Why Customers Hate Your Chatbot (And How to Fix It)

Nobody has ever complained about a chatbot that answered their question. The complaints, every one of them, are about chatbots that got in the way of a person trying to give a business money.

I have spent a lot of time reading transcripts, both from systems I built and from systems businesses asked me to look at after they stopped working. The failures are remarkably consistent. There are about seven chatbot mistakes that account for nearly all of the damage, and most of them are configuration decisions rather than limitations of the technology.

They are ranked below by how much they cost you, worst first.

Trapping people in a loop with no exit#

The single worst thing a chat window can do is refuse to let someone out. You know the pattern. The customer asks something specific, the bot offers four buttons, none of them relevant. They type the question again in plainer words. The bot offers the same four buttons. On the third attempt it says "I didn't quite get that, let's try again" and shows the buttons.

At this point the customer is not evaluating your service. They are working out whether it is faster to find your competitor.

The fix is a rule you can write in one line: after two consecutive failures on the same topic, stop trying and offer a person. Not on the fifth attempt. The second. A bot that says "I'm not getting this right, let me grab someone who can help, or leave me your number and we'll call you within the hour" recovers most of the goodwill it just burned. A bot that tries a third time does not.

While you are at it, allow free text everywhere. Buttons are fine as suggestions and terrible as the only input, because the question a customer wants to ask is never one of your four options.

Hiding the human like it's a state secret#

Some businesses deliberately bury the escalation path, on the theory that if people cannot reach a human, support costs go down. Support costs do go down. So does revenue, and it takes about six months for anyone to connect the two.

Customers who want a person want a person. Making them fight for it does not change their mind, it just adds a grievance to whatever they were originally calling about. The ones who give up do not become self-service converts. They become former customers, quietly, without telling you why.

Put the option in front of them. Mention it in the opening message, honour the word "human" or "agent" or "person" the instant it appears, and never respond to a request for a person with another attempt at self-service. When someone asks for a human and gets "I can help with that! What's your order number?", they are being told their request does not count.

The mechanics of doing the transfer well, rather than just doing it, are in the handoff. Doing it badly is only marginally better than not doing it.

Pretending to be a person#

Giving your bot a human name and a stock photo of a friendly woman called Emma feels harmless. It is not, because customers work it out. They always work it out, usually within two messages, and the moment of realisation costs you more than honesty ever would.

What follows is worse. The customer, now suspicious, starts testing. They ask something off-topic to see what happens. They get annoyed at Emma for wasting their time. Then they tell their partner about the company that lied to them about who they were talking to, which is a memorable story in a way that "the website has a chatbot" is not.

Say what it is. "Hi, I'm the assistant for Westside Plumbing" costs you nothing and buys you enormous latitude, because expectations drop to a level you can comfortably exceed. Nobody is offended by an assistant that answers well.

Answering a question nobody asked#

The customer types: "do you replace hot water cylinders in Titirangi". The bot replies with three paragraphs about the company's commitment to quality workmanship and a link to the services page.

This is what happens when a bot has been pointed at a marketing website and left to fend for itself. Retrieval finds the page with the most surface similarity, which is usually the one stuffed with the same words in a promotional register, and the customer gets atmosphere instead of an answer. The cause and the cure are both covered in training an AI agent on your business, and the short version is that a curated set of direct answers beats a full site crawl every time.

The related sin is the confident wrong answer. A bot that does not know something should say so. If yours has never once said "I don't have that information, let me get someone who does", it is not being accurate, it is being unsupervised.

Making people repeat themselves after the transfer#

Customer explains the whole problem to the bot. Bot escalates. Human arrives and says "Hi, how can I help today?"

Everything the customer just typed has been thrown away, and they now have to type it again to a second entity in the same window. This is the point where a merely irritating experience becomes a bad chatbot experience worth complaining about publicly, because the customer has done work that the business discarded.

Pass the full transcript. Every escalation should hand the human the conversation so far plus a one-line summary of what the customer wants, and the human's first message should demonstrate they read it. "Hi Sarah, I can see you're after a cylinder replacement in Titirangi and the bot couldn't confirm timing. I can" is a completely different experience to "how can I help".

Popping up before anyone has read anything#

The visitor lands on your homepage. Before the page has finished loading, a chat window slides up over the content with a cheerful "Hi there! Have any questions?"

They have not read a word yet. They have no questions. What they have is a box covering the thing they were trying to look at, on a phone, with a close button the size of a grain of rice.

Delay the trigger until someone has been on a page long enough to be reading it, or until they show intent by reaching a pricing or contact page. On mobile, use a small persistent button rather than an overlay. And test the close button on an actual phone, held in one hand, by someone who is not you. Placement and trigger timing are covered properly in adding AI chat to your website.

Being nine seconds slower than useful, at the wrong hours#

A chat widget that is only staffed 9 to 5 on weekdays is a chat widget that is absent for most of the week. The visitor at 8pm gets "leave a message and we'll be in touch", which is a contact form wearing a costume.

The whole point of putting an agent on the channel is that it does not sleep. If yours only works during the hours you were already answering the phone, it is solving nothing. The pattern of when enquiries actually arrive is worth checking in your own data, and the method for doing that is in most of your enquiries arrive after hours.

The chatbot mistakes ranked by what they cost#

How to find out whether yours is doing this#

Read fifty transcripts. Not the summary dashboard, which will report a containment rate and imply everything is fine. The actual conversations, in order, including the short ones where somebody typed one message and left.

Look for the point in each conversation where the customer's tone changes, because that is where your bot did something stupid. Nearly all of the chatbot mistakes above announce themselves this way. Short messages getting shorter, repeated rephrasing of the same question, all caps, the word "human" appearing anywhere. Each of those is a specific fixable failure with a timestamp attached.

Then fix the top three and read fifty more. Most chatbot best practices articles will give you principles, and this list of chatbot mistakes is no exception. Fifty transcripts will give you a to-do list, and the to-do list is worth more.

If you are earlier than that and still deciding what to build, the full picture of how these systems work, what they cost and how they hand over to a person is in the AI receptionist for small business guide. The technology is fine. Almost every complaint traces back to somebody choosing to make it obstructive.