How Much an AI Assistant for Handling Inquiries Costs: an Honest 2026 Breakdown

Ask five vendors and you will get five different answers: 30 thousand, 150 thousand, 700 thousand, "from a million" and "we need to discuss it". The uncomfortable part is that all five can be telling the truth. They simply mean different things by "AI assistant": anything from a bot that replies from a list of prepared answers to a system that reads your knowledge base, holds a conversation, creates deals in your CRM and hands the manager an already warmed-up customer.

This article is not about naming one number. It is about breaking the price into layers, seeing what exactly you pay for, and working out whether it pays off in your case. After it you will be able to talk to a vendor in the same language and notice when you are being sold something you do not need.

Why the spread is so wide

Imagine asking how much a car costs. The answer depends on whether you need to reach a summer house or haul cargo every day. Assistants are the same, except the boundaries are less obvious, because from the outside every solution looks identical: a chat window with someone answering in it.

The difference is hidden inside. It lies in where the assistant gets its answers, what it can do besides talking, and what happens when it is wrong. Let us go layer by layer.

Layer one: a scripted bot

The simplest option. You write the questions and answers in advance, the bot delivers them. There is no language model inside, so there is no "artificial intelligence" in any strict sense either — though it is very often sold under exactly that word.

• What it can do: answer frequent questions, capture a contact, book an appointment through buttons.
• What it cannot do: understand a question phrased differently from what you anticipated. The customer writes "are you open on Sunday" while the script has "opening hours" - and the bot does not connect the two.
• Price from published rates of Russian studios: roughly from 30 thousand roubles one-off for a simple FAQ bot, from 90 thousand for a bot with sales logic.
• Running costs: almost nothing, from about one and a half thousand roubles a month for hosting.

This is a valid option if you genuinely have a dozen typical questions and accept that the eleventh will get a wrong answer. For booking a barbershop it is enough. For handling inquiries about a complex service it is not.

Layer two: an assistant on a language model

Here there is a real language model inside. It understands a question phrased any way at all and answers coherently. The key difference from the previous layer: the assistant does not pick from prepared answers, it composes a new one.

This is where the main technical task appears — the one you are actually paying for. The model on its own knows nothing about your business: not your prices, not your delivery terms, not the fact that you do not work with companies. To make it answer about you rather than about the world in general, it needs your knowledge base and a way to search it. That is a separate piece of engineering.

• What it can do: hold a natural conversation, answer from your materials, clarify details, work around the clock.
• Price: roughly from 120 thousand roubles for a basic deployment, from 200 thousand for a version with integrations.
• Running costs: here a charge for model calls appears. It is measured by volume of text and, for a few hundred conversations a month, is usually counted in thousands of roubles rather than tens of thousands.

Do not ask a vendor "do you have AI". Ask "where does the assistant get facts about my company, and what does it answer when the fact is missing". The answer to the second question shows immediately whether anyone thought about it at all.

Layer three: an assistant built into your processes

Talking is not enough. Value appears when the assistant does something inside your systems: creates a deal in the CRM, assigns a task to a manager, checks stock, books a calendar slot, sends a reminder the next day.

Every such connection is separate work. Not because it is technically hard, but because you have to decide what happens when it fails. If the CRM is unavailable, the inquiry must not vanish - it should queue and go through later. If the customer changes their mind, the deal has to close. Details like these make up most of the budget.

• What it can do: carry an inquiry through to a result without a human, and pass on only what genuinely needs one.
• Price: from 500 thousand roubles upward, depending on how many systems are involved and what state they are in.
• A separate line item: if your CRM is a mess, that has to be cleaned up first. This is a normal situation, but it belongs in the budget from the start.

Layer four: the enterprise perimeter

The same thing again, but with the requirements large companies bring: data must not leave the perimeter, every action must be logged, access must be segmented, the system must survive failures and must be protected against the model inventing things.

By estimates published by RBC, setting up a language model together with a system to manage it starts at roughly 700 thousand roubles, and full enterprise deployments run into millions. At this level you are paying less for the assistant than for the infrastructure around it.

Your own build or a ready-made service

A fork everyone reaches. Both paths have a price you see immediately and a price that shows up six months later.

A subscription service

You pay monthly and configure it yourself through an interface. Almost no barrier to entry, live within a week. Good for testing whether people will write to a chat at all.

• Upside: fast, cheap to start, nothing to maintain.
• Downside: you live inside someone else rules. A non-standard integration is often simply impossible rather than more expensive.
• Main risk: customer data sits with the provider, and the provider changes the terms. Prices can rise, features can disappear.

Your own build

More expensive upfront, but the system is yours: any integration, your own data, no dependence on someone else roadmap.

• Upside: does exactly what you need and stays yours.
• Downside: needs maintenance. Without it the system decays within a year.
• When it makes sense: once the process is settled and it is clear what exactly you are automating.

A sensible order: a ready-made service for a couple of months to see the real questions customers ask, then your own build for what proved out. Starting with a custom build before you know what people will ask means designing blind.

How long it takes

Asked less often than price, and wrongly so: deadlines slip mostly because preparation was not accounted for.

