AI chatbots for your website and WhatsApp that know when to hand over
Foundreco builds AI assistants that answer from your own content, capture the lead into your CRM, book the call, and pass anything uncertain to a person. On your website, in WhatsApp, by SMS or on the phone.
Last updated:
What does a good AI chatbot for a business actually do?
A good AI chatbot answers questions from the business's own content rather than from general knowledge, captures the visitor's details into the CRM, books a call when the visitor is ready, and hands anything it is unsure of to a person. Foundreco builds these on OpenAI, Claude or Gemini, for websites, WhatsApp, SMS and phone, as a fixed-price project scoped after a free audit.
The failure everyone has met is the bot that answers confidently and wrongly, or loops "I didn't understand that" until the visitor leaves. Both come from the same design mistake: a model with no grounding in your content and no rule for when to stop. The build below is mostly about those two things.
Where can the assistant run?
| Channel | Good for | The thing to know |
|---|---|---|
| Your website | Answering pre-sales questions and capturing leads while the visitor is still on the page | It adds weight to every page, so it should load after the page does |
| Markets where customers already message businesses there | Outside a 24-hour window, only approved templates can be sent (see below) | |
| SMS | Reminders, confirmations and short replies | Runs on a phone number from a provider such as Twilio |
| Phone | Answering and booking calls in a natural voice | Covered as AI call and voice agents, with its own consent rules |
| Support inboxes with repeated questions | Covered in the project below |
On WhatsApp specifically, Meta's rule is that a business can send free-form messages only inside a customer service window of 24 hours after the customer last wrote; outside it, only template messages can be sent. A WhatsApp assistant has to be designed around that, or its follow-ups simply fail.
How Foundreco builds one
- Your content first. The pages, FAQs, policies and price lists the assistant is allowed to answer from, split into passages and indexed with embeddings so it retrieves the right one. LangChain or n8n handles the retrieval.
- A model chosen for the job. OpenAI, Claude or Gemini, picked on answer quality for your content and on running cost, not on brand.
- A confidence rule. If the retrieved content does not clearly answer the question, the assistant says so and offers a person, rather than guessing.
- The lead goes somewhere. Name, contact and the conversation itself land in HubSpot, GoHighLevel, Airtable or a Google Sheet, and the team is notified.
- Booking. When the visitor is ready, it offers times from your real calendar.
- Logs you can read. Every conversation is kept, so you can see what people ask that your website never answered.
That last point is underrated. The questions a chatbot cannot answer are a free list of what is missing from your site.
The same pattern, delivered: AI email support
AI email support and FAQ replies in n8n
Each incoming Gmail message is classified by intent, matched against a knowledge base using embeddings, and answered by a response model that returns a confidence score. Confident answers are sent automatically and logged; uncertain ones go to a review queue and the support team is alerted in Slack. A weekly digest of the logs goes out by email and Slack.
- New message
- Classify intent
- Search your contentembeddings
- Draft the answer
- Confidence check
- Reply and log
- Or: hand to a personSlack
A website chatbot is this workflow with a chat window in front of it instead of an inbox. The n8n automation page has the full build.
Why not just install an off-the-shelf chat widget?
For some businesses a widget is enough, and it is cheaper. This site ran one, and two things about it are worth knowing before you choose. It added about 588 KB to every page, which is more than the rest of the homepage's images put together, and on its free plan it hit its view limit and replaced itself with a red "temporarily deactivated" box, shown to every visitor, with nothing alerting anyone that it had happened. It was removed.
A built assistant costs more up front. In exchange it loads when you decide, answers only from your content, writes leads into your own CRM, and cannot switch itself off because a plan ran out.
What it connects to
What it costs to run
Two costs, and you see both before you commit. The build is a fixed price. Running it means AI model usage, which grows with the number of conversations, plus any channel fees: WhatsApp, SMS or a phone number. Those are estimated from your expected volume in the audit, so the monthly number is known in advance rather than discovered on an invoice.
Questions about AI chatbots
How much does an AI chatbot for a website cost?
Will the chatbot make things up?
Can the chatbot work on WhatsApp?
Which AI model should my chatbot use?
Can it book appointments and add leads to my CRM?
Send your ten most asked questions
You get back how an assistant would answer them, where it should hand over to a person, and what it would cost to build and to run each month.