminders, replies to customers, prepayments and debts after delivery, a morning summary for the head of the company.
• SUV AI. Agritech: every morning it tells the farmer when to irrigate and how much water to apply. Stage 5 of 5; the first paid contract covers 100 hectares at a large farm.
• EstIQ. Prices an apartment in Uzbekistan in 40 seconds, with a median error of about 10% on resale apartments, based on 125,000 properties from open market data. Stage 3 of 5.
• KARTOCHKA. Brings Uzum sellers’ product listings in line with the platform’s rules. Stage 2 of 5.

Each has one checkable result, and a client implementation starts from the same kind of brief. The rest of the Lab is on the home page.

Which AI is best for business

There is no best model for business, only the one that handles your task best on your data at an acceptable price. Compare candidates on five criteria.

1. Quality in Uzbek and Russian. Customers write Uzbek in Latin and Cyrillic script, switch to Russian mid-sentence and send voice messages; general rankings will not show how a model copes, so candidates are compared on your real messages.
2. Price per request. A provider’s smallest and largest models can differ in price by up to 100 times (see the cost section). For sorting requests by topic, test the smallest first: if it copes, there is no reason to pay more.
3. Where the data is processed. OpenAI, Anthropic and Google models run on the provider’s servers abroad, so personal data becomes a legal question (see below). An open model on your own server keeps the data inside but needs hardware, and its Uzbek must be tested separately.
4. Speed. A customer in a chat notices seconds; an overnight document run does not.
5. Context length. How much text the model takes in at once: a long contract, a month of chat history, a hundred-page policy.

Examples: ChatGPT and the OpenAI API, Claude from Anthropic, Gemini from Google, open models on your own server. We choose the model for the task; the keys are yours, on a provider account in your name. The model is one node in the workflow, so a better fit can be swapped in after a test on the same examples.

AI implementation stages

Implementing AI in business processes runs in five stages: a 48-hour audit, data and access, a 2-week pilot, handover, and the next process.

1. Audit — 48 hours, free. Describe one process in the questionnaire, no intro call needed. Our founder, Amir Negmatullaev, breaks it down himself and sends back a map: what to take off people, where a model is needed and where a rule will do, in what order and at what cost.
2. Data and access. We sign an NDA before any data is handed over, connect to your systems under your accounts and collect real emails, requests and documents with the correct result for each, to test the model on before it meets customers.
3. Pilot — 2 weeks. One process runs live. The “before” metric is written down in advance — say, time to the first substantive reply or the share of documents processed without correction. Miss it, and the stage is not paid for.
4. Handover. Code, access and documentation are yours, and everything lives in your environment. Support comes by subscription, or your own team runs the system.
5. Next process. We take the next step from the audit map; three or four connected processes make up the Shop floor package, 4 weeks.

What it costs to implement AI in your business

Implementing AI in one process costs from $1,200 for development plus from $300 a month for support; the model provider is paid separately, at their rates. The first audit is free. The price has three parts.

Development. Per project, not per hour: from $1,200 for one process end to end (Pipeline), from $6,500 for three or four connected processes (Shop floor), from $18,000 for a pilot of your own model on your data (Custom); the pricing block below lists what each includes. The range depends on the number of integrations and data sources. No discounts: a smaller budget means a narrower scope. In Uzbekistan you can pay in soum at the Central Bank rate on the invoice date.

The model. You pay the provider directly from your own account, per million tokens (a token is a piece of a word), with input and output priced separately (OpenAI, Anthropic, Google). The spread is wide: as of 27 September 2026, the smallest model in OpenAI’s GPT-6 line, Luna, costs $0.10 per million input tokens and $0.50 per million output tokens; the largest, Astra, $10 and $50. A rough example: 10,000 requests a month at 3,000 input and 300 output tokens each come to about $4.50 on the smaller model and about $450 on the larger one. We estimate your volume at the audit.

Support. From $300 a month for Pipeline, $700 for Shop floor, $1,200 for Custom: monitoring, repairs when data sources break, and changes. You can drop it and run the system yourself; a server for the workflows, if needed, is also in your name and paid to the host directly.

