Someone is reading your pricing page at 10 pm. Nobody is there to answer.
A homeowner has your site in one tab and a competitor's in the next, with one question about price, area, or whether you do that job. If the answer is buried, they close your tab. The widget answers that question from what your site and price list already say, and hands them a next step while they're still reading. Here is the order things happen in.
- Page load
The launcher waits. A small button sits in a corner of your site. It doesn't pop open and it doesn't message anyone first. When a visitor clicks it, the greeting says it's a virtual assistant before a single question is answered.
- First question
It reads your material. The closest passages from your own pages and documents go to the model with the question. If they don't settle it, the agent searches again, trying a second phrasing before it gives up.
- Before display
The reply is checked. The draft goes through the review pass described below, and only an approved reply reaches the chat window. The typing effect the visitor sees is added afterwards, on their screen.
- Wants a quote
It takes the contact. When someone asks for a quote, a callback, or something your team has to look into, it asks for a name and the best way to reach them, then saves the contact in HubSpot, or posts it to whatever system you use that takes webhooks.
- Handoff
A person takes over. If the answer isn't in your material, the visitor asks for a human, or money is in dispute, it hands off. A Slack alert brings you the reason, the contact, and a short summary. Alerts go to Slack. If you don't use it, they go to a channel I watch and I pass them on. The visitor sees a line saying a team member has been notified and will follow up.
A sample conversation
A fictional plumbing company, Sandhollow Plumbing, with an $89 diagnostic fee and a one-year warranty on its own work. Two questions the documents answer, one they don't.
A team member has been notified and will follow up.
What lands in your Slack
Handoff: Sandhollow Plumbing
Handoff requested (normal)
Reason: asks whether we'd redo another company's repair under our warranty
Customer contact: Tom Alvarez, 555-0134
Visitor asked about the diagnostic fee and Georgetown coverage (answered from the FAQ). A pipe repaired by another company in the spring is dripping again, and he wants to know if we'd fix it under warranty. Our documents only cover our own work, so this needs a person.
The format is the agent's own: urgency, reason, contact, then a one-paragraph summary. You call Tom back; the widget doesn't try to settle a warranty question it has no document for.
Three things people expect it to be.
Most AI chatbots for small business websites are either live-chat software with an AI add-on or a builder you train yourself. This is a managed agent: I set it up and run it, and it answers under your name.
It is not an inbox you answer from.
Nobody on your side watches the chat window. When a conversation needs a person, it leaves the widget: the alert and the contact come to you, and you reply by phone or email. If you have staff who want to jump into chats live, a live-chat product with an agent inbox is the better buy.
It is not a chatbot builder.
Self-serve builders cost less and suit anyone who enjoys tuning a bot. The money here buys the setup, the rehearsal, a check on every reply, and a person reading the conversations early on. If you'd rather do that work, a builder is a fair choice.
It is not a pop-up offer.
It won't slide open after ten seconds with a discount. Wording like percent-off, coupon, promo code, and limited-time offers is blocked in its replies by a rule in code. It answers the question the visitor brought, and that is all.
From signed agreement to a live widget.
A website widget doesn't need a phone number, so there is no carrier approval to wait on. The pace depends on two things: how quickly your intake comes back, and how many rehearsal rounds it takes before the answers sound like you.
- 01
One link covers the paperwork. Paperwork is one link. Read the terms, type your name to sign the service agreement, and pay the $850 setup invoice. Work starts when it clears.
- 02
You answer the intake. Your website, hours, service area, services, pricing notes, and booking link. How you greet people, the words you always use, the promises you never make, and whose phone the urgent ones go to. Upload price lists, flyers, or an FAQ document; PDFs and text files are indexed alongside the site.
- 03
It learns your site. The console reads as many as 25 pages from your domain, the useful ones first (services, pricing, FAQ, service area, warranty, financing, contact), and skips anything your robots.txt blocks. From that it suggests profile details, a draft FAQ, and review rules. I approve each suggestion by hand; nothing is applied on its own.
- 04
You read the rehearsal. The agent answers a set of customer questions drawn from your own material, with the real review switched on. You read them and flag any that sound off. Your fixes go in before a single visitor sees the widget.
