AI chatbot for business: how to choose one for your website

Most businesses that buy an AI chatbot for their website make the decision on a demo. The demo shows a friendly assistant answering a few well-chosen questions, and it looks finished. What the demo does not show is what happens when a customer asks something the bot has never seen, when the answer sits in a PDF nobody has updated since 2023, or when the visitor is angry and wants a person. Those moments decide whether an AI chatbot for business earns its place on your site or quietly damages it.

The stakes are not abstract. In February 2024 a British Columbia tribunal ordered Air Canada to pay damages after its website chatbot gave a customer wrong information about bereavement fares, and it rejected the airline's argument that the chatbot was a separate entity responsible for its own actions, as summarised by the American Bar Association. Whatever your bot says, you said it. This guide explains how to choose one: what to test, what to ask vendors, what the rules require, and where a chatbot stops being the right tool.

Quick answer

Choose an AI chatbot for your business by testing five things before you pay: whether it answers only from your approved content, whether it hands over to a human with the full conversation, whether it captures leads into your CRM, whether you can read and measure every conversation, and how your visitors' data is handled. Run it against your hardest real questions, not the vendor's demo script.

What an AI chatbot for business actually does

A business chatbot sits on your website, and sometimes in WhatsApp or your help centre, and talks to visitors in natural language. The current generation is built on large language models, which means it can understand a badly phrased question and reply in fluent sentences. That is the easy part. The hard part is making sure the reply is true for your business.

A language model on its own knows nothing about your prices, your opening hours, your returns policy or the terms in your contracts. It will still produce a confident answer, because producing fluent text is what it does. Useful business chatbots solve this by retrieving passages from your own documents and instructing the model to answer only from them. The approach is usually called retrieval-augmented generation, and we cover how it works for company knowledge in our guide to an AI knowledge base.

It is worth being clear about scope. A chatbot answers questions and collects information. It does not, by default, update your CRM, issue a refund, rebook an appointment or chase an invoice. Tools that take actions across your systems are a different category, and the difference matters enough that we wrote a separate piece on AI agents versus chatbots. If you are unsure which one you need, read that first, then come back here.

Start with the job, not the product

Vendors sell chatbots as general-purpose. Buyers get better results when they pick one narrow job and judge every product against it. A job is specific enough to measure. Here are the ones we see most often among small and mid-sized firms:

Write the job down in one sentence, with a number attached. For example: handle the thirty most common questions without a human and pass every sales enquiry to the team within one minute. A chatbot that cannot be judged against a sentence like that cannot be improved either.

Knowledge grounding: the feature that matters most

If you only test one thing, test whether the bot invents answers. Ask for the question your customers actually fear: a refund after the deadline, a price for something you do not sell, a policy that changed last month. Then watch what it does.

A well-built bot answers from your content and cites where the answer came from. When it does not know, it says so and offers a person. A poorly built bot fills the gap with something plausible. The Air Canada case turned on exactly this: the bot gave a customer a refund rule that did not exist, and the tribunal held that the airline was responsible for it. According to the same ABA summary, the tribunal also noted that a customer cannot be expected to double-check one part of a website against another.

Do not assume that retrieval removes the problem. A Stanford study of leading legal research tools found that AI products built on retrieval still hallucinated between 17% and 33% of the time, and the researchers concluded that retrieval reduces hallucinations compared with a general-purpose model but does not eliminate them. Legal research is a harder domain than a retail FAQ, so your numbers should be better, but the lesson transfers: grounding lowers the risk and testing tells you by how much.

A practical grounding test takes an afternoon:

Wrong answers are normal in the first week. What you are buying is the speed and ease with which you can fix them.

Human handover and the escalation path

Customers are wary of being trapped with a machine. A Gartner survey of 5,728 customers, published in July 2024, found that 64% would prefer companies did not use AI in customer service, and 53% would consider switching to a competitor if they learned a company was going to use it. The top concern was that AI would make it harder to reach a person. Gartner's advice was to tell customers when AI is in use, tell them they can reach a person, and make sure the human agent picks up where the chatbot left off.