50% reimbursement. On 18 August 2026 the President announced the “Artificial Intelligence — Partner for 10,000 Enterprises” programme, under which enterprises are to be reimbursed half their AI implementation costs. The payment procedure had not been published by late September 2026, so we make no promises about reimbursement; what is known is in our breakdown of the programme.

Data, security and the law

Uzbekistan’s legal framework for AI is Law ZRU-1115 of 21 January 2026: legally significant decisions affecting human rights and freedoms may not rely solely on AI output, and unlawful processing of personal data using AI carries a fine. The ban sits in Article 7¹ of the Law “On Informatization”, the fine in part two of Article 46² of the Code of Administrative Responsibility: 50–100 base calculation units, or 22–44 million soum since 1 September 2026 (ZRU-1115, Decree UP-115).

Read literally, the rule lets AI prepare a decision — gather data, sort requests, suggest an option — but forbids its output becoming the decision with no person in between. The AI ethical rules No. 3787, in force since 17 June 2026, add that “the final decision must be made by a human” (our breakdown of ZRU-1115). The rules, fines and figures, each with its primary source, are collected in our reference page AI in business in Uzbekistan: 2026 numbers and rules.

Where data may be stored. Since 27 March 2026, Law ZRU-1125 requires only biometric data, genetic data and data on users of telecom operators’ services to stay in Uzbekistan. Customer chats may go to an external model under one of three conditions, such as the provider’s country being on the list with adequate data protection: 49 countries and territories, the US only for companies operating within the EU-US Data Privacy Framework. The law’s other personal data requirements still apply (breakdown).

Does the model learn from your data? At OpenAI, data sent to the API has not been used to train models since 1 March 2023, unless you explicitly opt in (OpenAI). Anthropic’s commercial terms state that it may not train models on customer content (Anthropic). At Google it depends on the tier: on the paid Gemini API tier, prompts and responses are not used to improve products; on the free tier they are, and human reviewers may read them (Google). Not training is not the same as not storing: by default, OpenAI keeps abuse monitoring logs for up to 30 days.

What we do.

• Everything lives in your environment: your server, your n8n, your model keys.
• What goes to the model is fixed in the spec; sensitive fields can be left out.
• We sign an NDA before any data is handed over.
• We do not take your database, resell it or train anything unrelated on it — that is a separate clause in the contract.

This is not legal advice: check how these rules apply to your customer data with a lawyer.

Why AI projects fail to deliver

An AI project fails to deliver when the result was not written down before the start, or the model was given a task beyond its capabilities.

The “95% of AI pilots fail” figure comes from a preliminary 2025 MIT NANDA report built on interviews with representatives of 52 organisations and 153 surveys, where success was whatever respondents called success. Its funnel for enterprise AI systems: 60% of organisations evaluated them, 20% reached a pilot, 5% reached production.

Where results were measured, the picture is mixed. In a BCG experiment with 758 consultants, GPT-4 helped on tasks within its capabilities, but on a task beyond that frontier, consultants using AI were 19 percentage points less likely to reach the correct solution than those without it. The authors call the frontier jagged; only measurement shows where it runs through your process.

How we guard against this:

• the metric is one sentence written before the start: what we count, in what units, what number means success;
• the “before” is measured on normal work, not on impressions;
• quality is measured alongside speed;
• we time things instead of asking participants;
• if we miss the metric, the stage is not paid for.

Six pilot rules and the research behind them: our piece on the 95% of AI pilots.

AI implementation companies: how to choose

Choose an AI implementation company by its answers to specific questions, not by its website: everyone offers “turnkey AI”, from CRM partners and chatbot studios to freelancers and developers. Four signs of a contractor that sells results, not hours:

1. It names one process and a metric — and what happens if the metric is missed.
2. It puts everything in your name: code, access, accounts and model keys; you can ask for them in week two, not at the end.
3. It says where the data goes: which fields reach the model, which provider, where they are stored.
4. It names jobs it would turn down, including where AI is not needed.

Eight questions to ask a contractor, with what good answers look like, are in “Business automation companies: how to choose one”. We are a software developer: “AI CAPITAL FACTORY” MChJ, tax ID 313270615, IT Park Uzbekistan resident, a team of 13.