A few years ago, a website chatbot meant a scripted menu: press 1 for opening hours, press 2 to leave your email. Today's AI assistants understand the question behind the words, reply in the visitor's language and can look things up in your systems before they answer.
That makes them genuinely useful. It also raises the questions we hear in almost every first meeting: will it make things up? Is our data safe? What does it cost to run? Here is our honest answer, based on what we see when we build them.
What a good AI assistant does well
- Answers the questions you get every day. Opening hours, prices, policies, what is included, how to get there. If the answer exists in your own content, an assistant can find it and phrase it clearly, at 2 a.m. and in Greek, English or German.
- Collects the details before a person steps in. Dates, number of guests, product, budget. By the time an enquiry reaches you, it already has what you need to reply with a quote.
- Works inside your tools. Connected to your booking engine, shop or CRM, an assistant can check availability, look up an order or prepare a booking for you to confirm.
- Helps your own team. Some of the most useful assistants never face a customer. They answer staff questions from manuals, procedures and past emails, so new colleagues stop interrupting the one person who knows everything.
Where they go wrong
- They fill gaps with confident guesses. Language models are built to produce a plausible answer. If the right information is missing, a poorly set-up assistant may invent a price or a policy that sounds right.
- They repeat outdated content. An assistant is only as current as the content behind it. Last year's prices in a forgotten PDF become today's answer.
- They become a wall. Nothing frustrates a customer more than an assistant that will not let them reach a person.
None of these are reasons to avoid AI assistants. They are reasons to build them carefully.
How we build assistants people can trust
- Grounded in approved content. The assistant answers from the sources you choose, shows where an answer came from, and says it does not know when the sources are silent.
- A clear way to a human. Some topics always go to a person, and the visitor can ask for one at any time. The conversation so far travels with the hand-off.
- Tested before launch. We collect real questions from your inbox and phone calls, run them through the assistant and check the answers before anyone outside your team sees it.
- Limited, logged actions. If the assistant can create a booking or change an order, it does so within clear limits, every action is logged, and anything unusual waits for approval.
- Sensible data settings. We use providers and settings that keep your data out of model training, and we only send the assistant what it needs to answer.
What it costs to run
Most AI providers charge by usage, roughly in proportion to how much text goes in and out. For a typical small business assistant, the monthly running cost is usually small next to the staff time it saves, but it depends on traffic, on how much content each answer needs and on the model you use. A useful habit: use a fast, inexpensive model for routine questions and save the larger models for the conversations that need them.
How to start
- Pick one job and one channel. For example: answer pre-booking questions on the website. Not everything, everywhere, on day one.
- Write down your 30 most common questions and the answers you would give. This list becomes both the assistant's content and its first test.
- Tidy the content behind it. Remove outdated pages and files. This usually improves your website too.
- Pilot it internally first. Let your team try to break it for a week.
- Measure. How many questions it answers on its own, how many it hands over, and what visitors ask that you did not expect.
Done this way, an AI assistant stops being a gimmick and becomes a colleague that handles the repetitive questions, so your team can spend its time on the conversations that need a person.
Thinking about an assistant for your website, app or team? See how we build AI assistants, or tell us about your idea.