A breakdown of how an inbound lead system is actually built: which parts it consists of, what happens to an inquiry at every step, how long deployment takes and which numbers tell you it is working.
How a system differs from a chatbot
A chatbot is an interface. It can hold a conversation in one channel and that is where it ends: a human still copies the data over by hand, assigns the task to a manager and tries to remember who needs a call back. That is why businesses get disappointed — the bot is live, but the number of lost leads has not moved.
A system is a chain that carries an inquiry from the first message all the way to a CRM record and a follow-up reminder a day later. The conversation is just one node in it. The value is not that somebody replied to the customer — it is that the inquiry did not vanish afterwards.
A simple maturity test: ask to see what happens to a lead 48 hours after the first message. If the answer is “the manager will check the chat history”, it is not a system yet.
What the system is made of
The architecture is nearly always the same; only the content changes per industry. Five nodes, each solving its own job and each replaceable independently of the rest.
- Capture. Every channel — website, WhatsApp, Telegram, Instagram DM, email, phone — flows into a single entry point. From that moment the inquiry has one identifier you can trace it by.
- Understanding. The model reads the message in the context of the whole thread and the customer history: what the person wants, whether this is a repeat inquiry, which product is being discussed, whether it is urgent.
- Qualification. The system asks the missing questions — politely and one at a time, not as a form — and separates real leads from spam, job applications and idle curiosity.
- Recording. A deal is created in the CRM with fields filled in and the full conversation attached. Manual entry disappears as a category, and with it the typos in phone numbers.
- Follow-through. The manager gets a notification where they actually work. No reaction — escalation. Customer went quiet — a scheduled follow-up rather than one somebody remembered.
How deployment works
We do not start by choosing a model or a platform. We start with a walkthrough: we sit down and map how an inquiry travels through the company today — from which channel to which person, where it waits, where it is lost and who finds out about it.
- Week 1 — process walkthrough, access to channels and CRM, sign-off on the architecture and on the limits of the system’s responsibility.
- Weeks 2–3 — the first loop is assembled on one channel and one scenario. It is already live, but supervised.
- Weeks 4–6 — the remaining channels are connected, answers are tuned on real conversations, escalation rules and reporting are added.
- After that — iterative development. Every month we look at where the system stumbles and retrain it on exactly those questions.
A first working loop usually appears within two to six weeks depending on how many channels and integrations are involved. We advise against launching everything at once: on a single channel mistakes surface faster and cost less.
Which numbers tell you it works
The number of messages the AI handled is a metric for a slide deck, not for a business. Look at these instead.
- Time to first response — the median, not the average, and split out for nights and weekends. That is where most of the loss hides.
- Share of inquiries that reached the CRM. If it is not 100%, a manual step is still in the chain somewhere.
- Share of qualified leads in the total flow — it shows how much manager time was freed up.
- Conversion from inquiry to booking or meeting, before and after launch. This is the only number an owner actually cares about.
- Number of follow-ups sent on schedule. Before automation it is almost always zero.
Common mistakes
The first and most expensive one is trying to make the system sell instead of a human. AI is good at removing routine: capturing, clarifying, recording, reminding. Closing the deal and handling objections stay with the manager — and that is not a temporary limitation of the technology but a sensible division of labour.
The second is automating a process that is broken to begin with. If the company has no rule about who answers an inquiry and how fast, the system will simply deliver inquiries into the same void more quickly. The rule comes first, the automation second.
The third is hiding from the customer that they are talking to an assistant. It always comes out and it always costs trust. An honest line plus an instant handover to a human on first request works far better.
What drives the estimate
We do not publish a price list because no two deployments are alike: a system on one channel with one CRM and a system across six channels with telephony and inventory differ by an order of magnitude. The estimate follows the process walkthrough and depends on four things.
- How many channels inquiries arrive from.
- The number and maturity of integrations: modern CRMs have APIs, in-house systems often do not.
- Scenario complexity: booking a consultation is one thing, matching a property against a dozen parameters is another.
- The size of the knowledge base the system has to absorb — and what shape it is in today.
Frequently asked questions
- Do we need to change our CRM?
- No. The system plugs into what you already run — Bitrix24, amoCRM, HubSpot or an in-house tool with an API. We do not propose changing your tooling for the sake of the deployment.
- What happens if the system does not understand a question?
- It does not invent an answer. The unresolved inquiry is flagged and handed to a human along with a note about what stayed unclear. We review those cases regularly — they are the raw material for retraining.
- Will the customer realise they are talking to AI?
- We always recommend saying so plainly and offering a one-click handover to a person. Concealment gives a short-term effect and long-term reputational damage.
- Who owns the data and the keys?
- You do. Data stays in your infrastructure and your services, keys and tokens live on your side, access is granted minimally and revoked in one step.
