11:40pm
Website chatbot for a small business: the question that arrives at 11:40pm
A homeowner asked a heating company's website one question at twenty to midnight, and the answer came at nine. What a website chatbot for a small business, or a real estate office, really does in that gap. Why the number this category is sold on cannot be sourced. Who owns what it says, whether it is worth it for your kind of business, and what to do first.
We help local business owners decide whether a website chatbot belongs on their site, and set up an AI chatbot that is built to answer from the owner's own information, offer times from the real calendar and hand anything it cannot check to a person. Below is how to test one, what it does and what it will not do, whether it is worth it for your kind of business, and what to do first this week.
This page is for owners of a home services company, a dental, accounting or insurance office, a salon, spa or gym, or a real estate office: any business where a visitor lands on the website with a question and nobody is there to answer it.
An invented example shows the gap. At twenty to midnight on a January night, a homeowner in Fishkill is on her phone with a furnace that has started to bang. She finds the website of a heating and cooling company, and she has one question: do they come to Fishkill, and could someone look at it tomorrow.
There was a contact form. She filled it in, or more likely she did not, and she went to bed under two blankets.
In the invented example, the owner saw the form at 8am and called at 9. By then she had asked the same question on another company's site, whose chat answered in under a minute and offered a visit that morning. The first company never learned which call it lost.
That gap, between when people ask and when a business answers, is what this page is about.
In short
- Being first buys the conversation. In a 2011 study, trying within the hour was nearly seven times likelier to reach a decision maker than waiting one more hour. Of 2,241 companies sent a test inquiry, 23 percent never responded.
- A good assistant answers at any hour from what you approved, reads the live calendar, moves to text, and writes what it learns to the CRM. It is worth it only where questions arrive when nobody can answer and a person reads the transcripts.
- It does not close, and it should say so rather than invent an answer it cannot verify: a company owns what its chatbot says.
The speed claim in AI chatbot sales pitches, and where it comes from
One figure about reply speed circulates with AI chatbot pitches, and an earlier version of this very page rested on it: that seventy eight percent of customers buy from whoever answers first.
We went looking for the study behind it, because a claim doing that much work should have a document under it. We found no source for it: no report, no stated sample and no method that we could open. None of that proves the figure is wrong. It does mean we cannot tell you whether it is right, and a number nobody can check is a slogan wearing a percentage sign.
So this article does not use it. What follows rests on a different piece of work, one where the sample size, the companies and the definitions are all published and can be argued with. It is from 2011 and it stays here because we found no newer study that measures the same thing without selling software. It carries the argument, and this paragraph says its year.
The evidence
Relative odds of qualifying a lead, by how fast you respond.
Odds are indexed to the after-24-hours case. Source: Harvard Business Review, The Short Life of Online Sales Leads (2011), by James Oldroyd, Kristina McElheran and David Elkington, reporting their separate study of 1.25 million sales leads received by 29 B2C and 13 B2B companies in the US.
Cross-industry lead response research, not a real estate study. The paper publishes two ratios rather than three: contact within an hour was "nearly seven times" as likely to qualify a lead as contact an hour later, and "more than 60 times" as likely as waiting a day or more. The middle bar is those two divided, not a third figure they measured. The pattern is what carries over.
According to the same Harvard Business Review article, an audit of 2,241 US companies, each sent a web-generated test lead, found that 23 percent never responded at all, and among those that responded within 30 days the average response took 42 hours. Its separate study of 1.25 million sales leads found that firms that tried within an hour were nearly seven times as likely to qualify the lead as those that tried even an hour later. Look at what the chart is counting, because it is not sales. It counts whether the firm ever got into a real exchange with somebody who could decide anything. That is a much lower bar than closing, and a far higher one than a reply landing in an inbox. And it is the bar a contact form that nobody reads until morning cannot meet at twenty to midnight.
That is a smaller claim than the slide makes, and a more useful one. It does not say the fastest company is the best company, and it does not say speed closes work. It says that in that sample the firms who tried first were far likelier to reach the conversation, and the conversation is where a furnace repair, a cleaning, a first appointment or a house is sold. Everything you are good at happens after somebody picks up.
