An agent answers the client, books them in, opens a deal in your CRM and hands anything unusual to a person. We build custom AI agents for business around one process at a time: minimal permissions, a metric fixed in writing before the start, code and access that stay yours.
This is not a video: pick a scenario and watch the bot run the dialogue. It would walk your customer the same way.
Unlike an ordinary bot, an AI agent acts inside your systems: it checks free slots, creates bookings, opens deals and assigns tasks to staff. For us, AI agent development starts with one process rather than a choice of model: where the agent will take hours off your people, which actions it can be trusted with, and where it has to stop.
Below: what kinds of AI agents for business exist, how an agent differs from a chatbot and a workflow, how it is built, what makes it risky and what it costs. In the demo on this page, a salon bot does what an agent does in its simplest form: books a client, quotes a price and hands an off-script question to a specialist along with the conversation.
An AI agent is a language model that does not just answer but acts: it looks things up in the CRM, books the client, sets a task and hands the conversation to a person when a question goes beyond its limits.
In September 2025 the developer Simon Willison wrote that, in his view, the word “agent” finally has a definition agreed on widely enough to be useful: “An LLM agent runs tools in a loop to achieve a goal” (simonwillison.net). Word by word:
An example. At 23:40 a client writes on Telegram: “Can I come tomorrow afternoon?” The agent checks free slots, offers 13:30, creates the booking once the client agrees and sets a reminder. If the next message asks about an allergy to a material, the agent does not answer it; it passes the question to a specialist with the conversation attached.
Agents are also called AI assistants or “AI employees”. The second name is misleading: an employee answers for their decisions, an agent does not. The company that deployed it answers for what it says — that is where a tribunal in Canada and a court in Germany landed (both cases).
A chatbot walks the client through a pre-drawn script, a workflow runs a fixed chain of steps, and an AI agent chooses its next step itself — which is why it needs limits the other two can do without.
Anthropic, the company behind the Claude models, draws the line like this: in a workflow, the model and tools follow paths written in code in advance; an agent directs its own process and decides which tools to use (Building effective agents, December 2024).
| What it can do | Button chatbot | Workflow (n8n) | AI agent |
|---|---|---|---|
| Understands free text and voice | No, only buttons and keywords | Yes, if a model step is built into the chain | Yes |
| Who picks the next step | The script | A diagram drawn in advance | The model, within what is allowed |
| Actions in CRM, calendar, spreadsheet | By script: a lead, a booking | By diagram, always the same | Picks from an approved list |
| Off-script question | Dead end, unless a human handover is built in | Goes down a predefined branch | Recognises it and hands it to a person with the history |
| Predictable path | Fully | Fully | Less so: needs minimal permissions and a log |
| When to choose it | Menu, three-step booking, FAQs | Moving data, reports, notifications | Free-form conversation, many variants, actions in your systems |
The honest answer to “which one?” is that a bot or a workflow is often enough. In the same article Anthropic advises looking for the simplest solution, adding complexity only when needed, and accepts that sometimes the right call is no agentic system at all. It also notes that an agent’s autonomy comes at the price of higher costs and errors that compound.
We think the same way. If the process fits into buttons and a short script, a chatbot will do. If the steps are always the same, we build a workflow in n8n. An agent is needed when clients write in free text and voice messages, conversations branch dozens of ways, and something has to happen in your systems along the way.
Often the best option is a hybrid: the workflow holds the fixed steps and the model sits in it as one link that reads the text. Other kinds of automation, and where AI agents fit among them, are covered on the business automation page.
By task, agents fall into six types: sales, support, client requests, process supervision, documents and an internal knowledge base.
Answers a lead at once, including nights and weekends, asks questions from your script, qualifies the client and books a meeting, a viewing or a test drive. A hot lead goes to a manager as a ready card. How the agent fits into the funnel and the CRM, and which metrics to track, is on the sales automation page; for dealers tied to a manufacturer’s CRM, see car dealerships.
Handles the routine: order status, address, terms. Complaints, refunds and anything not covered by approved answers go to a person.
