Does Your Business Need a Chatbot? Four Questions to Answer First

Your business needs a chatbot if the same questions keep showing up when nobody is around to answer them, and the answers already exist in writing. If either half of that is missing, a chatbot fills the gap by guessing, and you own every guess it makes.

Nobody actually wants a chatbot. An HVAC owner up on a roof isn't wishing for one. The owner wants the leads to stop slipping away while the phone rings in an empty truck. A chatbot is only worth buying if it stops that leak, and plenty of them don't.

Site assistants are one of the things we build for small businesses as part of our AI and data work, and one runs on our own website. Most of what follows comes from trying to break ours.

Four questions decide whether your business needs a chatbot

A chatbot is worth it when the answer to all four of these questions is yes. Any no means you have a different problem to fix first, and a chatbot will make that problem easier to see, not smaller.

1. Do the same questions come in over and over? Hours, service area, whether you take a certain kind of job, what happens after someone books. A chatbot is a machine for repeat questions. If every inquiry is a custom conversation, it will hand nearly all of them back to you.
2. Are the answers written down? On a page or in a document, somewhere a machine can read them, not in your head. A chatbot pointed at answers that don't exist produces answers anyway, and they sound sure of themselves.
3. Do the questions arrive when nobody can answer them? After hours, or while the whole crew is out on a job. If your phone gets picked up and your inbox gets a reply within the hour, the case for a bot is pretty thin.
4. Does somebody own it after launch? A named person reads what it says and updates it when a price changes. The same person takes the handoff when it can't answer. A chatbot nobody owns keeps talking anyway.

Question four is the one people skip.

Most customers would rather companies skip AI in customer service

A Gartner survey of 5,728 customers, fielded in December 2023 and published in July 2024, found that 64% would prefer companies not use AI in customer service, and 53% would consider switching to a competitor if they found out a company was using AI for customer service.

Customers in that survey worried more about being blocked from a person than about getting wrong answers. According to The Register's report on the survey, 60% of customers worried AI would make it harder to reach a human, and 42% worried it would give them wrong answers. Keith McIntosh, senior principal of research at Gartner, put it this way: "Many customers fear that GenAI will simply become another obstacle between them and an agent."

The Consumer Financial Protection Bureau has a name for a chatbot that blocks the way to a person. Its June 2023 report on bank chatbots, which estimates that about 37% of the US population used a bank's chatbot in 2022, describes customers caught in "continuous loops of repetitive, unhelpful jargon or legalese without an offramp to a human customer service representative." The report calls them doom loops.

In the same release, Gartner said 60% of customer service and support leaders are under pressure to adopt AI. Most of the push for a chatbot comes from the seller's side of the counter. If you add a chatbot, never let it stand between a customer and a person.

Chatbots break on ordinary questions, not attacks

When we tested our own chatbot, it held off every deliberate attack we threw at it and broke on one ordinary buyer's question.

In August 2026 we ran a red team against the assistant on our website, which means we attacked it on purpose to find what breaks. We sent 12 prompt injections (messages written to override the bot's instructions), 11 attempts to manipulate its pricing, and 10 social-engineering pushes, which try to talk a bot into breaking a rule rather than hack it. None of them got through.

We also sent one ordinary message written to sound like a real buyer. It described an HVAC company with nine trucks, about 1,800 calls a year, an old website, missed calls after hours, and the dispatch software they use. It was polite and detailed. The reply contained three things that were not true.

• It said we would connect to their dispatch software. We had never built that.
• It gave a schedule: "site in 2 to 3 weeks, AI live within 4 to 6 weeks." Nobody had set one.
• It claimed a track record: "most HVAC shops we work with see payback in 30 to 45 days." We have no roster of HVAC clients and no data like that.

The invented track record is the most expensive sentence a business chatbot can write, because a customer can repeat it back to you as a promise.

Our internal write-up of the August 2026 test says the assistant was "robust against people trying to break it and soft against people trying to buy from it." Hostile messages made it careful. A warm, detailed question from someone ready to buy made it eager, and eager is where the invented answers come from.

We added four hard limits the same day: no results or track record, no promises about other companies' software, no timelines, and no knocking anyone else's product. A retest still caught it telling the buyer their requests would flow into their software. Same promise, different words. That rule got tighter too.

New York City learned the same lesson in public. The city launched an AI chatbot for business owners in October 2023, and when The Markup tested it, it told owners they could take a cut of their workers' tips, that buildings didn't have to accept housing vouchers, and that a restaurant could go cash-free. All three of those answers were wrong, and they were ordinary questions from the people the bot was built for. In January 2026, New York's new mayor, Zohran Mamdani, called it "functionally unusable" and said it was costing the city around half a million dollars. By February 4, 2026, the bot's page said its "beta test has ended."