So the question is not whether you are a better heating contractor than the one who, in the invented example, answered a minute after she asked. You probably are. The question is whether you were in the room.
The gap
They asked
11:40 pm
9 hours 20 minutes
You called
9:00 am
Nine hours and twenty minutes is not a slow reply. It is a conversation that happened without you.
The sum below is arithmetic, not a forecast. An invented example: an HVAC company gets 40 website inquiries a month and replies within a few hours, which the calculator treats as handing 40 percent of them to whoever answered first: 40 x 40% = 16 a month, and 16 x 12 = 192 conversations a year that began somewhere else. It stops there on purpose, because what a conversation is worth differs in every trade.
In your numbers
What is answering late costing you?
Form fills, chat starts, do-you-cover-my-town and is-this-still-available texts.
How fast do you reply, honestly?
Conversations a year that began somewhere else
192a year
- Inquiries through your siteyour 40
- 40a month
- Handed to whoever answered first40% of them
- 16a month
- Conversations that began somewhere else12 months
- 192a year
The arithmetic is on screen on purpose, because the easy version of this calculator is a lie. The reply curve is a judgement and not a measurement: an immediate reply is treated as winning roughly nine conversations in ten, a few hours about half, the next day about a fifth. It is shaped by the research quoted above rather than derived from it, and deliberately not by the unsourced 78% figure this article declines to use. It stops at conversations a year, because what one is worth differs in every trade, from a first dental visit to a furnace repair to a house, and a calculator that multiplies every missed inquiry by a job or a commission is selling you something. The two numbers you can drag are yours. Both are a place to start arguing from, not a measurement of your business.
What does an AI chatbot actually do on a small business website?
Strip away the marketing and an AI chatbot is a text conversation on your own website, answered by a language model, and connected to the systems that make the answer true. For a small business that means your approved answers, your calendar and a place to send the conversation. For a real estate office it can also mean the live listings.
The version we run does four things:
What it actually does
Four moves.
It answers immediately, at any hour.
Not a canned reply. An actual answer to the actual question, from the answers you approved.
It checks what is live.
Ask it whether a technician can come tomorrow and it reads the calendar and offers the times that are open right now, not the ones that were open when somebody last edited a page. For a real estate office it queries the live listings the same way.
It moves to a channel people open.
The confirmation goes out by text, to the number the visitor gave. People read texts. People do not read the fourth email from a business they have never used.
It captures what it learns.
Name, number, what the job is, where and when, and the transcript, all written to the CRM so your callback starts from what they said rather than from a blank record.
The last one is the one people underrate. A lead that arrives with a transcript attached is a completely different object from a lead that arrives as a name and an email address. Your callback starts from what they said.

What makes an AI chatbot answer true?
An AI chatbot's answer about something that changes, such as a time or a price, is true only when it comes from your live systems, and getting that right is the whole job. You will hear that an assistant is trained on your data. Ask what that sentence means, because it can describe two quite different things.
The first meaning is that the assistant was shown your website. Somebody pointed a crawler at your pages, and the text got chopped up and stored. When a visitor asks a question, the system finds the paragraphs that look most similar and asks a language model to write an answer out of them. That works for what your website already says. It is a better search box, and a better search box is genuinely worth having.
It also cannot tell you whether Thursday at 4:00pm is still open. That is because your website is a copy of the truth taken at some point in the past, and the question is about right now.
The second meaning is that the assistant is connected to the systems that hold the live answer. It calls them while the person is waiting. Asked about availability it reads the calendar and offers what the calendar shows when it asks. Asked about a price you have not published it says it does not have one, and books the visit. The difference between the two is not model quality, and it is not prompt writing. It is whether anything on the other end is actually live.
That difference matters in any trade where the facts change faster than the web page does. In an invented week, a technician calls in sick on a Tuesday, a salon chair opens on Friday, and a listing goes to contingent on a Tuesday afternoon. Every one of those events can leave an assistant that only read your website confidently wrong. And confidently wrong at eleven at night with nobody watching is exactly the failure that costs you a customer rather than a lead.