Klarna is a telling case. In its first month, its AI assistant handled 2.3 million conversations, two-thirds of the company’s customer service chats. Fourteen months later the CEO admitted that cost had weighed too heavily and quality had dropped (our breakdown). That is why, in a support pilot, we put a quality measure next to speed — for example, the share of repeat enquiries.
Clients of an accounting firm write on Telegram: “issue an invoice”, “send the reconciliation statement” — in the evening and at the height of reporting season. The agent checks the counterparty, tax ID and amount, reads the attached contract and passes a complete request to the assigned accountant. It does not quote tax amounts: that is the accountant’s job. Details are on the accounting firms page.
Watches not a single conversation but the whole process: who has which task, what is overdue, who needs a reminder, what to report to the manager. Our own system is an example. CorpVisor monitors the process of a manufacturing company in Samarkand: tasks by role, reminders, replies to clients, prepayments and debts after delivery, a morning summary for the manager.
Reads emails, PDFs and photos of contracts, pulls out bank details, amounts and dates, checks them for completeness and puts them where they are expected: a spreadsheet, the CRM or 1C. A field the model is unsure about is better flagged for a person than guessed.
Answers employees from your policies, price lists and instructions: what warranty this model has, how to process a return. A good first agent: it never sees clients, sends nothing outside, and an employee will catch its mistake before a client does.
We have no client case studies with numbers yet: the company has been operating since August 2026. So instead of other people’s logos we show our own systems and write the metric into the contract. Our second system is Valli. The Valli bot handles conversations with clients on Telegram and hands everything non-standard to a person — it is currently in pilot.
An AI agent has five parts: a model, tools, memory, rules and a human in the loop. The last four matter for reliability at least as much as the model does.
The agent does not decide anything that involves money or changes the client’s position: discounts, refunds, instalments, cancelling a deal, and medical or legal questions. There it collects the details and hands the conversation to a person.
It quotes prices and terms only from text you have approved; if the answer is not there, it says it will check. A clinic adds a stop-list of medical topics, agreed in writing before launch (how that works). For an accounting firm, tax amounts go on the stop-list.
This is not only caution. Article 7¹ of Uzbekistan’s Law “On Informatization”, added by Law ZRU-1115, bars relying solely on the conclusions of AI-based systems when legally significant decisions affecting a person’s rights and freedoms are made (lex.uz). What that means for a sales or support bot is covered in our breakdown of the AI law ZRU-1115.
The main risk of an AI agent is not a clumsy sentence but an action taken at someone else’s prompting: leaking data, promising too much, deleting a record. We close it with limits in code: minimal permissions, human confirmation for anything irreversible, an action log.
The developer Simon Willison described a “lethal trifecta”: an agent is dangerous when it has access to private data, exposure to untrusted content and the ability to communicate externally, all at once. He writes that with this combination an attacker can easily trick the agent into fetching private data and sending it to them (Willison’s post). A Telegram agent that can see your client base gets all three on day one: every incoming message is untrusted content, and every reply is a way out.
There are real cases. In 2023 a Chevrolet dealer’s bot in California, after two instructions from a user, agreed to a Tahoe for one dollar. In 2025 Replit’s agent deleted a production database during a code freeze, when nothing was supposed to change. Both are covered in our piece on the lethal trifecta.
OWASP’s list of risks for LLM applications calls this Excessive Agency. According to OWASP, the root cause is usually one or more of three things: excessive functionality, excessive permissions, excessive autonomy (OWASP LLM06:2025). Our measures follow from that:
We know of no complete protection against attacks hidden in message text (prompt injection), and we do not promise one. Limits narrow what can be stolen and where it can be sent; they do not make the model invulnerable.
The agent replies to the client in their language — Russian, Uzbek in Latin or Cyrillic script, or English — and understands voice messages.
What matters here:
We test Uzbek answers before launch on your real messages, not on a translation of a Russian script. You can see how it looks in the demo on the car dealerships page: the second scenario there is in Uzbek.
