The Client Asked for a Chatbot. I Built Them a Form.

"Can you add one of those AI chat things to the site? Everyone has one now."

I get some version of this message every few weeks. And I sell AI consulting. I build AI tooling for clients. Saying yes would be the easy, profitable, entirely defensible answer.

But Gartner surveyed 3,566 B2B and B2C customers in February and March 2026 and found that use of company-provided chatbots has been statistically flat since 2022. Not declining. Flat. Meanwhile use of third-party generative AI tools for the same customer service problems has nearly doubled in the past year, to the point where people are now roughly three times more likely to open ChatGPT to solve a problem with your business than to use the chat widget you paid for.

Sit with that for a second. Your customers already have an AI assistant. They brought their own. The widget in your bottom-right corner is competing with it and losing.

So when someone asks me "do I need an AI chatbot," my honest answer is almost always: probably not, and here is what you actually have.

Nobody Actually Wants a Chatbot

In my experience the request is never really about the chatbot. It is a solution someone reached for before naming the problem, and if you ask two or three follow-up questions the real thing falls out almost immediately.

There are five answers I hear, over and over, and they are not variations of each other. They are five completely different problems wearing the same costume.

"We get the same five questions constantly." Somebody on the team is answering "do you ship to Canada" for the four-hundredth time and would like to stop.

"We miss enquiries outside office hours." Someone lands on the site at 9pm, has a question, and leaves.

"Our contact form doesn't convert." Traffic is fine. Submissions are not.

"A competitor has one." Said out loud less often than it is felt.

"The boss saw a demo." This one arrives with the most urgency and the least specification.

I wrote a whole post about that last case, because a demo is a very specific kind of event with a very predictable aftermath: Your Client's CEO Just Watched an AI Demo. Here's What Actually Happens Next. The pattern is consistent enough that I can usually guess which of the five it is from the wording of the first email.

The diagnosis matters more than the technology. Four of those five problems have a cheaper, faster, more reliable fix than a bot. One of them sometimes genuinely needs one.

The Same Five Questions Deserve a Better Page

If your team answers the same five questions constantly, that is not a staffing problem. That is a content problem, and it is telling you exactly which five pieces of content are missing.

Here is the part that annoys me about the chatbot pitch for this case. A bot that answers "do you ship to Canada" needs a knowledge base containing the sentence "we ship to Canada." You have to write that sentence either way. The chatbot does not save you the writing. It just adds a retrieval layer, a monthly fee, and a chance of getting the answer wrong on top of the writing you were always going to have to do.

Write the sentence. Put it on the page. Put it in the place where the question gets asked, not buried in an FAQ accordion three clicks deep. If people ask about shipping on the product page, the shipping answer belongs on the product page.

There is a nice side effect. Text on a page gets indexed. It shows up in search, it shows up in AI overviews, it works for the visitor who never clicks a chat widget because they have been trained to assume chat widgets are useless. A chatbot's knowledge lives inside the chatbot. A page's knowledge lives everywhere.

Fix the Clock, Not the Interface

The after-hours argument is the strongest of the five, and it still usually points somewhere other than a chatbot.

The classic number here is 21x: leads contacted within five minutes are twenty-one times more likely to qualify than leads contacted after thirty minutes. It gets cited as Harvard research constantly, which is wrong, and worth correcting because the actual provenance is better. That figure comes from a 2007 MIT and InsideSales study covering more than 15,000 leads. The separate 2011 Harvard Business Review study by Oldroyd, McElheran and Elkington audited 2,241 US firms and found an average first response time of 42 hours.

Forty-two hours. That is the real problem in most small businesses, and it has nothing to do with the hours between 6pm and 9am. It is the enquiry that arrived Tuesday at 11am and got replied to on Thursday.

If that is your situation, a chatbot does not fix it. It just moves the delay. The bot collects the question at 9pm, the human still opens it Thursday, and now you have added a step where the visitor was told they were being helped and then was not. That is worse than an honest form that says "we reply within one business day" and then does.

Fix the response process first. Route enquiries to a phone, not just an inbox. Set an actual internal SLA. Tell the visitor on the form what to expect and hold to it. If after all that you still have real volume arriving at 2am from a timezone you do not staff, then you have a genuine case, and it is a case for automation of some kind rather than a case for a chat bubble specifically.

When the Form Is the Problem, Fix the Form

This is the one where the numbers are unambiguous, and it is where the title of this post comes from. The pattern I keep running into is a business convinced their contact form needs replacing with conversational AI, when what the form needs is deleting most of itself.

Baymard Institute's checkout research found the average checkout flow carries 11.3 form fields, and that most stores could complete the same order in about eight. HubSpot's data puts the average conversion cost at roughly 4.1% per additional field. The effect is not linear either: the drop from three fields to five is mild, and then somewhere past five it falls off a cliff.

And then there is Expedia, which is the story every form designer knows. They had an optional field labelled "Company" on their booking form. People misread it as their bank name, then entered their bank's address in the address field, then failed credit card verification. Removing that one optional field was reported to be worth around $12 million a year in profit. One field. Optional. Not even required.

I want to be careful here, because "fewer fields always wins" is a myth that CRO people have been pushing back on for years, and they are right to. Sometimes more fields raise conversion, because a quote request with real qualifying questions signals that you take the enquiry seriously. Sometimes a longer form filters out the tyre-kickers and the leads that do arrive are worth five times more. Field count is a proxy, not a law. The real variable is whether each field feels justified to the person filling it in.