That gives you a clear specification for handover. When you evaluate a chatbot, check each of these:

Lead capture and CRM integration

For many businesses the chatbot is a sales tool first. A bot that answers questions but lets the visitor leave without a name, an email address or a next step is a missed opportunity. Look at how lead capture works in practice.

The best patterns ask for contact details at the point of value, after the bot has helped, rather than as a gate at the start. They store the conversation with the lead, so the salesperson sees what the visitor asked about. They push the record to your CRM with sensible fields filled in, and they avoid creating duplicates when a returning visitor gives the same email again.

Ask which CRMs have a native integration and which rely on a generic webhook that your team must maintain, and whether the bot checks for an existing contact before creating a record.

Analytics: can you see what customers are really asking

A chatbot is also a listening post. Every conversation records what visitors wanted, in their own words, at the moment they wanted it. Many teams never look. Choose a product that makes looking easy.

The numbers worth tracking are modest. Containment, meaning the share of conversations resolved without a person, is the headline figure, but read it alongside satisfaction and re-contact rate, because a bot can contain a conversation simply by frustrating the customer into leaving. Track unanswered questions, since that list is a ready-made content plan for your website. Track handover reasons, conversion from chat to lead, and the topics that rise and fall each month.

Insist on access to full transcripts, search across them, and a way to flag a bad answer and trace it back to its source document. Be wary of products that show only a dashboard summary. If you cannot read the conversations, you cannot improve the bot, and you cannot answer a complaint about what it said.

Data protection and the rules that now apply

Three sets of obligations are worth understanding before you sign anything. None of them is a reason to avoid a chatbot, but each shapes what a good deployment looks like.

Telling people they are talking to AI

In the European Union, Article 50 of the AI Act requires providers of AI systems that interact directly with people to design them so that people are informed they are dealing with AI. According to the European Commission's FAQ on Article 50, the notice must come at the start of the first interaction, be clear and distinguishable, and meet accessibility requirements. The rule applies from 2 August 2026, and the Commission says the exception for cases where the interaction is obvious should be read restrictively. The practical conclusion is simple: label your bot as an AI assistant from the first message and do not give it a human name and photograph to imply otherwise.

Personal data and privacy information

Chat conversations contain personal data: names, emails, order details and sometimes things people should not have typed at all. The UK Information Commissioner's Office transparency guidance for AI says people should receive appropriate and timely privacy information about how AI processes their data, in plain language, and not buried in long terms and conditions. It also says every party in the supply chain should be able to show compliance with the transparency provisions of UK GDPR. The ICO notes that this guidance is under review because of the Data (Use and Access) Act, so check the current version when you write your own notice.

In practice, ask the vendor where conversations are stored, which sub-processors see them, whether your data is used to train shared models, how long transcripts are kept, and how a deletion request is handled. Put the answers in a data processing agreement. Our guide to GDPR and custom AI agents in the UK and EU goes through this in more depth.

Security of the bot itself

A chatbot that reads text from the public is a target. In December 2025 the UK National Cyber Security Centre warned in a post on prompt injection that large language models cannot reliably tell data from instructions, and that prompt injection may never be fully mitigated in the way SQL injection can be. Its advice was to reduce the likelihood and impact of attacks rather than assume they can be stopped.

For a website chatbot that translates into design choices. Give the bot the minimum access it needs. Do not connect it to systems where a manipulated reply could cause real harm. Keep customer records out of its reach unless the visitor has been verified. Put a human approval step in front of anything that moves money or changes data. A bot that only answers questions from public content is far lower risk than one wired into your back office.

Pricing models to understand before you compare

Chatbot pricing is hard to compare because vendors meter different things. Some charge per seat for your staff, some per conversation, some per resolved conversation, and some per message. Each model rewards different behaviour. A per-resolution price looks fair until you realise that the definition of resolved is set by the vendor. A flat monthly fee looks predictable until traffic grows and you hit a cap.

