Where an AI bot genuinely takes the load off
A bot works well when the answer already exists in writing and people ask for it dozens of times a month. Delivery terms, warranty, holiday hours, whether a model is compatible with a particular piece of equipment. Your team answers these from memory, and their memory is no better than a knowledge base.
The second solid use is qualification. The bot asks four or five questions and hands over a filled-in card: location, task, rough budget, timeline. The manager opens the conversation already knowing what it's about, instead of spending the first ten minutes on "tell me more".
The third is searching through a lot of information. A catalogue with several thousand items, technical documentation, internal procedures for staff. AI beats filters and site search here, because people ask in words, not part numbers.
- answers to repeat questions, pulled from your knowledge base
- order status by number - through an integration with your CRM or warehouse
- collecting details for an enquiry and handing it over with a ready summary
- internal answers for your own staff about documents and procedures
Where a bot annoys people and costs you leads
First: a bot standing at the entrance to a complaint. Someone didn't get their order, was charged twice, received a damaged product. They don't need a politeness generator, they need a human within a minute. Simple rule: words like "complaint", "refund", "not working", "scam" trigger an immediate handover, with no attempt to "clarify the details".
Second: a bot that makes things up. An invented price, deadline or warranty condition isn't a funny glitch, it's a promise you either have to honour or walk back in front of the customer. The bot should answer only from your own material, and be able to say plainly: "I don't know, passing you to a colleague."
Third: a bot with no exit. Three rounds of "could you rephrase that" and the customer closes the tab. Set a hard limit: two failed attempts, then a real person or a contact form.
Fourth: a bot pretending to be human. "Hi, my name is Anna" followed by a reply in 0.3 seconds gives itself away instantly and reads like an attempt to deceive. It's easier to say it's a bot and put a "talk to a human" button right next to it.
Does your business need an AI chatbot: check your own message history
You can work this out in one evening, before talking to any developer. Export the last month of conversations and sort the questions by topic - by hand, in a spreadsheet. Then look at the proportions.
If repeat questions with the same answer for everyone make up less than a third, you don't need a bot. It's cheaper and faster to build a decent FAQ page and set up quick-reply templates in your messaging app. If those questions are more than half and the volume is growing, there's something worth automating.
Count separately how many enquiries arrive outside working hours. "The bot answers at night" sounds convincing, but if five people message at night per month, payback will take a while. We regularly conclude that what a client needs isn't a bot but a rebuilt pricing page.
- export the last 200-300 conversations from messaging apps and email
- group the questions and write out the top 10 by frequency
- mark the ones where the answer is identical for every customer
- calculate the share of enquiries that arrive outside working hours
Scripted bots and AI are different tools
A scripted bot is buttons and fixed branches. It's predictable, cheaper to build, and physically incapable of inventing anything. For ten standard questions and booking an appointment that's enough - AI would only add cost and risk.
You need AI when customers write in free-form paragraphs: "here's my situation, would this work for me...". For most businesses the sensible option is a hybrid: buttons for the main paths, with AI stepping in where the question didn't fit the script.
What to ask a developer before you pay
Ask for a demo on your data, not abstract examples: take twenty real questions from your export and see what the bot says. The interesting ones are those with no answer in the knowledge base - that's where you find out whether the bot lies or admits it.
- where the bot gets its answers: your knowledge base or the model's own head
- what it does when it doesn't know - ask to see that scenario
- where conversation logs are stored and who reviews them weekly
- who pays for model requests and what happens if traffic spikes
- what the handover to a human looks like, and what happens after hours
- how to switch the bot off with one button if it starts answering wrong
What makes up the price
There are four parts to count: development and integrations, preparing the knowledge base, paying for model requests, and support. The second and fourth usually get forgotten, and the knowledge base is your work: someone has to gather current terms, prices and procedures into one document and keep it updated.
Our site lists starting prices for Telegram bots and web services so you have a sense of the order of magnitude. The exact quote comes after a brief: a bot that answers from documents and a bot that reaches into your CRM, warehouse and payments differ in effort by a wide margin.
Measure payback in specifics, not "hours saved": how many conversations a month the bot closes without a human, how many minutes of your team's time that used to take, and how many enquiries you previously lost because the reply came the next day.
Don't start by picking a platform, start by exporting your own conversations: if repeat questions are under a third, you need an answers page, not a bot. If they're over half, hand the routine to the bot and route complaints, money and unusual requests firmly to a person, with a limit of two failed attempts. At EFIMOV DEV we usually start with that export - it honestly shows whether you need AI, whether a scripted bot is enough, or whether the problem isn't the messaging at all. Starting prices for bots and web services are on our site, and the exact quote is calculated after a brief.
Frequently asked
Can an AI bot fully replace a salesperson?
No, and it shouldn't be the goal. It handles repeat questions and collects details for an enquiry. Anything involving money, a complaint or an unusual situation should go to a person - otherwise the bot starts negotiating and making promises on your behalf.
What do I do if the bot invented a price and the customer is quoting it back?
You're the one who has to sort that out, so the problem gets solved before launch: the bot answers only from the knowledge base, and points to a page or product card for prices and timelines. Also keep conversation logs - without them you won't even know what it said.
Where should the bot live - on the website or in a messaging app?
Where people already message you. If 80% of enquiries come through messaging apps, a website widget won't change much. Look at a month of your own enquiry stats rather than general arguments about which channel is popular.
How much time does support take after launch?
Plan on reading the logs regularly - at least weekly at first. Customer questions shift, the knowledge base goes stale, and a bot left unattended for six months starts confidently quoting last year's terms.