First-line AI agent: when a bot actually pays off and when it's just a toy

A first-line AI agent pays off where managers are drowning in a flow of repetitive first-contact inquiries, and almost nowhere else. Gartner predicts that over 40% of AI agent projects will be canceled by the end of 2027 due to rising costs and unclear business value. The fate of a particular bot is decided before launch: is there enough volume, has the cost of the routine work been calculated, and where do the conversations end up.

Below is a breakdown in numbers: how an agent differs from a button bot, why the speed of the first response is worth real money, how to calculate payback before implementation, what the Klarna story teaches, and the signs that a bot will remain a toy.

How is a first-line AI agent different from a button bot?

A button bot walks the customer through a rigid menu and breaks on the first open-ended question. An AI agent understands a question asked in the customer's own words, answers from the company's knowledge base, and steers the conversation toward a goal: a qualified lead in the CRM. Customers feel the difference immediately: instead of "press 1 to find out our opening hours," they get an answer to what they actually asked, with prices, timelines and terms.

Both work around the clock, but differently: at night a button bot collects phone numbers, while an agent answers questions and qualifies leads. An inquiry left at 2 a.m. gets its first meaningful response right away, without waiting until morning.

First-line support is just one type: what other AI agents for business exist and how they work is covered in a separate article.

It's also worth knowing the limits in advance. Complex negotiations, haggling and unusual situations stay with people; the agent's job is to take the mass of first-contact inquiries off the managers' plates so they can focus on selling, and to hand the conversation to a person at the right moment, along with its history.

Why is the speed of the first response worth money?

Because a lead goes cold within minutes. The classic MIT and InsideSales.com lead response management study showed that contacting a lead within the first 5 minutes rather than 30 makes you 100 times more likely to reach them and 21 times more likely to qualify them. An AI agent wins exactly this window: it responds within seconds, at 2 a.m. and on weekends, when a live sales team is physically unavailable.

The other half of the problem is how companies themselves respond. A Harvard Business Review audit covered 2,241 US companies: the average response time to an online inquiry was 42 hours, and 23% of companies never responded at all. Those that responded within an hour were nearly 7 times more likely to qualify the lead than the rest.

These studies are years old, from 2007 and 2011, and their conclusions have only become harsher since: a customer who messages you is messaging two competitors at the same time. Whoever replies first, and with substance, wins.

Where an AI agent pays off: five signs

The economics of an agent work out where there is a large flow of first-contact inquiries with recurring questions: real estate, services, retail, e-commerce. If managers answer the same twenty questions day after day while warm prospects wait in a common queue, an agent pays back the investment within months. Check yourself against the signs below: the more of them match, the faster the payback.

• Managers spend a noticeable part of their day on "just asking" questions, the ones already answered on the website and in the price list.
• The time to first response is measured in hours, and some leads go to competitors in the meantime.
• At night and on weekends, inquiries pile up unanswered until Monday morning.
• Incoming inquiries aren't qualified: warm and cold leads are handled the same way, in the order they arrive.
• Managers enter customer data into the CRM by hand, or don't enter it at all.

Our case with a property developer was built on exactly these signs: a flow of inquiries from people who want to "figure things out first," and managers drowning in first-contact work. The first-line bot explains, answers common questions and warms prospects up, and puts qualified deals into the CRM, so managers only work with warm leads.

If you're still choosing what to automate first, start with the big picture: we laid it out in the article “Business automation: where to start”.

How do you calculate an AI agent's payback before implementation?

Treat it like any other investment. The cost of manual first-line handling follows a formula: number of first-contact inquiries per month × average handling time × employee hourly rate + an estimate of leads lost to slow responses. Compare the total with the cost of implementation and support, and you get the payback period. A project without a clear payback period is better left unstarted, however fashionable AI may be.

An example of the arithmetic:

1. Count the flow: say, 900 first-contact inquiries a month, 700 of which are routine questions about prices, timelines and terms.
2. Estimate the time: at 10 minutes per inquiry, that's about 117 hours of managers' work a month.
3. Multiply by the rate: at 800 rubles an hour, that's roughly 93,000 rubles a month just for answers that are already in the price list.
4. Add the lost leads: how many inquiries go cold during hours of waiting; in most calculations this is the biggest line.