There is a second half to it, and it is the harder half. It is what happens when the system does not have the answer. The fix is not a cleverer model. It is a rule that when a lookup returns nothing, the assistant says it does not know and offers a person instead, rather than writing something that only sounds right. That behaviour has to be built and it has to be tested. When you are testing one, ask it something it cannot possibly know and watch what it does. That answer tells you more than an hour of demo.
What should you check before an AI chatbot goes on your website?
Four things are worth checking before an AI chatbot goes on your website, because they are yours to live with afterwards. None of this is legal advice.
What it says is what your business said. In 2024 a Canadian tribunal decided a case against an airline whose website chatbot had given a customer wrong information about bereavement fares. The airline argued the chatbot was a separate entity responsible for its own actions. According to the law firm McCarthy Tétrault's summary, the tribunal rejected that, called it "a remarkable submission", and held that a chatbot is still just part of the website and the company is responsible for all the information on it, whether it comes from a static page or a chatbot. It is a Canadian tribunal's decision, not New York law. It is still the clearest picture of what a customer, and a judge, will assume: the answer on your site is your answer. That is why the assistant answers only from what you approved, and why a price, a diagnosis or a promise goes to a person.
Do not let it pretend to be a person, and understand why that is a rule and not a preference. California has had a bot disclosure law since the first of July, 2019, which is the date the chapter gives itself. Under Business and Professions Code section 17941, it is unlawful "to use a bot to communicate or interact with another person in California online, with the intent to mislead the other person about its artificial identity for the purpose of knowingly deceiving the person about the content of the communication in order to incentivize a purchase or sale of goods or services in a commercial transaction". The escape hatch is one sentence long and it is free: "A person using a bot shall not be liable under this section if the person discloses that it is a bot." The disclosure has to be "clear, conspicuous, and reasonably designed to inform".
Whether that reaches a business in the Hudson Valley whose visitor is in Los Angeles is a question for an actual lawyer, and we are not one. What is not in doubt is that the compliant behaviour and the decent behaviour are the same behaviour, and it costs one sentence.
It is going to cost you page speed, and page speed is a number Google publishes. A chat widget is third party JavaScript that loads on your pages. Google's own Core Web Vitals thresholds say that "to provide a good user experience, pages should have a INP of 200 milliseconds or less" and that largest contentful paint "should occur within 2.5 seconds of when the page first starts loading". A heavy widget eats into both, and phones are the place to measure it.
Somebody has to be able to use it with a keyboard. A chat window is a part of your website like any other, and a visitor who cannot use a mouse still needs to get in and out of it. The test is two minutes and needs no vendor, as the third check below shows.
None of the four needs a meeting or a budget to look into. Here is how to check three of them yourself, on whatever is answering on your site tonight.
Three you can run tonight
All three take an afternoon, and none of them needs us.
Open your own chat window and read the first thing it says.
The test is not whether a disclosure exists somewhere in your terms. It is whether a stranger on a phone would know, inside one line, that they are talking to software. If you have to go looking for it, it is not clear and conspicuous, and the fix is one sentence you can write this evening.
Time your busiest page twice, and keep both numbers.
Once before the widget goes on and once after, on a phone rather than on your desktop, because that is where your traffic is. The thresholds are published and the test costs nothing. If the vendor cannot tell you what their script weighs, that is your answer, and the good ones load nothing at all until somebody taps the bubble.
Tab into the chat window, then try to tab back out of it.
Keyboard only, no mouse, on the page a buyer would actually land on. If focus goes in and will not come out, a visitor using a screen reader is stuck on your website with no way forward. That is not a slow page. That is somebody who cannot leave your front door.
What does it take to run an AI chatbot?
What an AI chatbot takes from the owner is mostly three things, and only one of them is software. We print no price on this page.