What genuinely kills forms, in my experience, is rarely field count anyway. It is error handling. It is the form that clears everything you typed when validation fails. It is the phone number field that silently rejects the international format. It is the submit button that gives no feedback so people click it four times. It is the form that never says what happens next.

I built a static Astro site's contact form recently and wrote up the whole approach, because getting this right without a server is a genuinely fiddly problem: How to Add a Contact Form to a Static Astro Site (Without SSR). None of the hard parts were AI. All of the hard parts were error states and expectations.

There is a broader version of this too. A form that fails silently is one of a whole family of things that look completely fine on your site and are quietly costing you money: Your Website Looks Fine. Here Are 8 Invisible Problems Still Costing You Customers.

Where a Chatbot Genuinely Is the Right Answer

I am not going to give this one a token paragraph. There are cases where the answer to "do I need an AI chatbot" is a straight yes, and I have built for some of them.

High query volume against a real knowledge base. Not five questions. Five hundred, with documented answers that already exist in a maintained system. The bot is a search interface over content you have, not a substitute for content you have not written.

Transactional lookups. Order status. Booking references. Shipment tracking. This is the strongest case of all, and Gartner's Eric Keller put his finger on exactly why the rest disappoint: most company chatbots are still built to answer questions, while customers increasingly expect AI to take an action for them. A bot that reads your order database and tells someone where their parcel is provides something no FAQ page can. A bot that paraphrases your FAQ page provides less than the FAQ page.

Appointment booking that replaces a phone queue. If your alternative is holding for six minutes, a bot that books the slot is a straight upgrade.

Multilingual first-line support. If you get enquiries in six languages and staff two, this is genuinely transformative and it is the case I see undersold most often.

Displacing a real staffing cost. If you have three people doing first-line triage and the bot handles 40% of it, there is a number you can put on that. If nobody currently does the job, the bot is not saving anything. It is a new expense with a new maintenance burden.

Notice what all five have in common. There is a system behind the bot. Data, a booking engine, a documented knowledge base, a payroll line. If your chatbot has nothing behind it except a language model and some scraped web copy, you have built a wrapper, and I have opinions about those: 6 Signs That AI Product Is Just ChatGPT in a Trench Coat

The Costs Nobody Quotes at the Demo

The demo takes four minutes. The bot answers three softball questions perfectly. Nobody mentions the rest of it, which is where most of the real cost of an AI chatbot actually lives.

Somebody has to build and maintain the knowledge base. This is the real project. Not the integration, which takes an afternoon. Every price change, policy update, new product and seasonal hours change has to reach the bot, and the day it stops reaching the bot is the day the bot starts confidently giving out last year's prices.

Somebody has to own the escalation path. When the bot cannot help, where does the person go? If the answer is "back to the contact form," you have added a step to the thing you were trying to improve. Escalation is the part that is always underspecified and always the part that breaks.

Hallucinated answers become your liability. This is not hypothetical any more. A German court has already ruled on who carries the can when an AI system says something false about a business, and the reasoning cuts both ways depending on who deployed the system: Google's AI Overview Can Now Lie About Your Business. A German Court Just Said That's Google's Problem, Not Yours. When it is your bot on your domain making the claim, you are the deployer. If it quotes a price you do not honour or promises a delivery date you cannot meet, that is your problem.

And people may simply not want it. OnePoll surveyed 6,000 consumers across the US, UK and Canada in April 2026 and found 85% prefer speaking to a real person and 5% prefer AI, with 59% reporting frustration with AI agents, up from 54% six months earlier. I will flag the obvious caveat: that survey was commissioned by AnswerConnect, which sells human answering services, so it is exactly the result they wanted. Treat vendor-funded statistics on both sides of this with suspicion. The chatbot vendors publish equally convenient numbers in the other direction. But the direction of travel in that data matches what Gartner found independently, and it matches what I see in analytics.

What I Actually Tell People

The question is not "do I need an AI chatbot." The question is "what is the specific thing that is going wrong, and what is the cheapest reliable fix for it."

Sometimes the answer is a bot, and when it is I will build it and I will be glad to. More often the answer is three paragraphs of honest copy on the right page, a form trimmed from nine fields to four with error messages that do not lose your typing, and a rule that enquiries get answered inside four hours instead of forty-two.

That is a less exciting invoice. It is also the one that works, and it takes about a tenth as long.

I have a more structured version of this diagnostic for the broader question of whether AI belongs in your business at all: Do You Actually Need an AI Consultant? An Honest Flowchart

And the honest uncertainty, since I would rather say it than pretend: I do not know how long the current numbers hold. If third-party AI assistants keep growing at the rate Gartner measured, the whole question might reframe itself into something else entirely within two years, where the useful work is making your business legible to somebody else's AI rather than running your own. I am watching that closely and I do not have a confident answer yet.

What I do know is that in 2026, most of the people asking me for a chatbot do not have a chatbot problem. They have a copy problem, a form problem, or a response-time problem, and those are all cheaper to fix.

If you want someone to work out which one you actually have before you buy anything, that is most of what I do at Wunderlandmedia. I am a one-person agency in Augsburg, I build AI tooling for clients when it makes sense, and I will tell you when it does not.