Which models
Primarily Anthropic's Claude models — our own product Genix ERP runs on them. Where a task calls for a different model, we use it; the choice follows your task and budget.
AI agent development — for real work in your business
Not demo chatbots: the agent answers a lead, reads a document, posts a record to Odoo or 1C — under your rules and your approval. Works in Uzbek and Russian.
What we build
Custom AI agents
For sales, customer support, and back office: replies to leads, documents, invoices, reports — each agent is built around your process.
AI assistant over company data (RAG)
Answers from your documents, price lists, regulations, and ERP data — from your own base, not the open internet.
LLM integration into existing systems
Odoo, 1C, custom ERP or CRM: the agent reads and writes real records — with guardrails, and human approval where required.
Telegram, web, and voice agents
The agent lives where your customers are: a Telegram bot, a chat on your site, or a phone line.
Agentic workflow automation
Email → document → ERP record chains run without manual retyping: the agent reads the document and posts it to the right system itself.
AI copilots for staff
An accountant's or sales manager's copilot inside your own system: it searches, drafts documents, suggests options; a person makes the call.
How it works in your industry
One concrete scenario per industry — to show where the agent plugs in and what it does.
- Retail & trade
A customer asks on Telegram: "Any fridges under 5 million soums?" — the agent pulls stock and prices from Odoo, replies, places the order, and hands over to a manager only at payment.
- Automotive
A customer wants a service appointment — the agent checks free slots, books one, and sends a reminder. Our Telegram auction bot for Carland already runs in this channel.
- Manufacturing
A supplier invoice arrives as a PDF — the agent reads it and posts the receipt to 1C or Odoo; if the numbers don't match, it raises the question with the responsible person.
- Services
Leads come from the site, Telegram, and phone — the agent logs each one in the CRM, sends the first reply, and books the meeting; the manager sees a deal that's already prepared.
- Construction
A buyer asks about a building, floor, and price — the agent answers from the price sheet and unit chart, then hands the interested buyer to a manager with the full conversation history.
How it plugs into your existing systems
An AI agent is not a separate program — it is built on top of the systems you already run.
Odoo and 1C: the agent reads stock, prices, and counterparties; it creates orders and documents under your rules.
Custom ERP or CRM: if there's an API, we connect to it; if not, we build the integration layer ourselves — that is our core work.
Write access always comes with guardrails: what the agent does on its own and what needs human approval — we define together.
Every action is logged: what the agent did and why — auditable at any moment.
How we deliver
Discovery & analysis
We find where an agent saves the most time: lead flow, documents, information lookup. The output is one concrete task and a measurable criterion.
A 2–4 week pilot
A working agent on a narrow but real task — on your data, in your process. The pilot's scope, timeline, and price are fixed in a written plan.
Integration & rollout
Once the pilot proves itself, we wire the agent fully into your workflow: systems, access rights, staff training.
Support & monitoring
We watch the agent's answers, tune the rules, and add new tasks. Quality metrics come in a regular report.
Technology and security — straight answers
Data security
Your data is not used to train models. The agent accesses only the data we agree on, and every action is logged.
On-premise option
If required, we deploy the solution on your own servers — data never leaves the company. What runs where is spelled out in the written plan.
Uzbek and Russian
Agents work in the languages your customers write in: Uzbek, Russian, English when needed. We tune tone and terminology to your industry and verify quality on real conversations during the pilot.
Why you can trust us with this
Genix ERP is our AI-native ERP built on Claude: multi-agent architecture, RAG, and write-path guardrails — running in production. Client-facing AI has already shipped: the AI assistant at Markab Store and the Telegram auction bot at Carland. Six years and 50+ projects inside ERPs, POS, and accounting systems mean we know exactly where an agent has to plug in.
Frequently asked questions
How much does an AI agent cost?
We start with a 2–4 week pilot — far cheaper than a full rollout, and it shows up front whether the spend pays off. The exact figure depends on task scope, integrations, and channels — it is fixed in a written plan after analysis, before any work starts. Model usage (API fees) is added on top: it scales with usage and appears as its own line in the estimate.
How long does it take?
The first working agent — the pilot — launches in 2–4 weeks. Full integration can take several more weeks depending on the number of systems, access rights, and training scope. Timelines are fixed in the written plan.
How is our data kept secure?
Three rules. First: your data is not used to train models. Second: the agent accesses only agreed data — its rights are no wider than an employee's. Third: every action is logged. If your requirements are stricter, we deploy on your own servers.
Which AI models do you use?
Our primary choice is Anthropic's Claude — our own product Genix ERP runs on it, so we know its capabilities and limits from practice. If a task calls for a different model — for price or hosting reasons — we propose it with the reasoning in the open. The model choice is recorded in the written plan.
Does the agent understand Uzbek?
Yes. Modern models work in Uzbek and Russian — the agent replies in whatever language the customer writes. We tune tone and industry terminology together and verify answer quality on your real conversations during the pilot.
How does it connect to our existing system?
Odoo, 1C, and any system with an API we connect to directly; if there is no API, we build the integration layer ourselves — that has been our work for six years. What the agent may read and write is restricted at the access-rights level.
What happens if the agent makes a mistake?
Guardrails exist for exactly this: critical actions — payments, large discounts, deleting documents — never run without human approval; on a low-confidence answer the agent hands the conversation to a person; every action is logged, so a mistake is easy to find and the rule easy to fix. Where the agent answers alone and where a person steps in — we define together during the pilot.
What are the pilot terms?
A pilot is 2–4 weeks, one concrete task, real data. Before it starts, the written plan fixes scope, timeline, price, and the success criterion. At the end you keep a working agent and real numbers: continuing or stopping is your call.