The first is what it is allowed to look at. An assistant answering out of your approved answers is the smaller job. An assistant that reads a live calendar, or live listings for a real estate office, has to be connected to the system that holds them. That is a data project with its own rules, its own refresh behaviour and its own failure modes. The second is what happens after the conversation ends. A transcript that lands in an inbox is worth a fraction of a transcript that lands on the right contact record in the CRM you already use, with the appointment on the calendar. That plumbing is often the larger half of the build. It is the half people forget to ask about, because it is invisible in a demo.
The third is the reading. Somebody has to go through what people actually asked, once a week at first, because the questions your visitors ask show what they want to know, and because that is how a wrong answer gets found.
On timing, the build is not the part that needs the most thought. Writing the page of answers and settling what the assistant may and may not say takes the thought. And the first month tells you more than the build does, because it shows what people really ask at eleven at night, which nobody could tell you in advance.
What an owner provides is short: the answers it may give, the calendar with the times it may offer, a place for the conversations to land, and about twenty minutes a week to read them. We suggest twenty, in your own name, for as long as it runs. If nobody will do that, leave it switched off.
What a person still approves: the page of what it may answer and every change to it, any quote for a particular job, any discount, any promise, and the first reply to every lead it hands over.
How do you test an AI chatbot before you buy it?
Five questions test an AI chatbot before you buy it, and you can ask all of them inside a demo window without anybody's permission. Take them to any vendor, including us.
Ask it something it cannot know. Make up a street and ask what a repair on it would cost. What you want back is an admission that it has no way to confirm that, and an offer of a person. If it produces a number, then you have just watched it invent one.
Ask it something that changed this week. A new opening hour, a closed day, a status change, anything with a date on it. This is the question that separates an assistant reading a copy of your website from one reading the live calendar.
Ask it something a competitor's customer would ask. Then look at where the answer came from, and whether you are comfortable with that being said in your name to somebody you have never met.
Try to break the handoff. Say you want to speak to a person. Say it rudely. The assistant's job at that moment is to stop selling and get out of the way. An assistant that argues has failed the test.
Then close the laptop and open your phone. Load the page on mobile data rather than the office wifi, and count the seconds before you can type. A demo may run on a desktop over a fast connection, so test on yours.
Or have us run it
Send us the address of your site and we will put all five of those questions to whatever is answering on it now, then send you back what it actually said.
No charge, and nobody calls you unless you ask us to.
Is an AI chatbot worth it for a small business?
An AI chatbot is worth it when questions arrive often enough that a missed one is noticed within a week, and when a person will read what it says. These are the screening questions of our AI assessment, our own method rather than a published standard, and three of them keep the idea or cross it off:
- Does the job happen often enough that a failure shows within a week? Visitors ask something on most days, and a slow reply costs you a visit you can name.
- Can the rule be written down for a new hire? Your hours, your service area, what a first visit involves and what you never quote over the phone fit on one page.
- Does a wrong answer land where a person sees it first? For a chat that means the risky answers never reach the customer alone: a price, a diagnosis or a promise goes to a person, and someone reads the week's transcripts so the rest is caught.
Then the stop question, which asks what breaks if the job stopped: if nothing breaks, stop doing it and do not automate it. Then check whether a tool you already pay for does it, since a website builder or a booking tool you already have may include a chat window you have not switched on.
Score the job's impact from 1 to 5. At 1 or 2, leave it alone: a clear contact form and a promise of when you will reply, kept every time, does the job. The chat reaches a customer, so it carries risk, and a person approves what it may say before it goes live and reads the conversations after.
An invented example: a salon gets 6 website questions a week, and an owner who answers each in 4 minutes spends 6 x 4 = 24 minutes a week (the owner's estimate). That is a leave-it-alone answer. A heating and cooling company with 40 inquiries a month and most of them arriving after six in the evening is a different job, and the answer is to run the test above on it.
The three questions are run on your own last thirty website messages before anyone looks at a tool.
What should an AI chatbot never do?
An AI chatbot has four limits, and here is the plain version.
It does not close. It does not read a room, it does not know when a customer is stretching the truth about their timeline, and it has no instinct for what is really going on behind a question. Those are the reasons you have a business.
