AI
What Is AI Implementation for Businesses in Uzbekistan?
AI implementation is not adding a chatbot. It is embedding intelligent agents into real business workflows - lead qualif...
Databricks has pointed in the right direction: an AI agent should not work “in general,” but on your data, with access rights and quality checks. For companies in Uzbekistan, this is a chance to close sales, loans, and operational issues faster — if they put their data in order first.
It’s Friday, purchasing is closed, and on Monday in Yunusabad the fast-moving rice is missing again. The director of a retail chain in Tashkent does not want a lecture about artificial intelligence. He wants an answer: where did the chain snap — in the forecast, the shipment, receiving, or on the shelf?
Usually, a question like that starts a manual hunt through 1C, Excel, Telegram, and calls to store managers. The launch of Databricks Agent Bricks is interesting not because of the brand, but because of the shift in approach: an AI agent stops being a toy chat window and becomes a working service that answers based on the company’s data.
For businesses in Uzbekistan, the conclusion is direct. The agent should work next to ERP, CRM, contracts, and reports, see only the data it is allowed to see, and pass quality checks like any internal product. Databricks has not brought a magic button. It has simply formulated a mature pattern well: agent, data, permissions, tests, and an action log in one loop. Start with a narrow process — inventory, a loan application, a clinic schedule, maintenance requests — and you will see value faster. Connect AI “to all documents” and you will get an expensive fantasist with a confident voice.
Why this is not just another chatbot
A chatbot answers a question. An AI agent closes a small piece of work: finds records, checks a rule, asks for clarification, assembles a conclusion, and leaves a trace in the log. The difference is not academic. You see it the moment a buyer asks: “Show me the products where sales are growing, but the supplier has missed deadlines for the second month in a row.”
In Uzbekistan, this is especially sensitive. Data is often scattered between 1C, Odoo, Bitrix24, Excel, and chats. The same city may be entered in Russian, Uzbek, Cyrillic, and Latin. The agent has to live in this mess, but not multiply it. That is why I read the Databricks launch as a signal: the market is tired of pretty demos and has moved on to engineering trust.
Market signal
7.2%users of generative AI
17.8%global benchmark
62ndrank in AI readiness
Uzbekistan is not overheated yet — and that is a plus
According to Microsoft data cited by The Times of Central Asia, the share of generative AI users among Uzbekistan’s working-age population grew from 5.7% in the first half of 2025 to 7.2% in the first quarter of 2026. The global level of 17.8% is still far away. For business, this is not a reason to sit still, but a rare moment to enter without panic and inflated budgets.
Companies that bring their data into order now and launch 1–2 applied agents will be counting the savings a year from now. The rest will buy licenses, then discover that the agent cannot recognize a customer because the same customer was entered into the system in three different ways.
Harsh, but fair
Do not start with the phrase “let’s build an agent for the whole business.” Start with one expensive question that employees answer every day while losing hours.
The first payback often appears where data explains an operational break.
Where it pays off first
A good first scenario is easy to recognize: it repeats often, mistakes cost money, and the data already exists somewhere — even if it is not in perfect shape.
🛒
Retail and distribution
The agent breaks down stock gaps, sees supplier delays, and prepares a short summary for the buyer without manual table-hopping.
💳
Banks and microfinance
It gathers borrower data, checks internal rules, and highlights risk before the application goes to manual review.
🌱
Agriculture and water
For farms, the agent can read logs, sensors, and agronomist recommendations instead of retelling a generic knowledge base.
🩺
Clinics
It helps the administrator find an appointment slot, check insurance rules, and avoid sending the patient back and forth between reception and doctors’ offices.
⚡
Energy and service operations
The agent analyzes requests, emergency logs, and regulations so an engineer can find similar cases faster.
The real cost is not the model, but trust in the data
The most common mistake in projects like this: the company argues about which model to choose, while the product catalog contains duplicates, access rights have not been cleaned up for years, and contracts are stored as scans without proper recognition. An agent on that data does not “sometimes miss.” It confidently assembles the wrong picture.
Before a pilot, we at Celion ask for three things:
- a source map — where the agent gets its facts from;
- a permissions matrix — who is allowed to see what;
- test questions with reference answers — how quality will be measured.
