AI for business: where to start and where it actually pays off

If you run a company, the question is rarely "should we use AI?" any more. It is "where do we start, and where does it actually pay off?" The short answer: AI pays off first on repetitive, text-heavy work you already do every day: answering the same customer questions, sorting incoming leads, reading invoices and delivery notes, drafting routine content. It rarely pays off when you bolt it onto a process that was already broken.

This is the practical version of that answer: where AI is worth it in a company of two to fifty people, which ready-made AI tools to try first, when you need a custom AI agent connected to your systems instead, what to consider under Kazakhstan law, and a 30-day plan for a first pilot.

Where AI in business actually pays off

The value is rarely in doing something clever. It is in taking over a task that eats an hour a day from a person who has better things to do. Six places where that is usually true:

• Customer replies. Opening hours, delivery status, prices, returns. An AI support agent answers these from your own materials and order data at any hour, in WhatsApp, Telegram and on the website, in Russian and Kazakh, and hands the rest to a person with the full conversation. If many of those questions arrive by phone, an AI receptionist that answers your calls does the same job on the line. When the questions are simple and repeat every day, a plain rule-based chatbot is often enough to start.
• Lead qualification. An enquiry arrives from a form or a messenger; the agent asks the two or three questions you would ask, records the answers in amoCRM or Bitrix24 and books a call only when the lead fits.
• Document processing. Invoices, delivery notes, applications. The model reads the PDF or photo, extracts the fields you need into your accounting or ordering system, and flags anything it is unsure about.
• Bookkeeping preparation. Matching receipts to transactions, categorising expenses, preparing the monthly pack for your accountant. Not filing the return, preparing it.
• Content drafts. Product descriptions for the website and Kaspi, newsletters, job ads, replies to reviews on 2GIS. A first draft in your tone that a person edits.
• Internal search. "What did we agree with this supplier?" An assistant that answers from your documents instead of a colleague hunting through folders.

Notice what is not on the list: strategy, pricing decisions, anything where being wrong is rare but expensive. AI works best where the task is frequent, the answer is checkable and a person can step in.

Use cases compared: effort, payoff and risk

Read the table by its last column: the right first project is the one whose main risk you can live with while you learn, not the one with the biggest payoff.

What an AI agent is, in one paragraph

An AI agent is a language model with a goal, tools and rules. Instead of only answering a question, it completes a task: it reads the message, looks something up in your CRM or order system, decides what to do next, and either finishes the job or hands it to a person with context. A chatbot follows a script; a chat model answers whatever you paste into it; an AI agent is connected to your systems and acts within limits you set. The longer explanation is in what an AI agent is and what it can do.

Ready-made AI tools: a starter kit by task

Most companies do not need a new platform to start. They need one tool per job, and often that tool is already inside software they pay for. Below is a starting list by task. It is not a ranking and nobody paid to be on it; these are widely used options. Prices and plans change often, so check the vendor's current page, and read its data terms before you paste in anything about customers.

How to choose AI tools

• Start from the task, not the tool. Write down the one job you want off someone's desk, then look for the feature in the software you already use. A new subscription is the last option, not the first.
• One tool per task, one owner per tool. Three overlapping chat assistants mean three sets of data terms and nobody who knows how to work with them.
• Business terms, not a personal account. Personal and free accounts may use your inputs differently from business plans. Pay for the business tier where customer data is involved and read the data processing terms.
• Test on last week's real work. Give the tool ten real emails, invoices or tickets and compare its output with what your team did. That tells you more than any demo.
• Know where the tool stops. Off-the-shelf tools work inside their own product. When the task has to read a Kaspi order, write it to your CRM and reply in WhatsApp in one go, a single tool will not do it. That is the point where a custom agent makes sense.

Off-the-shelf AI tools vs a custom AI agent

When off-the-shelf tools are enough

The most useful tools are usually already inside your software: the assistant in your email client, the reply drafts in your helpdesk, the AI in your CRM. If the task lives in one system, needs no data from elsewhere and a person checks the output anyway, use what you have. The cost is a subscription per seat, not a project.