It does not know things it cannot verify. If a visitor asks whether a repair will cost a certain amount, or whether the seller of a house will take a contingency, then the honest answer is that nobody knows until somebody looks or asks. The assistant should say so and book a call or a visit. An AI that invents an answer to look competent is worse than no AI at all, because you will find out about the invention from a furious customer.
It does not replace the person at the front desk. It takes the repeatable first questions off that desk and hands over the conversations that need a voice, and the person stays.
And it does not pretend to be a person. Ours says in its first line that it is an AI assistant.
The measure of an AI assistant is not how human it sounds. It is whether the answer was correct, whether it was immediate, and whether a real person showed up when it mattered.
The teardown
Watch it handle the 11:40pm question.
The pattern, not a transcript from a real client.
What the visitor saw
Do you come to Fishkill? My furnace is banging and I would rather not wait.
Visitor / 11:40 pm
Yes, Fishkill is inside our service area. I cannot tell what the noise is from here, and I will not guess at a cause or a price. A technician can look at it. The first open visits are tomorrow at 9:00am or 1:00pm. Would one of those work?
Assistant / 11:40 pm
9 works.
Visitor / 11:41 pm
Booked for 9:00am. What name and number should the technician use? I will text you the confirmation.
Assistant / 11:41 pm
Sam, 845-555-0134.
Visitor / 11:42 pm
Thank you, Sam. The confirmation is on its way by text, and the technician will have this conversation before the visit.
Assistant / 11:42 pm
What happened behind it
- 11:40 pm
Answered
No wait, no form, no canned reply.
- 11:40 pm
Declined to guess
No diagnosis and no price: a person looks at it.
- 11:41 pm
Read the live calendar
9:00am and 1:00pm were open, so those were offered, and 9:00am is now held.
- 11:42 pm
Sent by text
The confirmation, to the number the visitor gave.
- 11:42 pm
Written to the CRM
Name, number, the job, the town, and the transcript.
- 8:30 am
A person starts from it
The technician reads the conversation before the 9:00 visit.
Which businesses an AI chatbot fits, and which it does not
An AI chatbot fits a business where questions arrive when nobody can answer them: evenings, weekends, or a front desk that is always with somebody. That is a home services company, a dental, accounting or insurance office, a salon, spa or gym, and a real estate office with listings to search.
Three kinds of business do something else. A business that gets a handful of website visits a week needs a clear contact form and a kept promise, not a chat window. A business where a person already answers every inquiry within minutes has no gap to close. And a business where nobody will read the weekly transcripts should leave it switched off, because a wrong answer nobody reads is a wrong answer that can be given again.
An invented example: a one-person piano tuner with forty visits a year keeps one clear page with a phone number and a line saying he calls back the same evening.
Most businesses fall into one of these groups, and the group decides the next step.
Who this is for
A good fit, a poor fit, and what to do instead
A good fit
Questions that arrive when nobody can answer: a home services company, a dental, accounting or insurance office, a salon, spa or gym, a real estate office.
A poor fit
A handful of website visits a week, a person who already answers every inquiry within minutes, or nobody to read the transcripts.
Instead, for a poor fit
One clear page with a phone number and a promise of when you reply, kept every time.
Either way
A person approves what it may say, and a person reads what it said.
What to do first with an AI chatbot
The first step with an AI chatbot is one job, taken a week at a time, ending with a decision written down.
- Week 1: read the last thirty messages that came through your website and mark the time each arrived and what was asked. Write the page of answers you would give a new hire: hours, service area, what a first visit involves, what you never quote.
- Week 2: put the five questions from "How do you test an AI chatbot before you buy it?" to whatever answers on your site now, or to a vendor's demo, and write down what each said.
- Weeks 3 and 4: run the test in "Is an AI chatbot worth it for a small business?". If it passes, approve the page of answers and the first line that says it is an assistant, and read every conversation for a month. If it fails, keep the contact form and write that decision down.
