AI Drafts, Humans Decide: The Only Safe Way to Put AI on Your Website
The fear is specific and it is fair. A guesthouse owner pictures an assistant on her website confirming a room that is already taken. A dentist pictures one telling a patient that a procedure is painless. A plumber pictures one quoting a price he would never have given for a job he has not seen.
So some owners ban the whole idea, and others let a chat widget loose on their website and hope. Both are the wrong answer to a sensible question. The right answer is about where the AI sits: let it draft and let it answer from what you have published, and make it impossible for it to commit you to anything a person has not decided.
A tool and a workflow are different things
A chat window on its own gives you generation with no rules around it. Whatever comes out can go anywhere, and the only protection is that somebody notices in time, which on a busy Friday is no protection at all.
A workflow puts the generation inside limits. What the AI is allowed to draw on is controlled. What it can say on its own is decided in advance. What a person has to approve before a customer sees it is written down. The limits are worth going through one at a time, because together they are the answer to the fear.
Limit one: it answers from what you published
An assistant that answers from your own pages is a very different animal from one that answers from the internet. Asked what a check-up costs, it should quote the figure on your price list. Asked whether you are open on a public holiday, it should read the hours you wrote.
Asked something your site does not cover, it should say so and hand over instead of filling the gap. A short “I do not have that, but the practice can tell you, and here is how to reach them” costs you nothing. A confident invention costs you a complaint.
Limit two: it never commits you
Draw the line around anything that binds the business. A discount. A promise about a medical outcome or a legal position. A date you may not be able to keep. A price for work nobody has looked at.
For standard work the assistant can quote the published price, because you published it and you meant it. For anything bespoke it should refuse politely and route the person to a human. Refusing to guess at a number is not a weakness in an assistant. It is the single most valuable thing it does.
Limit three: the customer knows what they are talking to
Tell people. An assistant with a human first name should say in its first message that it is an AI, before anything else happens, because a customer who works it out later feels tricked and a customer who was told at the start does not mind at all.
This is not a legal nicety. It changes how people use it. Somebody who knows they are talking to software asks clean, factual questions and takes the answer at the right weight.
Limit four: nothing goes out in your name unread
The same rule covers AI-written page copy and blog posts. Read it before it goes up, particularly anything that touches somebody’s health, money or legal position, and particularly anything containing a number.
Generated copy has a habit of producing claims that sound like facts. A sentence about recovery times or success rates can appear in a draft without any source behind it, and once it is on your site it is your claim, written in your voice, on your letterhead.
The two objections you will hear
From the cautious member of staff: checking it takes as long as writing it myself. That is true of a raw chat window, where the output arrives as a wall of text to be fact-checked from scratch. It is not true inside limits. When the assistant can only draw on pages you wrote, review becomes a spot check of the handful of things that could have gone wrong, not a line-by-line hunt for inventions.
From whoever is keenest: the limits slow us down. Yes, by minutes. A price that gets looked at takes a look it would not otherwise get. Those minutes are the entire product. The alternative is speed right up until the week you spend apologising to a customer, after which the whole thing gets switched off and you are back to answering the phone yourself.
Say in public how you use it
Write your position down where a customer can read it. Which tasks the AI does. What it is never allowed to decide. What happens to what a visitor types into it. Who they can speak to if they would rather speak to a person.
Customers ask about this more than they used to, and a short honest page turns an awkward question into a reason to trust you. Amysa publishes its own AI position for the same reason, including which models sit behind the assistant and what happens to the information a visitor types in.
The takeaway
Write your own limits down before you switch anything on. You should be able to answer four questions about any AI on your website: what can it say without a person, can it commit the business to a price or a promise, does the customer know it is not human, and what does it do when it does not know. Any “not sure” is where the trouble will come from.
Amysa builds to those answers not around them: the assistant says it is an AI, answers from what the business has published, quotes published prices, and hands a question it cannot answer to a person. Ask anyone selling you AI to be that specific.
KEEP READING
Related field notes
Tell us what still happens on the phone.
We will tell you what the site could take off your hands and what that costs. The conversation is free.