When you need an agent connected to your systems

The moment the task spans systems (read the message, check the order in the shop, update the CRM, reply in WhatsApp), a generic tool stops working: it cannot see your data or act on it. That is where a custom agent earns its keep: built on your knowledge base, connected to the systems you already use (1C, Kaspi, amoCRM or Bitrix24, MoySklad), with hand-off rules and a log of every action. This is AI automation in the practical sense: not a new platform, but your existing systems doing more on their own. A custom AI agent starts from 480 000 ₸, the price is fixed in the contract, with a six-month warranty. If you need several systems connected without the conversational part, that is business process automation.

Rule of thumb: a tool for a task inside one system, an agent for a workflow across several.

Data and the law: what to consider in Kazakhstan

None of this is a reason to wait, but it is a reason to set things up properly from the start:

• Know which data leaves your systems. Anything about an identifiable person (a name, a phone number, a conversation) is personal data under the Law of the Republic of Kazakhstan "On Personal Data and Their Protection".
• Keep the database in Kazakhstan. The law requires personal data to be stored in a database located in the country (Article 12, paragraph 2). We keep the agent's conversation logs and requests on a server in Kazakhstan.
• Document the transfer to the model. Large models run abroad, so a request containing customer data is a cross-border transfer: it is named in the consent to data processing. How to set up the consent is in our guide to personal data on a website.
• Send only what the task needs. An invoice reader does not need the customer's full history.
• Tell people. The Law "On Artificial Intelligence" No. 230-VIII requires users to be informed when a service is provided using AI systems (Article 21, paragraph 1). Customers should know they are talking to an assistant and how to reach a person.
• Keep a human route for decisions that matter, such as refusing an instalment plan or rejecting an application: for decisions based solely on automated processing of personal data, the AI law refers to the personal data legislation (Article 21, paragraph 4). And decide up front how long logs are kept.

We build this into the architecture by default. These are an afternoon's decisions, not a project of their own; check the details of your case with a lawyer.

How to use AI in business: a 30-day plan for one task

The question of how to use AI in business gets much easier when you refuse to answer it for the whole company at once. One task, one month, one decision.

Week 1: pick one repetitive task

It should happen daily, take text or documents in and text or data out, have an answer you can check, and be cheap to correct when wrong. Write down what it costs today: items a week, minutes per item, who does it.

Week 2: measure the baseline and set the bar

Record volume, time per item, error rate and response time for a normal week. Decide what "good" means before you see any results: for example, half the items handled without a person and no more errors than today. Write the hand-off rule: which cases always go to a human.

Weeks 3–4: pilot with a person in the loop

Run the AI on real items, but a person checks every output before it goes out, or the agent runs in shadow mode alongside the current process. Log every correction. Update the knowledge base or the rules once a week, not every hour.

Day 30: decide

Three honest outcomes. Expand: remove the check for the clear cases. Fix: the task is right but the data or rules need work, so run two more weeks. Stop: the payoff is not there, and you learned that for the price of a month. Only then pick the second task.

Mistakes to avoid

• Automating a broken process. If nobody in the company can state the returns policy, the agent cannot either. Fix the process, then automate it.
• No human handover. Every agent needs a route to a person and a rule for when to use it. Customers forgive "let me connect you to a colleague" far more readily than a confident wrong answer.
• No measurement. Without the baseline week you cannot tell whether the pilot worked.
• Starting with the hardest task. The impressive project fails slowly and expensively; the boring one succeeds in a month and funds the next.
• Treating launch as the end. Prices and policies change; somebody has to own the knowledge base or the answers drift.

What it costs

Off-the-shelf tools cost a subscription per user; many business software packages now include AI features in higher plans, so check what you already pay for before buying anything new. A custom AI agent is a project: an agent on one channel starts from 480 000 ₸, a voice agent for calls from 540 000 ₸. The price is fixed in the contract, with a six-month warranty, and the code and access stay with you. Model usage is billed separately, at cost with no markup, and depends on how many requests the agent handles.

If you are not sure yet which task to start with, that is what AI consulting is for: we look at your real work and tell you where AI pays off and where simple rules or a tool you already own will do.