Building an AI chatbot that genuinely answers your customers' questions doesn't start with the technology, but with your own information. A modern chatbot uses an existing language model and looks up the answer in your FAQs, terms and documents. Built well, it answers the questions that keep coming back, day and night, and hands the rest to a colleague. This article explains how that works, what a chatbot can and can't do, where the costs are and how to start small without putting your customer relationships at risk.

The short version

  • A good chatbot answers from your sources, not from a language model's general memory. That keeps answers accurate and checkable.
  • Handing over to a person is part of it. When in doubt, for complaints or personal situations, your team takes over.
  • The costs come in three parts: the build, language model usage per conversation, and keeping your information up to date.
  • Start small: one channel, the twenty questions you get most often, and measure the results before you expand.

From button menus to an assistant that reads your sources

The chatbots of a few years ago ran on fixed rules: if the bot spotted the word "delivery", it showed the standard answer about delivery times. If a customer phrased the question slightly differently, it got stuck. That's where chatbots got their bad name.

An AI chatbot works differently. It understands a question in plain language, even with typos or in a different order. It then searches your own information for the passages about that question, such as your knowledge base, your terms, product information or manuals. The language model writes an answer based on exactly those passages. This principle is called retrieval: look it up first, then answer.

The difference is big. The bot doesn't need to know every question in advance, and when your information changes, the answer changes with it. Want a broader view of where AI saves time in a business first? Read AI in business: practical applications.

What a good chatbot can and can't do

Where a chatbot is strong

  • Recurring questions: opening hours, delivery times, return policies, prices, how something works.
  • Searching lots of information: finding the right answer across dozens of pages of product information or specialist knowledge.
  • Always being available: also in the evening and at weekends, when your team is off.
  • Asking follow-up questions: helping a customer pin down the question and passing it on, summarised, to your team.

What a chatbot isn't for

  • Decisions with consequences: approving a refund, changing a contract, or giving medical or legal advice.
  • Complaints and emotions: an angry customer wants a person, not a politely worded standard answer.
  • Information that doesn't exist: if the answer isn't anywhere in your sources, the chatbot can't give it properly either.

Handing over to a person

A chatbot with no way out frustrates customers more than no chatbot at all. So agree up front when a conversation goes to a colleague: when the bot can't find the answer in its sources, when the customer asks for it, or for topics you want to handle yourself. The colleague receives the conversation so far, so the customer doesn't have to repeat anything. Outside office hours the bot can record the question and say when an answer will follow.

Limiting made-up answers

Language models can give convincing-sounding answers that are wrong. With customer questions, that's the biggest risk. You limit it with a few clear rules:

  • Answer only from your own sources. The chatbot uses your information, not whatever the model once read somewhere.
  • Show the sources. A link to the page the answer came from lets customers and your team check it.
  • "I don't know" is allowed. Better an honest "I don't have an answer to that, let me pass you on" than a guess.
  • Fence off sensitive topics. Decide which questions the bot always passes on, such as payments or complaints.
  • Spot-check. Regularly read back conversations and improve the sources where answers fall short.

Privacy, GDPR and transparency

A chatbot processes whatever customers type, and that sometimes includes personal data. So pay attention to:

  • Which data goes to the model provider. The customer's question and the passages found are sent to the language model. Choose a provider and settings where that data isn't used to train models, and put it in a data processing agreement.
  • No more data than needed. For general questions the chatbot doesn't need to see a customer file. Connect only what's truly necessary.
  • How long conversations are kept. Agree a retention period and mention the chatbot in your privacy policy.
  • Make clear it's AI. Users should know they're talking to an AI assistant, not a colleague. The European AI Act asks for that transparency too.

Off-the-shelf widget or custom-built?

Off-the-shelf widget

There are chatbot services where you upload a few documents and paste a widget onto your site. That fits when you mostly get general questions, your information sits on a handful of pages and you want to test quickly whether customers use a chatbot at all. You are tied to what the package can do, though, and integrations with your own systems are often limited.

Custom-built

A custom chatbot pays off when answers depend on your own systems, such as the status of an order, when you have a lot of specialist information, when the chatbot has to work inside your own app, or when you want to control exactly how data is processed. It's a bigger project, but the chatbot then does what your process needs rather than what a standard package allows.

What does an AI chatbot cost?

There's no fixed price for "a chatbot", because the scope varies widely. The costs, however, always come from the same three parts:

  • Build and setup. Gathering and organising your information, building the chatbot, setting up the handover to your team and any integrations with your systems. This is a one-off cost.
  • Language model usage. Language model providers charge per amount of text processed. So you pay per conversation, and the total grows with the number of questions.
  • Maintaining your information. When your prices, terms or products change, the sources have to follow. That's mostly your own team's time, plus the occasional technical update.

At Gain Impact you get a fixed price and deadline for the build after a free intake. How such a fixed price works is explained in How much does it cost to build an app?; the same principles apply to a chatbot.

How to start small

  1. Pick one channel, for example the chat on your website, not WhatsApp, email and your app all at once.
  2. Collect the twenty questions you get most often and check that the answers are written down somewhere and up to date.
  3. Set the boundaries: which topics always go to a person, and what does the bot say when it doesn't know?
  4. Test with your own team before customers see the chatbot, and read back plenty of conversations in the first weeks.
  5. Expand based on what you see: more topics, a second channel or an integration with your systems.

Tip: the questions your chatbot answers with "I don't know" are gold. They show exactly which information is missing from your website, including for customers who never use a chatbot.

How do you measure whether it works?

  • Questions resolved on their own: what share of conversations ends without a colleague having to step in?
  • Handovers: how often does a conversation go to your team, and on which topics? That's where the next improvement lies.
  • Satisfaction: ask with one click after the conversation whether the answer helped.
  • Time saved for your team: fewer recurring questions in the inbox and on the phone.

From practice: Rik at Roof Connect

Roof Connect is the largest digital trade platform for the Dutch roofing industry. Alongside the website and the iOS and Android app, we built Rik, an AI assistant that answers questions about roofing, insulation and materials. Roofers ask their trade questions right on the platform where they also read their news and find suppliers. We also co-own an AI product of our own, Woonbot.

How we approach it

We build on existing language models and tailor them to your information and processes, rather than developing a model from scratch. The AI supports your team, with clear limits where a person reviews or approves. Your data stays yours: before we start, we make clear agreements on storage, processing and access. We prefer to start small, with one task as a pilot, and measure what it delivers. After a free intake you get a fixed price and deadline.

Curious which questions in your business a chatbot could answer first? Read more about custom AI integrations, see our process or plan a free intake.