The market gives an upper benchmark: according to a study by SberAnalytics and Sber Business Soft (November 2025, 559 respondents), AI assistants on the first line of support and in call centers cut labor costs by up to 30–40%, and 39% of Russian companies already use AI agents and assistants.

We do this calculation at the first consultation: we go through the flow of inquiries, find where leads and hours are being lost, and set a goal in money terms. Every custom development project of ours starts with it.

What does the Klarna case teach?

That an agent wins at scale and loses on quality if it's left without people. In its first month, the AI assistant of fintech company Klarna handled 2.3 million conversations and took on two-thirds of support chats, a workload equivalent to 700 agents. A little over a year later, the company brought people back to the line.

The first-month figures were impressive: resolution time fell from 11 minutes to under 2, repeat inquiries dropped by 25%, and the estimated impact was 40 million dollars in added profit for 2024. Then complaints about formulaic answers in complex cases piled up, and in May 2025 Klarna started hiring live agents again.

"Really investing in the quality of the human support is the way of the future for us." — Sebastian Siemiatkowski, CEO of Klarna, in an interview with Bloomberg, May 2025.

The lesson for a smaller business is the same as for a fintech giant: the setup that works is a hybrid. The agent handles the mass of routine inquiries, a person picks up the complex and sensitive ones, and the handover between them happens without a break, with the full conversation history. Projects of the "let's replace support entirely" kind and the "let's put up a bot to tick a box" kind fail equally often.

When does an AI agent remain a toy?

When it's deployed because it's fashionable, with no flow of inquiries, no knowledge base and no goal in money terms. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027 due to rising costs, unclear business value and inadequate risk controls. The failure scenarios repeat so consistently that you can spot them before the start.

• There's no flow. With fifty or so inquiries a month, the agent saves a few hours, and the payback calculation from the section above doesn't work out.
• "Agent washing." Gartner estimates that among thousands of "AI agent" vendors, only about 130 offer real agentic capabilities; the rest are rebranded button bots and scripts.
• Nobody owns the knowledge base. Three months later the agent is quoting an outdated price list, and customers notice before the owner does.
• Conversations don't reach the CRM. The bot talked to the customer, the history stayed in the messenger: time is saved, but there's no revenue.
• The goal is phrased as "implement AI," with no metric. There's nothing to prove the effect with, and the project is the first to be cut when the budget shrinks.

Each of these is fixed at the design stage, before money is spent on implementation.

Why is an agent without a CRM money down the drain?

Because a conversation without a deal is worth nothing. An agent that talked to a customer and left the history in a messenger saved a manager's time and lost a sale: nobody will see the inquiry, nobody will set its status, no reminder will fire. What pays off is the whole chain, where each link has its own role.

1. The agent runs the conversation, answers from the knowledge base and collects the data: what's needed, when, what budget.
2. A qualified inquiry automatically becomes a deal in the CRM, with the fields filled in and the full chat history attached.
3. The manager picks up where the bot left off, without "the customer wanted something."
4. The whole funnel is visible: from the first message in the messenger to payment.

This is also where the answer lies to why so many companies say their "CRM doesn't work": managers don't fill it in by hand. An agent removes this problem by design: data gets into the system without a person. Why CRM stalls and how to fix it is covered in a separate breakdown, and how to connect the bot, the CRM and accounting into a single system is shown in the article on integrations.

What about hallucinations and reputational risk?

Restrict the agent to an approved knowledge base and build escalation to a person. An agent connected to a language model "as is" will sooner or later invent a price or a condition for a customer. A working setup is built differently: answers come only from the company's price lists, policies and descriptions, and when asked something outside the base, the agent honestly says it will check and hands the conversation to a person.

• All conversations are logged: you can see where the agent coped and where the knowledge base needs filling in.
• Critical scenarios (complaints, legal questions, haggling) are routed to people right away, with no attempts to "push it through" with the bot.
• The knowledge base grows from the logs: every unanswered question becomes a new knowledge base article.

