The chat bubble in the bottom corner has become default furniture on business websites, the way autoplay music once was. The decision worth making is blunter than the trend: can the thing answering actually state your hours, your service area, your prices, and what your warranty excludes? Most cannot, and a bot that knows nothing is worse than no bot.
Three species of widget, two of them empty
Most chat widgets are one of three things. A button tree — canned choices leading to canned paragraphs, a phone menu wearing a text interface. A lead form in costume — whatever you type, the answer is "leave your email." Or a general-purpose AI dropped onto the page with no connection to the business behind it. That last one has read most of the public internet and none of your paperwork. It can discuss roofing in general while knowing nothing about your crew's schedule or the bulletin that changed your fastener spec last spring. Its fluency is exactly what makes it dangerous, because it will answer anyway.
What a bot that knows things is built from
The version worth building is grounded in your documents. The price book, the warranty terms, the service policies, the install specs — collected, cut into passages, and indexed so the system can find the right pages the moment a question arrives. The bot is then instructed to answer from those pages and to show which ones it used. Think of a fast-reading new hire who is required to walk to the filing cabinet before speaking. This carries a precondition the sales calls skip: the filing cabinet has to be right. If your documents are outdated or contradict each other, or half of them are bad scans, the bot inherits the mess with confidence added. Cleaning the document set is real work, and it is usually the first phase of the project rather than a footnote to it.
The bot's words are your words
In 2024, a Canadian tribunal decided a case worth knowing by name: Moffatt v. Air Canada. The airline's website chatbot told a customer he could apply for a bereavement fare after booking, while the airline's own policy page said otherwise. Air Canada argued the chatbot was responsible for its own statements. The tribunal rejected that outright and made the airline pay the fare difference — a bit over eight hundred Canadian dollars. The money was trivial; the principle was not. A business is bound by what its bot tells a customer. Which means the governing question is not what a bot can discuss but what it is allowed to discuss. Prices, warranty scope, anything with dollars attached — those answers either come from retrieved, current documents, or the bot declines and hands the person to a human.
The first deployment should face your own staff
A bot answering employees and a bot answering strangers carry different risk. Your office manager can smell a wrong answer about the install spec; a first-time visitor cannot. So the sequence that works runs inward first. If the front desk interrupts a senior tech several times a day to ask what is on page 40 of a manual, that interruption already has a cost you can feel — and it is the safest possible first job for a grounded system, because every wrong answer gets caught by someone who knows better. Point it at customers only after it has spent a season being corrected by people who read the originals.
When the honest answer is no
Some businesses fail the test cleanly, and it saves them money. If everything you would want the bot to know fits in twenty pages, you do not need retrieval infrastructure — you need those twenty pages written clearly and a good FAQ, at a fraction of the cost. If your actual leak is the phone — calls ringing out while every pair of hands is full — a website widget is aimed at the wrong door entirely. If nobody in the building will own the documents, the bot decays into confident wrongness on a schedule set by your next price change. And a chatbot that takes a question, then routes the customer's follow-up into an inbox nobody opens, has recreated the original problem with extra steps.
The question behind the question
What decides this is not the technology; it is the shape of your inbound traffic. Grounded bots earn their cost on repeat questions with authoritative written answers — the same twelve questions arriving hundreds of times, each answerable from a document you already maintain. They are weakest on questions the documents do not cover, because the failure mode on absent information is not silence but a plausible guess. A well-built system is configured to say "I don't know" and hand off, and the research on that behavior shows it can be reduced, not eliminated. So the bot gets scoped to what is written down, and a human stays one tap away for everything else.
Where to start
Run a two-week tally before talking to any vendor. Every question that comes in — by phone, by email, or at the counter — goes on a list, with a checkmark if a current written document answers it. Many checkmarks against repeating questions: you have a genuine candidate, and the checked documents are its reading list. Repeating questions with no checkmarks: you have a documentation project first. Either way, the tally costs nothing and settles the argument with your own numbers.