This is also how our AI readiness assessment starts. In its interview we ask for one job written in two minutes under input, do, output and your check, then handed to a colleague who does not know it, noting where they stall. Each job is then scored for impact and risk from 1 to 5, with four checks: it binds the business, it reaches a customer, it feeds another system, it can damage trust. The owner leaves with each job marked to automate, to keep a person on, or to leave alone.
An invented example: a heating company's owner writes the answer to "do you come to Fishkill" in two minutes. Her dispatcher stalls at "do we cover Wappingers after six", and that answer becomes the first line of the page.
A suggested schedule
One job, a week at a time, a decision written down
Week 1
Read the last thirty website messages; write the page of answers a new hire would need.
Week 2
Put the five test questions to whatever answers on your site now, or to a demo.
Weeks 3 and 4
The worth-it test; then approve the answers page and read every conversation, or keep the contact form and write the decision down.
A schedule to start from, not a deadline.
Common questions about AI chatbots, answered honestly
What is an AI website chatbot for a small business?
An AI website chatbot for a small business is a text conversation on your own site, answered by a language model. The useful kind is connected to the systems that hold the real answers, such as your calendar, and the weaker kind only reads your website's text. A visitor types a question the way they would ask you. It answers in sentences rather than menus, and when it reaches the end of what it can confirm, it is built to book time with you instead. The part that makes it worth having is not that it talks. It is that it is awake at eleven at night and, when it is connected, it answers from something current.
Do I need an AI chatbot on my website?
You need one when questions reach your website at hours when nobody can answer them, and not before. If the website gets a few visits a week, or a person already answers every inquiry within minutes, a clear contact form and a promise of when you reply, kept every time, does the job. The test in "Is an AI chatbot worth it for a small business?" decides it with your own numbers. According to the Census Bureau's Business Trends and Outlook Survey, less than 20 percent of firms with four or fewer employees reported using AI, so choosing not to is a normal answer.
How much does an AI chatbot cost for a small business?
We print no price on this page. Hosted chat tools publish their own, so compare what each includes: whether it reads a live calendar or only your website text, where the conversations go, and who reads them each week. What it takes from the owner is set out in "What does it take to run an AI chatbot?".
What is the difference between live chat and an AI chat assistant?
Live chat puts a person behind the window, so it answers only while that person is working, and an AI chat assistant answers at any hour from what you approved. Many offices want both: the assistant for the first questions at night, and a person for what it hands over. Live chat for a real estate website follows the same rule, a person when someone is there and an assistant when no one is.
Can an AI chatbot work on a real estate website?
An AI chatbot on a real estate website can answer questions about the office and search the live listings during the conversation, so the homes it quotes come from the feed rather than from a page written last month. An assistant that reads only the website's text can tell a visitor only what the site already said. Anything about a seller's terms, such as whether the owner will take a contingent offer, goes to the agent, because nobody knows until the agent asks.
Does it work on the MLS, or only on what my website already says?
Both exist and they are not the same purchase, and the MLS only matters to a real estate office. An assistant reading your website can only tell somebody what your website already told them, which is a better search box. An assistant connected to the listings feed reads status, price and availability from the feed rather than from a page written last month, and it is only as current as the feed. Ask any vendor which one they are selling, and ask them what happens the afternoon a listing goes contingent.
How is it different from the chatbot I already have?
If yours is a decision tree with four buttons, you have a menu. The difference is whether it can answer a question nobody scripted in advance, and whether the answer comes from live data or from a page written eight months ago. Ask yours something specific about a job or a listing. If it offers you a phone number and a contact form, you have a form with a friendlier shape.
Will it annoy my visitors?
A bad one will. A pop-up that demands a phone number before it has been useful is an advert in a chat widget's clothes. The one worth running answers a question first and asks for a number once it has earned it. Look at your own bounce rate before and after rather than guess.
My customers want a human, not a bot
They do, eventually. What they want late at night is to know whether you cover their town and whether someone can come tomorrow. An assistant that answers that and then books a visit with you is not standing between you and the customer. Software that helps a customer and software that traps one with no way through to a person are different products.
Do I have to tell people it is an AI?
