AI Chatbot for Your Site: Leads at Night and Weekends
How many leads a slow reply costs, what an AI bot on RAG really handles, how to build one in 2–3 weeks, and when a plain form is enough.
Short answer: an AI bot isn’t “a toy with artificial intelligence” — it’s a way to close three specific funnel leaks: answering out of hours, answering in the first seconds, and the repetitive questions that eat employee hours. Payback is a five-minute math: missed enquiries × conversion × average order value. It ships in 2–3 weeks if you have a knowledge base and a CRM.
Where the money leaks
60–70% of enquiries in services and e-commerce arrive — or need a reply — out of hours and on weekends. A customer won’t wait until morning: they open a competitor’s tab. A bot doesn’t “replace sales” — it stops leads from going cold.
Count the losses before you build
The formula is simple: enquiries per day × share without a fast reply × conversion to order × average order value × 30.
| Parameter | Store (example) | Agency/service (example) |
|---|---|---|
| Enquiries per day | 60 | 12 |
| No reply in first 5 min | 25% | 40% (evenings, weekends) |
| Conversion of a “cold” lead | 3% instead of 12% | 5% instead of 25% |
| Average order value | $40 | $1,100 |
| Monthly loss | ≈ $650 of revenue | ≈ $11,000 of revenue |
This isn’t scaremongering — it’s why the first bot is deployed where there’s a stream of repetitive questions, not where it’s “trendy”.
Checklist: is your business ready for a bot
Answer “yes” to at least 5 of 7 — and the rollout will pay off:
- Repetitive questions: prices, availability, timelines, delivery, order status, warranty terms.
- A knowledge base exists, even if messy: price lists, policies, FAQ, manager email history.
- Leads come from several channels: site, Telegram, WhatsApp, Instagram, email, phone.
- Managers physically can’t reply in the first minutes.
- There’s a CRM or sheet where the bot can write leads (otherwise they vanish “in the chat”).
- Enquiries come at night, on weekends, in peak season.
- You’re ready to delegate part of the decisions: qualify, reschedule a visit, send a quote template.
What different “bots” can do — and what you don’t need
| Type | What it does | When it’s enough | When it’s not |
|---|---|---|---|
| Buttons & scripts | branching “price → package → lead” | 5–10 routine questions, a narrow service | free-form questions |
| Form widget | captures a contact, alerts a manager | low volume, replies are on time | you need an answer right now |
| RAG bot on a knowledge base | answers from your documents, semantic search | high volume of routine enquiries | deep expertise, bespoke quotes |
| AI agent with actions | creates a deal, checks stock, places an order, drafts a reply | processes documented, API/CRM exists | processes live “in people’s heads” |
If your processes aren’t documented, start with an audit and a map, as in the AI automation guide — otherwise the bot is “pretty but useless”.
Architecture: how it works without magic
- Knowledge base. Price lists, service descriptions, FAQ, policies, chat history → cleaned, chunked, indexed in a vector store (Qdrant or pgvector).
- Model. OpenAI or Claude: understand the question, assemble an answer from retrieved chunks, ask clarifying follow-ups. The model swaps in one config line — no vendor lock-in by design.
- Escalation rules. The bot hands to a human: payments, complaints, non-standard quotes, “let me talk to a person”, three low-confidence answers in a row — with full context.
- Integrations. CRM (create deal and contact), warehouse/ERP (stock), site and messengers (Telegram, WhatsApp), telephony and email.
- Logs and dashboard. Which questions come in, which the bot fumbles, how many dialogues it closed, how many leads reached a deal, token spend.
- Security. Access roles, no personal data in prompts, data stays on your own store.
Launch in 5 steps and 3 weeks
Week 0 — analyze dialogues. Export 300–500 real enquiries from the last 2 months and cluster the topics. That’s your spec, not “what should we ask the AI”.
Week 1 — clean the knowledge base. Outdated price lists and “three versions of one price” are the main cause of bot hallucinations. One source of truth.
Week 2 — prototype on 20 questions. Internal test: managers ask real questions, you review answers and fix the prompt and documents.
Week 3 — production and integrations. Widget on the site + Telegram, lead capture into CRM, escalation, logging, dashboard.
Ongoing — retrain. Every week review the top-10 “bad” dialogues: the fastest way to raise quality.
What kills the project
A knowledge base of stale documents: the bot confidently lies about prices and timelines.
A bot “in a vacuum”: without CRM capture, leads dissolve and the effect is zero.
Trying to fully replace the manager: customers feel the dead end and go to whoever answers.
No metrics: without % of closed dialogues and leads that reached a deal, you won’t know if it paid off.
Public access to an “internal” bot with no guardrails: a commercial-data leak.
Economics: rollout and monthly costs
| Line item | Typical range | Note |
|---|---|---|
| Rollout with RAG and integrations | from $1,807 (₴75,000) | includes dialogue audit, knowledge base, dashboard |
| API tokens | $50–$400/mo | depends on volume: 1 dialogue ≈ $0.10–$0.25 in model calls |
| Hosting and vector store | $25–$100/mo | for your own data and logs |
| Support and retraining | $200–$500/mo | weekly dialogue review, new documents |
Token spend is visible in the dashboard from day one: you always know the cost of a closed dialogue and can compare it to the cost of a manager’s minute.
FAQ
Will the bot start lying to customers about prices and timelines?
Without RAG and document control — yes. With RAG the model must answer from retrieved chunks of your base, and anything not there goes to escalation. That’s why week 1 in the plan is source cleanup, not “AI tuning”.
We get 5 enquiries a day — is it worth it?
Usually no: payback appears from 20–30 enquiries a day, or with a very high order value (10 enquiries at $1,100 each is already a case). At low volume, start with an automated flow “form → CRM → messenger alert → auto email”: days of work, same effect on lost leads.
Can I just install an off-the-shelf bot from a marketplace?
You can, if there are under ten questions and they don’t change. The limits: no your-data in answers, no stock/CRM integration, no dialogue audit. A boxed bot is an “auto-responder”, not an assistant on your base.
How much of the first line does it take over after launch?
In practice: the bot closes 60–80% of routine enquiries, the manager handles only escalations. That’s not headcount cuts — it’s reallocation: from answering “how much” to negotiation and deals.
The one-line rule
An AI bot pays back when you count not “the technology” but three numbers: how many enquiries go unanswered, their conversion, and the average order value. Everything else is technical — a matter of 2–3 weeks and transparent token economics.
Order an AI bot rollout → — we’ll start by analyzing your real dialogues and calculating lost revenue, then propose a solution. Service details are on the AI integration page.
bot MVP from $723 (₴30,000) · AI · Standard from $1,807 (₴75,000) · AI · Premium from $3,614 (₴150,000)
Need the same solution for your business?
AI agents inside your product: a chatbot for your site and messengers, automated request and document handling, smart answers and suggestions. OpenAI, Claude and RAG over your own data — wired into your CRM. And a ready-made DEXA Chat: enquiries and the whole conversation land in your Telegram.
- An AI chatbot for the site, Telegram and WhatsApp
- Automation of repetitive tasks on n8n / Make
- RAG: answers from your documents, voice and doc recognition
Get a free website audit in 48 hours
Speed, SEO, conversion — a 12-page report with a prioritized fix plan. No obligation, no spam: we show what to improve even if you never hire us.