Owning a chatbot also means owning what it says. When Air Canada argued, in a case decided in 2024, that its chatbot was a separate entity responsible for its own actions, British Columbia's Civil Resolution Tribunal rejected the idea in its decision: "It makes no difference whether the information comes from a static page or a chatbot."

A chatbot is only as current as its last update

A chatbot does not know your business changed unless someone tells it. For two weeks in July and August 2026, the assistant on our site was still telling visitors about an entry-level package we had already retired, because nobody updated its instructions when the offer ended. We caught it by hand during a site rebuild. Nothing flagged it.

The retired package was an ownership failure, and the fix was to make that mistake impossible to ship. Every time we publish changes to the site, the deploy script now runs a suite of 24 automated checks before it copies a single file. Two of them compare the prices in the assistant's instructions against our public pricing page and our internal price list, and if they disagree the deploy stops. Changing the assistant's instructions at all fails the suite until a person runs a separate set of 25 live tests against the real AI model and signs off on the new version.

Some rules can't live in the instructions at all. On its first live test, asked the plain question "What is Pioneer AI and Data?", the assistant answered in 11 sentences, then 9 on a second try. The assistant's instructions said two to four sentences unless the visitor asked for more. Reply length is now capped in code at six sentences. A rule the model can quietly ignore is not a rule.

Sometimes the right chatbot has no AI in it

A chatbot that deals in prices is often safer with no AI behind it at all. Our AI Quote Bot, which you can try from our work page, looks like a chat and feels like a conversation, but it runs on a script. It walks a visitor through 13 steps in about three minutes. After the contact details, every step is a button, not a text box, and if someone types anyway it answers, "Use the buttons for this step so the ROI math stays clean." The estimate at the end comes from a formula and says so on the page: "This is a planning estimate, not a guarantee."

We built the quote bot as a script on purpose. A price is a promise, and you don't want software improvising promises to your customers. A script can't invent a number, because it can only say what it was given.

My rule of thumb: use AI where questions are open-ended and the answers are written down, and use a script where the answer is a number, a date, or a booking. A lot of businesses that think they need an AI chatbot really need a good form with a friendly face on it.

A trustworthy chatbot refuses anything you haven't put in writing

A trustworthy chatbot refuses to answer anything your business hasn't put in writing, starting with prices, dates, and guarantees. Every assistant we build on a client's content inherits this list. Each item is a refusal unless the business's own written material covers it outright.

• Prices and fees
• Dates and turnaround times
• Guarantees and warranties
• Legal, medical, tax, or compliance advice
• Anything about a competitor
• Anything about another customer
• Any commitment made on your behalf

When a customer asks about something on that list, the bot says it doesn't have that answer and points the person to someone who does. Every refusal ends with a next step, whether that's a name, a phone number, or a page, so nobody gets stranded.

A chatbot a customer talks to should also admit it's a bot. Ours is not allowed to claim to be human, and in California it has been illegal since July 2019 to use a bot that misleads people about being a bot in order to sell them something.

A chatbot should leave a way to reach a person even when it breaks. If the AI service behind ours goes down, the visitor gets a fixed message with a direct email address, written in code with no AI involved. That fallback is how our commitment to keeping people in the driver's seat holds up during an outage.

Our assistant also forgets on purpose. Each message to it stands alone, with no conversation history, which makes it much harder for someone to walk the bot somewhere over ten messages.

When the answer is no, fix the cheaper thing first

If you answered no to any of the four questions, you don't need a chatbot yet. You need whatever the missing yes was.

• No repeat questions? Your customers need a conversation, and that's you. Make yourself easier to reach.
• Answers not written down? Put plain-English answers to your most common questions on a page on your site. It costs nothing, customers can read it at 11pm, and it is exactly what a chatbot would need later anyway.
• Questions don't arrive after hours? Then the fix is speed. A form that promises a reply time, and keeps it, beats a bot.
• Nobody to own it? Then wait. Somebody has to read what it says, and if that somebody is nobody, you'll hear about it from a customer.

If all four came back yes, cost is the next question, and we laid out real numbers, ours included, in what AI actually costs a small business.

The honest version of a chatbot is smaller than the one in the sales deck. It answers what's written down and says so when it can't. When it matters, it gets a person on the line. That one's worth having. The one that makes up a payback period for a nine-truck HVAC company is not.