Say yes and stop thinking about it. There are statutes in this area, and at least one of them, in California, makes disclosure the thing that removes liability. Beyond the law it is simply the right call: an assistant that introduces itself costs you nothing, and being caught pretending costs you the only thing you actually sell.
What happens when it gets something wrong?
Somebody has to be reading, and the business owns the answer either way, as the airline case above shows. Set it up so every conversation is stored and so you or somebody on your team reads the week's transcripts, at least at first. A wrong answer that nobody reads can be given again. And the person who finds it for you will be a customer.
Can I add an AI chat assistant to my existing website?
Yes. It is a small piece of code that goes into the site the business already has, whether it came from a website builder or was built by hand, so the site does not need rebuilding as long as it lets a piece of code be added. It connects to the calendar and the CRM the business already runs when they are systems it supports.
Can we talk about a website chatbot on a call if my business is outside New York?
We are based in New York, and the call works for an owner anywhere. The chatbot goes into the website the business already has, as the answer above says, so the work needs no visit.
Ten questions
What did you learn? Check it in two minutes
Ten questions on what this page taught. It takes about two minutes.
Take the quiz
1 of 10
In the Harvard Business Review audit of 2,241 US companies sent a web test lead, what share never responded?
- A.37 percent
- B.23 percent
- C.16 percent
- D.42 percent
Show the answer
B. 23 percent The audit found that 23 percent of the companies never responded at all; see "The speed claim in AI chatbot sales pitches, and where it comes from".
2 of 10
What does the 2011 response-speed chart count?
- A.Closed sales
- B.Replies landing in an inbox
- C.Whether the firm got into a real exchange with somebody who could decide
- D.Visits to the website
Show the answer
C. Whether the firm got into a real exchange with somebody who could decide It counts the conversation, a lower bar than closing and a higher one than an inbox reply; see "The speed claim in AI chatbot sales pitches, and where it comes from".
3 of 10
A visitor asks something the assistant cannot know. What is the right behaviour?
- A.Give its best guess, labelled as a guess
- B.Quote a typical price
- C.Repeat the question back
- D.Say it cannot confirm that, and offer a person
Show the answer
D. Say it cannot confirm that, and offer a person When a lookup returns nothing, the assistant says it does not know and offers a person instead; see "What makes an AI chatbot answer true?".
4 of 10
In the 2024 airline chatbot decision, who was responsible for what the chatbot said?
- A.The chatbot, as a separate entity
- B.The software vendor
- C.Nobody, because it was automated
- D.The company, because the chatbot is part of its website
Show the answer
D. The company, because the chatbot is part of its website The tribunal held that a chatbot is still part of the website and the company answers for all the information on it; see "What should you check before an AI chatbot goes on your website?".
5 of 10
Under California's bot disclosure law, what keeps a person from liability?
- A.Disclosing that it is a bot, clearly and conspicuously
- B.A line in the terms of service
- C.Giving the bot a first name
- D.Asking the visitor to agree first
Show the answer
A. Disclosing that it is a bot, clearly and conspicuously A person using a bot is not liable if they disclose that it is a bot, in a way that is clear and conspicuous; see "What should you check before an AI chatbot goes on your website?".
6 of 10
What does the third question of the worth-it test ask?
- A.Whether the software is cheap
- B.Whether a wrong answer lands where a person sees it first
- C.Whether the vendor is well known
- D.Whether competitors have one
Show the answer
B. Whether a wrong answer lands where a person sees it first Every conversation is stored and someone reads the week's transcripts; see "Is an AI chatbot worth it for a small business?".
7 of 10
A salon gets 6 website questions a week and answers each in 4 minutes. What is the verdict?
- A.Automate it at once
- B.Leave it alone: a clear contact form and a kept promise do the job
- C.Switch on a pop-up
- D.Buy two tools
Show the answer
B. Leave it alone: a clear contact form and a kept promise do the job 6 x 4 = 24 minutes a week is a leave-it-alone answer; see "Is an AI chatbot worth it for a small business?".
8 of 10
How much time a week do we suggest for reading the conversations?