An agent is built in five steps: choose one process, describe the scenarios and limits, connect the systems, run a two-week pilot with a metric, and hand the agent over into your environment.
That is how we build custom AI agents. The overall plan, from choosing a process to handover, is in our step-by-step automation guide; where a model pays off and how to choose one is on the AI implementation page.
An agent for one process starts at $1,200 to launch plus $300 a month, with a 2-week timeline; the final price depends on the number of integrations and data sources.
An agent that reads a price list and answers in one channel sits near the lower end. One that writes to the CRM, calendar and 1C and pulls data from several channels costs more.
One process end to end is the Pipeline package: up to three integrations, 2 weeks. Three or four connected processes with your own dashboard and team training is the Shop floor package: from $6,500 plus $700 a month, 4 weeks. Your own model on your data is the Custom package. The full packages and timelines are in the pricing block below. The monthly fee covers monitoring, fixes when data sources break, and changes.
The model itself is not included: you pay the provider directly, at its rates, with your key. How much that comes to each month depends on the number of conversations and the model you choose.
Prices are per project, not per hour. In Uzbekistan you can pay in UZS at the Central Bank rate on the invoice date. There are no discounts: if the budget is smaller, we narrow the scope. In August 2026 Uzbekistan’s President announced that half of enterprises’ AI implementation costs would be reimbursed, but as of late September the payment procedure has not been published — what is known is in 50% of AI costs reimbursed.
For a small business, an AI agent makes sense when clients write a lot, in free text and outside working hours, and the conversation ends in an action — a booking, a reservation, an order. Otherwise a ready-made bot or a simple workflow is usually enough.
An agent makes sense if:
Something ready-made will do if:
If the job can be done with a ready integration or an off-the-shelf model in two days, we will say so: there is no point paying us for development. And we do not take on work where success cannot be measured.
For a salon, a clinic or a Telegram shop, a sensible first agent is booking plus answers from the price list, as in the demo on this page. Describe your process in the quiz or message us on Telegram: within 48 hours we will tell you whether you need an agent or a bot is enough.
We sell a deadline, not a scope of work. If the deadline cannot be named, the task has not been broken down yet — and the first step is exactly about that.
We break one of your processes down step by step and hand back a map: what can be taken off people, in what order and what it costs. Free, and no intro call.
We build one pipeline and run it live. The "before" metric is fixed in writing before the start, not invented afterwards. If we do not hit it, the stage is not paid for.
The code, the access and the documentation are yours. Everything lives inside your own perimeter. Support by subscription or with your own hands — your call; we do not keep you on a hook.
Prices are per project. For Uzbekistan, payment in UZS at the Central Bank rate on the invoice date is possible.
+ $300 per month. One process end to end: data collection, processing, notifications, reporting. Up to three integrations.
+ $700 per month. Three or four connected processes, your own dashboard, team training, priority support.
+ $1,200 per month. Your own product: your data → your model. Satellite, geo, valuation, forecast. Code handover. Price is for the pilot; production is quoted separately.
The first audit is free. 48 hours, no intro call.
A Chevrolet dealer's bot agreed to a $1 Tahoe, and Replit's agent wiped a production database. The three properties that make an agent dangerous, and how to cut that trifecta in a Telegram bot.
In 2024 Klarna said its AI assistant did the equivalent work of 700 full-time agents; in 2025 its CEO admitted quality had dropped. The release, the CBA and Salesforce stories, and a better pilot metric.
The line about a person with AI gets credited to Jensen Huang, the economist Richard Baldwin and HBR. The earliest version we found is a radiologist's 2017 tweet. Four wordings, and data on 5,172 support agents.
What AI can do in Uzbek as of September 2026: two scripts and mixed-language messages, the token “language tax” (measured on FLORES-200, method included), which services offer Uzbek speech-to-text and text-to-speech according to their official language lists, and how to test a vendor in 30 minutes.
Six questions about how things are set up at your place today. The output is a diagram: what can be taken off people, in what order and what it costs.