Tone is part of reputation too: according to Zendesk CX Trends 2025 (a survey of about 10,000 consumers and CX professionals in 22 countries), 64% of consumers are more likely to trust AI agents that communicate in a friendly, human way.

A separate point for companies in Russia is data. Customers' personal data ends up in the knowledge base and the conversation logs, so the system has to run on servers in Russia and comply with Federal Law 152-FZ, Russia's personal data protection law. We put this in the contract: hosting in Russia, access controls, and sensitive data never leaves the client's perimeter.

What should be handed to a person when the agent makes a mistake?

The manager should pick up the conversation where the agent stopped, without asking the customer to repeat themselves. That's why context is passed along with the conversation, and the handover itself is recorded in the CRM.

• The whole chat and a short summary: who the customer is, what they asked, what the agent has already answered.
• The reason for the handover: a question outside the knowledge base, a complaint, haggling, a request to talk to a person, or low confidence in the answer.
• The customer's record from the CRM, if they're already in the database: past orders, the assigned manager, open deals.
• Everything the agent has already promised: a callback time, a quote, sending a document. Otherwise the promise gets lost between the agent and the person.

The customer, meanwhile, needs an honest answer: the conversation has been passed to a specialist, who will reply during business hours or by a stated time. Every such handover is worth reviewing when you read the logs. A recurring reason means the knowledge base is missing an article or an escalation rule is set too broadly.

How to launch an AI agent: six steps

A launch takes weeks, and it rests on one resource: your knowledge base. Connecting an agent to Telegram or WhatsApp is technically quick; the long part is gathering and checking the price lists, terms and standard answers it will use when talking to customers. After that comes a staged rollout, starting with part of the traffic.

1. Build the knowledge base: price lists, terms, policies, common questions and answers. Everything managers repeat to customers every day.
2. Connect the channels: Telegram, WhatsApp, VK (Russia's largest social network), a website widget, Avito (Russia's largest classifieds marketplace), the ones where your customers actually write to you.
3. Link the agent to the CRM: a qualified inquiry immediately becomes a deal with its history.
4. Define the escalation rules: which topics and triggers get passed to a person immediately.
5. Launch on part of the traffic and read the logs: where the agent coped, where it made mistakes, what's missing from the base.
6. Keep adding to the base and expand coverage: with each log review, the agent answers more accurately.

Steps 1 and 5 take the most calendar time and are underestimated most often. A knowledge base without an owner goes out of date within a quarter: assign someone responsible before launch, not after the first complaint.

Build in-house or with a contractor?

Look at the cost of a mistake and the depth of the integrations. Today an agent prototype can be put together in an evening even without a programmer, and at that level it's useful to experiment yourself to understand what's possible. The line falls where customer data, payments and the CRM link come in: beyond that point you need engineering backed by contractual responsibility.

This "tool versus engineering" line applies to any AI in business, not just first-line agents, and we covered it in detail in the article on vibe coding. The logic is the same: a tool gives you speed of experimentation, engineering gives you reliability in production.

There's a middle path too: a standard first-line AI agent as an off-the-shelf solution from our catalog, a quick start adapted to your process, and when the business runs into its limits, the system is extended through custom development.

If the payback signs in this article match your situation, tell us about your task: we'll calculate the cost of manual first-line handling in money terms, design the agent → CRM → dashboards link, and deliver the result under contract, with data kept in Russia and support after handover. The team has 300+ automation and AI projects behind it.

Sources

• MIT / InsideSales.com: Lead Response Management Study, 2007 — the impact of response speed on contacting and qualifying a lead
• Harvard Business Review: The Short Life of Online Sales Leads, March 2011
• Klarna: press release on its AI assistant — 2.3 million conversations and two-thirds of support chats in the first month, February 2024
• Bloomberg: Klarna brings live agents back to customer service, May 2025
• Gartner: over 40% of agentic AI projects will be canceled by the end of 2027, June 2025
• CNews: SberAnalytics and Sber Business Soft study on automation and AI in Russian companies, January 2026
• Zendesk: CX Trends 2025 — consumer trust in AI agents