1. Collecting the knowledge base: from a few days to a couple of weeks. Entirely on your side, and most often the bottleneck.
2. Development: from one week for a simple bot to six or eight weeks for a system with integrations.
3. Testing on real inquiries: two to three weeks. Not skippable — this is where the questions nobody anticipated show up.
4. Reaching a steady state: another month. The assistant is tuned on real conversations and you correct the wording.

So from "let us do it" to "runs without supervision" usually takes two to three months even on a simple project. Promises of a week apply to the first layer described above.

Costs nobody mentions at the first meeting

Development cost is what appears in the proposal. Then come the things that surface later.

1. Preparing the knowledge base. Most companies have no document stating what they sell and on what terms. The information lives in people heads and in chat threads. Turning it into coherent text takes days, and somebody has to do it.
2. The first month of fixes. The assistant meets real customers and inevitably runs into questions nobody foresaw. This is not a defect, it is a normal stage. Budget time and money for it.
3. Model call charges. They grow with your customer flow. A pleasant problem, but worth estimating in advance: ten times the conversations means roughly ten times the bill.
4. Support. Models get updated, integrations break when the CRM changes, requirements shift. An unattended assistant decays within six months.
5. Training your team. Managers need to know what to do with an inquiry the assistant hands over, and how to correct it when it gets something wrong.

How to work out whether it pays off

What matters is not "how much does it cost" but "how much does doing nothing cost". The formula is simple and needs three numbers you already know about your business.

• How many inquiries arrive outside working hours. Check a month of statistics: evenings, nights, weekends. For most businesses that is a quarter to a third of everything.
• What share of them is lost. A customer answered in the morning has often already gone to whoever answered immediately. Use your real conversion from inquiry to conversation.
• What you earn on one deal. Average order value times margin.

Multiply: night-time inquiries per month, times the share lost, times conversion to a deal, times profit per deal. The result is what you lose every month simply because nobody answers at night. Compare it with the deployment price and the payback period becomes obvious.

If the formula says you lose 20 thousand roubles a month, an assistant for 500 thousand pays back in two years. That is a poor investment, and an honest vendor will tell you so themselves.

How to tell whether it works

Counting conversations is pointless: the number grows on its own and says nothing. Look at other things.

• Share of conversations completed without a human. The key figure. If the assistant escalates nine out of ten, it is not working, it is forwarding.
• First response time at night and on weekends. This is what the whole exercise was for.
• Conversion from conversation to inquiry. Compare with what managers achieved before.
• Share of conversations where the assistant answered incorrectly. Only measurable by reading transcripts by hand. Nobody gets this figure automatically, and a vendor promising it out of the box is bluffing.

Agree target values before work starts and put them in the contract. Otherwise, two months in, the argument about whether it works will come down to feelings.

When an AI assistant is not needed

There are more such cases than is usually admitted.

• Few inquiries. At five a week a manager copes and there is nothing to automate.
• Every deal is unique. Project work negotiated from scratch each time maps poorly onto scripts.
• No described process. If nobody inside the company can explain what happens to an inquiry after it arrives, there is nothing to automate yet. Process first, tool second.
• The problem is not response speed. If customers leave because of price or product quality, an assistant will not fix that — it will only deliver the bad news faster.

What to ask a vendor before signing

A short list that saves months. The answers separate those who have done this before from those learning at your expense.

1. What does the assistant answer when it does not know? The right answer is that it says so honestly and hands over to a human, rather than inventing something.
2. Where are customer conversations stored and who has access to them?
3. What happens to an inquiry if the CRM is unavailable at that moment?
4. What will running costs be at the current volume, and at twice the volume?
5. Who owns the knowledge base and the configuration after handover? Can you move to another vendor without starting from scratch?
6. What is included in support and what counts as billable extra work?

Frequently asked

Will customers realise they are talking to software?

Probably yes, and that is fine. The problem is not that they realised, but whether you were pretending. An assistant that introduces itself honestly and answers quickly annoys people less than a manager who replies a day later. Trying to pass software off as a person becomes obvious by the third message and leaves a bad taste.

Will we have to lay people off?

In practice something else happens: routine comes off the managers and they start working the deals that genuinely need a human. If your inquiry flow grows, you will more likely need the same number of people producing more. Headcount reduction as the goal of a deployment usually means the maths was done wrong.

What if the assistant is rude or promises something we cannot deliver?

This is the real risk, and it is solved by constraints rather than promises: the assistant must not be able to quote prices that are not in the knowledge base or give guarantees. Ask the vendor how exactly that is prevented. "We asked the model not to" is a bad answer — a request is not a constraint.

Can we start small?

You should. Take one channel with the most inquiries and one task: for example, answering price questions and booking a consultation. Within a month you will see what people actually ask, and the next step will rest on data rather than assumptions.

In short

The price of an AI assistant is set not by how "smart" it is but by how much has to be built around it: the knowledge base, integrations, failure handling, support. A simple scripted bot costs tens of thousands and solves a narrow task. An assistant that genuinely carries an inquiry to a deal costs hundreds of thousands and requires that you have a described process.

Start not by choosing a vendor but with two numbers: how many inquiries you lose outside working hours and what each one is worth. Until you have those, every price will look equally arbitrary.