Do not rely on a review site for numbers, because vendors change them often. Take the figures from the vendor's own pricing page on the day you decide, and ask for the definition of every billable unit in writing. Then model three scenarios: your current traffic, double it, and a peak month. Add the cost of the human time needed to maintain the bot, because a chatbot that nobody owns degrades within weeks. Our article on how much an AI agent costs a small business lays out the cost items beyond licence fees.

Off-the-shelf, platform builder or custom build

There are three routes, and each fits a different situation.

An off-the-shelf chatbot tool, usually bundled with a help desk or website platform, is the quickest to launch. You paste in your website address, it learns the content, and you go live within a day. It suits businesses with simple, public information and a standard support process. The limits appear when you need unusual handover rules, deep CRM logic or control over where data lives.

A platform builder gives you more control. You assemble the knowledge sources, the conversation rules and the integrations yourself, usually with a visual editor. It suits teams with someone who can own the configuration and a willingness to test properly. The risk is that the build becomes a side project that nobody finishes.

A custom build makes sense when the chatbot has to work inside a specific process, such as qualifying an enquiry against your own criteria, pulling live availability, or writing structured records into a system of record. We compare these routes in detail in custom AI agent versus off-the-shelf tool. The honest answer for many small firms is to start with the simplest tool that passes your sixty-question test, and move up only when you hit a limit you can name.

How Klevere approaches website chatbots

Klevere is an international AI agency. We have deployed more than 500 AI agents across over 50 projects in 12 industries, and we keep a 98% client retention rate. That experience shapes how we treat chatbot projects, and it is mostly about restraint.

We begin with the job and the evidence, not the tool. That means reading your enquiry history, identifying the questions that account for most volume, and checking that approved answers exist for them. If the answers are not written down, the first piece of work is writing them, because no model can compensate for a missing policy.

We then build the bot to answer only from approved sources, to say plainly when it does not know, and to hand over to a named person with the full transcript. We design the escalation rules with your team, since they know which complaints are urgent and which topics should never be handled by software. Where the project calls for actions rather than answers, such as updating a record or booking a slot, we scope it as an agent with limited permissions and a clear approval step. Our AI agent development service and the AI support agent describe how that looks in practice, and our overview of AI customer support covers where it fits in a service operation.

Finally, we stay on after launch. Review of unanswered questions, bad answers and handover reasons is where the value comes from, and it is the part most vendors leave to the customer.

Frequently asked questions

What is the best AI chatbot for a small business website?

There is no single best product. The best choice is the one that passes a test on your own questions, hands over to a person cleanly, and connects to your CRM. For simple public information, a bundled tool is often enough. For qualification logic or live data, a configured or custom build is a better fit.

How much does an AI chatbot for business cost?

It depends on the pricing model, which may be per seat, per conversation or per resolution. Always take figures from the vendor's own pricing page on the day you buy, and ask how each billable unit is defined. Add staff time for maintenance, because an unowned chatbot gets worse over time.

Can an AI chatbot replace my customer service team?

Not entirely, and customers often do not want it to. It can absorb repetitive questions, but complaints, exceptions and sensitive cases need people. Gartner's survey found that customers worry most about being unable to reach a human, so keep that route visible and fast.

Do I have to tell visitors they are talking to AI?

In the EU, Article 50 of the AI Act requires people to be told they are interacting with AI, from 2 August 2026, unless it is obvious. Even where you are not covered, disclosure is good practice and builds trust. Label the bot clearly from its first message.

Is an AI chatbot safe to use with customer data?

It can be, if you control what it can access. Confirm where conversations are stored, who processes them, how long they are kept, and whether they train shared models. Limit the bot's permissions, put people in front of sensitive actions, and give visitors clear privacy information.

What is the difference between a chatbot and an AI agent?

A chatbot mainly answers questions and collects details. An AI agent can take actions across your systems, such as updating records, booking appointments or sending follow-ups, within limits you set. Many businesses start with a chatbot and add agent abilities once the answers are reliable.

If you would like a second opinion before you commit to a product, a free AI audit will show where a chatbot would help, where it would not, and what to fix first.