AI chatbots for business: what they bring and how to implement them

An AI chatbot is a conversational assistant that answers questions from website visitors — or from a company's clients and staff — in natural language, using a language model and, above all, the company's own information: services, prices, opening hours, documentation or policies. Done well, it responds at any hour, in several languages, and turns enquiries into qualified leads. Done badly, it makes up answers and scares clients away. Below we cover what it brings, which types exist and what it costs, using the one we built for our own website as an example.

What does an AI chatbot bring to a business?

A chatbot's value isn't "having AI" — it's solving three very specific problems that almost every small business has:

• Out-of-hours support: a large share of website visits come in the evening, at weekends or from another time zone. Without an immediate answer, many of those visitors leave.
• Repeated questions: ballpark prices, timelines, areas you cover, how the service works. An assistant answers them consistently well and frees up your team's time.
• Lead capture and qualification: instead of a cold form, a conversation that understands what the client needs and hands the sales team a useful summary.

On top of that, (anonymised) conversations are a goldmine: they tell you what your clients ask, which doubts hold back a purchase and what content your website is missing.

Case study: how the assistant on Somia Digital's website works

Before recommending it to anyone, we built it for ourselves. Our assistant is the same one you see in the corner of this website, and it's built on these principles:

• Three languages: it replies in English, Spanish or Catalan depending on the page and the user's language.
• It knows our real content: services, FAQs, rates and contact details come straight from the website's content, so when we change a rate, the assistant changes it too.
• Clear rules on pricing: it always talks about "from" prices and never invents fixed quotes.
• Context-aware: the greeting and suggested quick replies change depending on the page you're on (arriving from the SEO service page is not the same as arriving from a blog post).
• Qualifies and sends leads: when a visitor wants us to get in touch, the assistant collects their details, rates the lead as hot, warm or cold based on the interest and urgency shown in the conversation, and emails us the conversation so we can reply with context.
• Abuse limits: a maximum number of messages per session and per day, bot protection and size-limited conversations, so costs always stay under control.
• Privacy and consent: the chat history is kept in the browser session and disappears when the tab is closed; our server keeps no copy. That doesn't mean it never leaves the browser: each message is sent to Google's Gemini API to generate the reply, and if you ask us to get in touch, the conversation is included in the internal email we receive. Analytics (which topics are asked about, never the text) are only enabled if the user has accepted cookies.
• A route to a human: at any point it offers WhatsApp, email or phone.

You can try it right now on the Somia Digital assistant page: ask it what one of your clients would ask and see how it answers.

What types of chatbot are there?

Not all chatbots are the same, and choosing the right type is the most important decision. There are three main families:

1. FAQ bot: works with a decision tree or predefined answers. It's cheap and predictable, but limited: if the question isn't on the list, it can't answer. Useful for very narrow use cases.
2. AI assistant with its own knowledge (RAG): a language model that, before answering, looks up the company's information (website, catalogues, manuals, policies) and bases its reply on it. RAG stands for retrieval-augmented generation. It's the option we recommend for most SMEs: it understands questions phrased in a thousand ways and answers with real data.
3. Agent with actions: as well as answering, it can do things: book an appointment, check an order's status in the ERP, open a support ticket or send a quote. It adds a lot of value, but requires integrations, well-defined permissions and more testing.

The underlying technology can come from different providers — models from Google, Anthropic or OpenAI, depending on the case. The choice depends on cost per conversation, quality in the languages you need, speed and each provider's privacy terms.

How do you implement an AI chatbot, step by step?

1. Define goals: customer service, lead generation, internal support or a mix. Each goal changes the tone and the rules.
2. Prepare the knowledge: review and organise the information the assistant will be able to use. If your website or documentation is out of date, the bot will be too.
3. Write the rules: what it can and can't say, how to talk about prices, when to hand over to a person, how to handle personal data.
4. Integrate it: on the website, with email or your CRM for leads, and with analytics while respecting consent.
5. Set limits: message quotas, bot protection and a maximum conversation size.
6. Test with real questions, including awkward or off-topic ones, before going live.
7. Review it regularly: read sample conversations, fine-tune answers and add missing knowledge.

How much does an AI chatbot cost?

The cost has two parts. The first is implementation: preparing the knowledge, defining the rules, integrating and testing. It depends heavily on scope: a website assistant using existing content is a well-defined project; an agent connected to your ERP is a development project. The second is usage: every conversation consumes a small amount of credit from the model provider. With efficient models and a well-designed context — reused across conversations — the cost per conversation is usually a matter of cents. That's why limits matter: the risk isn't normal use, it's abuse.

How do you know whether the chatbot is working?

An assistant should be measured like any other sales tool. These are the metrics we review every month on ours:

• Open rate: what percentage of visitors interact with it, and from which pages.
• Most-asked topics: they reveal what the website doesn't explain well enough and what content needs creating.
• Leads generated and their quality: how many contacts arrive this way and how many turn into quotes.
• Answer ratings: a simple "helpful" or "not helpful" button on the last reply gives very valuable clues.
• Handovers to a person: if they're too frequent, knowledge is missing; if there are none, the bot may not be offering that route clearly enough.

Which mistakes should you avoid?

• A bot without its own knowledge: a generic model will confidently say things about your company that aren't true. It must always be anchored to your information.
• Never offering a human: the frustration of not being able to talk to someone is what hurts your brand the most.
• Hiding that it's an AI: since 2 August 2026, the EU Artificial Intelligence Act generally requires people to be informed when they are interacting with an AI system. More information from the European Commission.
• Ignoring data protection: you need to explain how conversations are handled, ask only for necessary data and respect consent for analytics.
• Launching it and forgetting it: an assistant improves when it's reviewed. Without maintenance, it goes stale.

Does your business need a chatbot?

If you get lots of repeated questions, out-of-hours visits or clients who speak different languages, probably yes. If your website doesn't yet explain clearly what you do, fix that first: an assistant can only be as good as the information it has. If you'd like to see what it could do for your company, try our assistant, tell us about your case or message us on WhatsApp. We'll help you choose between an FAQ bot, an assistant with its own knowledge or an agent, depending on what you really need.