- A.Twenty minutes
- B.None
- C.Two hours
- D.Five minutes
Show the answer
A. Twenty minutes We suggest twenty minutes a week, in your own name, for as long as it runs; see "What does it take to run an AI chatbot?".
9 of 10
What happens in week 1 of the roadmap?
- A.A widget is installed
- B.A domain is bought
- C.The five test questions go to a vendor
- D.The last thirty website messages are read and the page of answers is written
Show the answer
D. The last thirty website messages are read and the page of answers is written Week 1 reads the messages and writes the page of answers a new hire would need; see "What to do first with an AI chatbot".
10 of 10
Which is a real reason to leave the assistant switched off?
- A.Nobody will read the weekly transcripts
- B.The website has a footer
- C.The owner prefers the phone
- D.The widget has the wrong colour
Show the answer
A. Nobody will read the weekly transcripts A wrong answer nobody reads is a wrong answer that can be given again; see "Which businesses an AI chatbot fits, and which it does not".
Where does an AI chatbot go wrong?
An AI chatbot fails in three ways, each avoidable:
Three ways it fails
All avoidable, all common.
It is not connected to anything.
An assistant with no calendar and no live data can only talk in generalities, and a visitor works that out in about one question.
The handoff is a dead end.
It qualifies a hot lead beautifully and then drops them into an inbox nobody watches. The value was never the captured email. It was the booked visit.
Nobody reads the transcripts.
They are the most useful sales material you will ever own, because they are a record of the exact questions your customers are asking, in their own words. Most people never open them.
The system
What it is connected to.
An assistant is only as good as what it can reach. This is every hop between the question and the booked call.
- The question: Your own website, any hour
- The assistant: A language model, given limits
- Live facts: Your calendar and answers
- The reply: Sent by text
- The record: CRM, with the transcript
- The handoff: A booked visit, a person
What to do next about an AI chatbot
Our own AI chatbot is built this way, and the home page shows where it sits: asked something hard, it either answers or says it cannot and offers a call, and both of those are the correct behaviour. The build itself, and what happens to a lead in the minutes after the conversation ends, is written out on the AI chat assistant page.
The free half of this needs nobody, us included: your own site opened on a phone at night, the way the homeowner in the invented example did, with her question asked, the taps counted, and a note of whether it ever answers or just asks for your email address.
In the invented example at the top, the homeowner with the banging furnace was never coming back to check whether anyone replied. She was home at 9 when another company's technician arrived. The only real question is whose.
Whatever a visitor asks it, the answer you want is either a real one or an honest no.
A free website: some businesses get a plain website built free, by application. A plain site has no chat or services pages; an AI chat assistant can be added later if the business wants one. Apply through the form on the AI chatbot for small business page, and say in it that you are applying for the free website.
Sources
- Harvard Business Review, The Short Life of Online Sales Leads, by James Oldroyd, Kristina McElheran and David Elkington, on hbr.org, 2011. Read November 14, 2026; kept for the argument, because no newer independent study of response speed was found.
- McCarthy Tétrault, on Moffatt v. Air Canada, 2024 BCCRT 149, a Canadian tribunal decision, the firm's summary, February 19, 2024. Read November 14, 2026; kept for the argument, the case in this post about who owns a chatbot's answer.
- California Business and Professions Code sections 17941 and 17943, on bot disclosure, section 17941, current as of January 1, 2026. Read November 14, 2026.
- Google, Core Web Vitals thresholds, web.dev, a live document. Read November 14, 2026.
- US Census Bureau, AI use by businesses, from the Business Trends and Outlook Survey, December 2025 to May 2026, the story, May 26, 2026. Read November 14, 2026.
Sources checked November 14, 2026.
Written with an AI assistant and checked by a person.
Keep reading
All three leads look the same. Two are worth your morning. 
July 31, 2026
All three leads look the same. Two are worth your morning.
AI voice agent for small business: answering the calls an owner misses 
July 30, 2026
AI voice agent for small business: answering the calls an owner misses
The answer was wrong in March. It was still wrong in October. 
August 27, 2026
The answer was wrong in March. It was still wrong in October.