- 05
You paste one line. The snippet is a single script tag. Most site builders have a box for it in the footer or header settings, or it goes in your tag manager. The server is set to take chats from your own site's address, and on a phone the chat opens full screen.
- 06
Live, with a second reader. For the first weeks every reply is copied to me as it goes out, along with the reviewer's verdict. An answer that reads wrong turns into a correction or a house rule, so the next visitor doesn't get it.
Once it's running, a read-only report link shows how many conversations it handled this month and the usage so far. The words of each conversation are dropped after 90 days; the counts stay.
Every reply is checked before the visitor reads it.
One model writes the answer. A second one compares it with your configuration and the passages it was drawn from: every price, fee, hour, phone number, link, service area, warranty, and policy. The fixed rule for this agent is plain: any fact about hours, prices, fees, coverage, or policy has to come from your own documents or settings. Deterministic checks run alongside the model and don't depend on its judgment.
Only your links
A link that isn't in your configuration, your documents, or the visitor's own message blocks the draft. The widget can't send a visitor somewhere you never listed.
Objections need quotes
The reviewing model has to quote the wrong words in the draft and the passage of yours they contradict. An objection it can't back with both quotes doesn't block the reply.
No advertising
Offer language in a reply is caught by a fixed list of phrases, not by asking the model's opinion. The answer stays an answer.
Never first
It replies only to someone who opened the conversation. There's no setting that makes it start one.
A rejected draft gets one rewrite. If the rewrite is rejected too, the visitor gets a plain holding line instead, thanking them and saying a team member will follow up, and I get the alert rather than your Slack, because a blocked answer usually means something in the setup needs fixing. The agent is also instructed never to ask for card numbers, passwords, or ID numbers, and to treat anything a visitor types as information, not as instructions to follow.
The greeting's disclosure isn't decoration. Anthropic's usage policy requires a consumer-facing chatbot to tell people they're talking to AI at the start of each session. The widget's first line introduces it as the virtual assistant; if you'd rather it say 'AI assistant' in so many words, that's a one-line setting. State rules on bots and on collecting contact details vary, so check your own situation with your counsel.
Your website first. Text messages if you want them.
The widget is the product. If you also want the same agent answering texts, the console can point a phone number's incoming messages at it. Tell me you want this and I set its replies to text style: short, with only your booking link.
WhatsApp and Facebook Messenger are a different story. The agent has a reply style for both, but neither is connected as a channel today. If one of those matters more to you than your website, raise it when we talk and I'll tell you plainly where it stands.
$249 a month with $50 of usage in it.
$249 a month, which includes $50 a month of AI usage; heavier chat volume is billed beyond it. The $850 setup is billed once, before anything is built, and pays for the site training, the rehearsal, and the watched first weeks. The terms, as printed:
- Month to month, cancel any time
- 30-day cancel for any reason
- No free trial
- If you add text messaging, the number ports out with you
- Conversation records kept 90 days
Running several of these? The 7 paid agents are $1,200 a month as one bundle. How the included usage works, what published alternatives charge, and the break-even arithmetic are on the pricing page.
Count last week's repeat questions first.
Go through last week's email and texts and count the messages that asked something your site already answers. Say it's fifteen, at four minutes each. That's an hour a week, a bit over four hours a month, and some of those people didn't wait for your reply. Now put a value on one job you win because the answer arrived while they were still on your page. If that job is worth more than $249 to you, one a month covers the fee. If you can't name one, don't buy it.
A fit if…
- Your inbox and phone get the same handful of questions every week: price, area, hours, whether you do that job
- The answers already exist in writing, on your site or in a price list
- People read your site in the evening, when nobody is at the desk
- You'd rather get a summary and a phone number than read every chat
Skip it if…
- Your site gets a few visitors a day. There aren't enough questions to justify it.
- Your answers aren't written down anywhere. It can only repeat what you've given it, so write the FAQ first.
- Most questions are about a specific order or account. It can't see those; it takes details and opens a ticket.
- You want to talk to visitors yourself in real time. Buy live chat with an inbox.
No testimonials here. The proof is what you watch yourself: the rehearsal, and the review on every reply.
The website is only one door. Calls that ring out are another, and missed-call text back covers that one. For the whole path an inquiry takes, and where a human belongs in it, see how to automate lead follow-up.