AI Chatbots for Customer Enquiries: What They Can and Cannot Do

Chatbots have a reputation problem earned over a decade of bad implementations, and the arrival of far more capable language models has not entirely fixed it. The technology improved faster than the way businesses deploy it.

This covers what an AI assistant on your website can genuinely do now, where it still fails, and how to scope one so it helps customers rather than standing between them and an answer.

What has actually changed

Older chatbots worked from decision trees. They recognised a few phrases, offered buttons, and collapsed the moment someone phrased a question unexpectedly. The experience of being asked to rephrase three times before being offered a contact form is what most people mean when they say they dislike chatbots.

Language-model-based assistants understand phrasing far better. They can read your actual content, answer in natural language, handle follow-up questions that refer back to earlier ones, and deal with a question asked in a way nobody anticipated.

What has not changed is that they are only as good as the information they are given, and that they will produce a confident answer whether or not they have the facts.

What they do well

Answering questions that are already documented. Opening hours, service areas, what is included, how a process works, delivery times, policies. If the answer exists on your site or in a document you can supply, an assistant can retrieve it reliably and phrase it appropriately.

Handling the first contact outside working hours. A meaningful share of enquiries arrive in the evening. An assistant that answers the straightforward questions and captures details for the rest converts attention that would otherwise be lost by morning.

Qualifying and routing. Asking a few natural questions to understand what someone needs, then directing them to the right page or passing a structured summary to the right person. This is often more valuable than answering, because it reduces the back and forth before a human conversation starts.

Reducing repetitive enquiries. Most businesses answer the same ten questions endlessly. Handling those frees time for the conversations that need a person.

What they still cannot do reliably

Quote prices in anything but the simplest cases. If pricing depends on assessment, site conditions or scope, an assistant guessing a figure creates a commitment problem. Ranges and the factors that drive cost are fine; a specific quote is not.

Make commitments. Promising availability, agreeing a deadline or confirming something is covered are all decisions with consequences. These need a human or a system with real data behind it.

Handle complaints well. An upset customer wants to be heard by someone who can act. Routing them through an assistant first usually makes it worse, and it is the single most common cause of the goodwill damage these tools cause.

Answer what it was not given. A model asked something outside its supplied information will often produce a plausible answer anyway. On a business website, a confidently wrong answer about a policy or a product is worse than no answer.

Scoping one properly

The difference between a useful assistant and an irritating one is almost entirely scope.

Define what it answers, drawn from your actual content rather than from general knowledge. Define what it must never attempt: pricing commitments, availability, anything regulated, anything involving a complaint. Define how it escalates, and make that path short and obvious rather than buried.

The rule worth holding to is that it should never be harder to reach a person than it would be without the assistant. An assistant that intercepts people trying to make contact is a net loss regardless of how many questions it answers.

Grounding it in your own information

An assistant should answer from your content: your service pages, your policies, your documentation. That is what keeps answers accurate and consistent with what the rest of the site says.

This has a useful side effect. Setting one up usually reveals that your own content does not actually answer common questions, which is worth fixing regardless of whether the assistant launches. If the assistant cannot find the answer, neither can your customers, which is the problem covered in why pages fail to convert.

Keep it current. An assistant confidently quoting last year’s policy because nobody updated its source material is a liability, and this is where the maintenance cost sits.

Being honest with customers

Tell people they are talking to an automated assistant. Attempting to pass it off as a person damages trust when they work it out, and they usually do.

Being upfront also sets expectations usefully: people ask simpler questions of an assistant and are less annoyed when it cannot help, provided reaching a person is easy.

Where it fits against other automation

A customer-facing assistant is a more visible project than most automation and carries more risk, because mistakes are seen by customers rather than staff. For many businesses, the better first project is internal: getting enquiries captured, acknowledged and routed quickly, which produces a measurable improvement in response time without any customer-facing risk.

If response speed is the problem you are solving, an automated acknowledgement with a realistic timeframe often achieves more than a chatbot, and is far simpler. The prioritisation is covered in which processes to automate first.

Measuring whether it helps

Track the things that indicate value rather than volume. How many conversations were resolved without a person. How many produced a qualified enquiry. How often people asked for a human, and how quickly they got one. Whether total enquiries went up or down after launch.

That last one matters most. If enquiries fell after deploying an assistant, it is intercepting people rather than helping them, and the deployment is costing you business regardless of how many questions it answered.

Deciding whether you need one

An assistant is worth considering if you receive a high volume of repetitive questions, if a meaningful share of enquiries arrive outside working hours, or if people routinely contact you to ask something the site should have answered.

It is not worth it if your enquiry volume is low, if every enquiry is different, or if your customers expect a personal response from the first contact. In high-value service businesses, the first interaction is often part of what is being bought.

If you are unsure which applies, tell us what questions you get most often and how enquiries reach you now. We will tell you whether an assistant would help, whether better content would solve it more cheaply, and what it would need to be scoped to. Our AI automation services cover both, and the wider framework is in our complete guide to AI automation.