Yes, it is boring. But later the director does not have to argue with AI in the style of: “Why did you show branch margin to an employee who should not see it?”
An AI agent is only as good as the company is honest about its data, permissions, and exceptions.
Celion
What the Databricks approach teaches us
The lesson is not that everyone urgently needs Databricks specifically. For large banks, holdings, and retail chains with serious analytics, such a platform may fit well. For mid-sized businesses in Tashkent, the same discipline on a lighter stack is often enough: a data warehouse, access through API, document search, roles, and quality tests.
The point is tougher than it sounds in presentations. The agent should not live as a separate “smart chatterbox” off to the side of the IT system. It should work inside the data loop, where it is clear where the answer came from, who requested it, and why that answer can be trusted.
The main cost of an agent is not the model, but trust in the data sources.
A 30–45 day pilot plan
This kind of pilot does not prove “AI in general.” It checks whether the agent will remove a specific business pain.
- Choose an expensive question For example: why a product disappears from the shelf, why a loan application got stuck, or where customer requests get lost.
- Connect two sources ERP and a document database are enough to start. Ten integrations in the first month almost always slow the project down.
- Collect 50 checks Write down real employee questions and reference answers so you can evaluate the agent with numbers, not impressions.
- Configure permissions Sales should not see salaries, a branch should not see another branch’s margin, and a contractor should not see internal regulations.
- Launch with a human nearby In the first weeks, the agent suggests an answer, while an employee confirms it, corrects it, and trains the system on real cases.
What I would not do at the start
I would not give the agent direct permission to change data in ERP in the first month. Read-only access, drafts, recommendations, and a clear log only. Writing back to the system comes later, when answer quality stays consistently at the agreed level and the business understands the cost of an error.
I also would not start with a “corporate assistant for everyone.” It looks pleasant on a slide and is hard to measure in real work. A normal first pilot in Uzbekistan often runs not into the cost of the model, but into 60–150 mln soums of integration and preparation work. That money should be spent on a process where time or money saved can be calculated without gymnastics in Excel.
A 30–45 day pilot works best around one measurable process.
What to remember
If you remove the noise around the Databricks launch, practical conclusions remain.
- The agent needs your data. Without ERP, CRM, documents, and access rules, it remains an ordinary chatbot with a nice interface.
- Start narrow. One process with money at stake is better than an abstract assistant for all employees.
- Test answers in advance. A set of test questions and reference answers is needed before launch, not after the first scandal.
- A platform will not save chaos. Databricks, cloud, or a local stack will help only where the company is ready to describe sources and responsibility.
- Access rights are part of the product. If the agent sees too much, it is no longer innovation — it is a business risk.
FAQ
Does a company in Uzbekistan have to implement Databricks?
No. Databricks makes sense where there are large data warehouses, complex analytics, and strict access requirements. For a mid-sized business, it is often more reasonable to build the same principle on a lighter architecture: a database, API, document search, roles, and quality checks. The sign on the door is not what works. Data discipline is.
Can an agent be connected to 1C, Odoo, SAP, or Bitrix24?
Yes, if there is access through API, exports, or an intermediate storage layer. But the connection itself does not make a product. First you need to decide which tables are needed, who has the right to see them, how to handle duplicates, and what to do if the data in the system is outdated.
Will the agent work properly with Russian, Uzbek, Cyrillic, and Latin?
It can, but you need to test it on your data. In Uzbekistan, the same client, address, or product is often written in several variants. That is why the pilot must include real documents, chats, and directories — not sterile examples from a presentation.
How safe is it to give corporate data to an AI agent?
Security depends on the architecture. Data can be masked, stored inside your own environment, restricted by permissions, logged through request journals, and kept from sending sensitive fields to external services. The most dangerous option is when employees upload contracts and reports into public tools on their own, without rules.
When does an agent like this start saving money?
The first metrics are usually visible in 4–8 weeks if the scenario is chosen correctly. Good indicators are fewer manual reconciliations, faster request processing, a lower number of errors, and fewer calls between departments. If the metrics are not defined before the pilot, the project quickly turns into an expensive demo.
Discuss a pilot
At Celion, we design AI agents around real business data: ERP, CRM, documents, permissions, and quality metrics. Write to us — we will break down one process and show where an agent can create an effect without showroom magic.
