Many companies lose clients in the evening. Not because the offer is wrong - because at 10:40pm, when the client finally has time to compare, there is nobody on the other side.
Short answer: an AI agent is not a chat bubble. It is a front-line worker trained on your offer, your prices and your process. It answers immediately, qualifies the enquiry and books a slot, so in the morning you have a lead rather than a missed message.
Kluczowe wnioski
- An AI agent wins on reply time rather than answer quality - the client wrote to several firms and talks to whoever answers first.
- The difference from a chatbot is the source of its knowledge: scripted rules against your offer, prices and process.
- Qualifying before the call is worth more than the answer itself - your salesperson rings prepared.
- Implementation starts by writing down what you already say on the phone.
What an AI agent does on your site
An AI agent draws on a knowledge base about your business - the offer, prices, process, lead times - and holds a conversation like a competent employee. It is not infallible, though, so it matters that it says plainly what it does not know and hands the matter to a person.
- It answers questions about scope and price in the client's own words.
- It qualifies: it tells a browser from a buyer.
- It books meetings straight into your calendar.
- It records the conversation and hands it to your CRM - with a privacy notice for the client, in line with GDPR.
The difference from an ordinary chatbot comes down to one thing: a chatbot knows a script, an agent knows your company. Which is why the first question at implementation is not "which model" but "what exactly do you tell clients on the phone". If you are still choosing between an off-the-shelf tool and something custom, we compare the options and prices in A chatbot for your website.
Why reply time beats answer quality
Somebody looking for a contractor or a flat often writes to several companies at once and talks to whoever answers first - not always to whoever answers best. Research on response times, such as the study published in Harvard Business Review in 2011, shows that a fast reply clearly raises the chance of a conversation.
Answer
A specific reply on price, timing or scope, in seconds.
Qualification
Budget, timing, location. Three questions, not a questionnaire.
Booking
A free slot from your calendar, confirmed by email.
Handover
The whole conversation in your inbox or CRM before morning coffee.
That is why an AI agent earns most in industries where enquiries arrive outside office hours: evenings, weekends, the commute home. We describe the mechanism in detail for real estate in how agents get clients, and for medical services in aesthetic clinic websites.
What the agent should do, and what it should refuse
The most common implementation mistake is trying to teach the agent everything. An effective agent does four things and declines the rest:
- Answers questions with unambiguous answers - scope, indicative price, lead time, service area.
- Collects what the salesperson needs - budget, timing, location, scale.
- Books a slot in a real calendar, with confirmation.
- Hands over when the question goes past what it knows.
And, more importantly, it does not negotiate price, promise deadlines, or advise on anything carrying professional liability. In a law firm or a clinic, the agent's job is to take the details and book a consultation, not to give advice. That boundary is part of the implementation rather than a limitation of the model.
Where its knowledge comes from
Three sources, in this order:
- Your website - offer, prices, service descriptions, FAQ. The better it is written, the less work is left.
- An internal knowledge document - what is not on the site but gets repeated on the phone anyway: exceptions, lead times, how far you travel, what you do not do.
- Conversation history - after the first month you can see what clients actually ask, and that is the best source of corrections.
The last point is why an agent nobody reviews after a month stays exactly as good as the day it launched.
What implementation looks like
- Week 1: writing the knowledge down - offer, prices, common questions, the limits of what it may answer.
- Week 2: configuration, calendar and inbox connected, tested on real questions.
- Week 3: live on the site, watching the first conversations.
- After a month: reading transcripts and fixing where the agent asked something pointless or handed over too early.
This is not a quarter-long project. The longest part is usually the first, because it means collecting knowledge that currently lives in two people's heads.
What it costs and when it pays back
In our model the AI agent runs on the AI Employee subscription - 4,500 zł a month, LLM token costs included. Before you work out the payback, do it honestly: what counts is the margin, not just the revenue from a client, and only the clients you would not have won without the agent. If you get a few enquiries a month and answer them promptly, an agent usually will not pay off.
FAQ
Will clients know they are talking to AI?
They should - and must. Since 2 August 2026 the EU AI Act requires that a person talking to an AI system is clearly told so, so calling it an "assistant" is not enough. Clients rarely mind, as long as they get a specific answer faster than a human would give it tomorrow.
What if the agent does not know something?
It says so and hands the matter on, having first taken the contact and the context. That beats an invented answer, and it is the difference between an agent that was configured and one that was merely switched on.
Will it replace a salesperson?
No. It replaces an empty inbox at 10:40pm. The salesperson gets a lead with a budget, a timeline and a booked call, instead of starting from "hello, how can I help".
Do I need a new website for this?
No, it runs on the site you have. But if that site has no prices, no description of the process and no answers to common questions, the agent has nothing to learn from - and the first step is sorting the content out rather than deploying anything.
Summary
An AI agent is not a gadget. It answers one specific loss: the enquiries that arrive when nobody is there. It wins on reply time, qualifies rather than merely replying, and hands the salesperson a finished lead. Start by writing down what you already say on the phone - that is the whole training set.