What an AI chatbot can do today
Modern chatbots use large language models (LLMs). Unlike the old bots with a tree of canned answers, they understand questions asked in people's own words, can ask follow-up questions and answer in full sentences.
Typical uses in a company:
- customer service – answers about the offer, prices, delivery, returns, opening hours,
- internal assistant – searching procedures, contracts, technical documentation,
- sales support – qualifying enquiries and collecting details before a salesperson gets in touch,
- first-line helpdesk – solving repetitive tickets and handing harder ones over to a person.
How a chatbot knows your company: RAG
A language model doesn't know your price list or terms and conditions on its own. The best way to make it answer from company knowledge is RAG (Retrieval-Augmented Generation).
- Indexing – company documents (website, PDFs, knowledge base, FAQ) are split into passages and stored in a semantic search engine.
- Retrieval – when a user asks a question, the system finds the passages that best match its meaning, not just its words.
- Answer – the model receives the question together with those passages and writes the answer based only on them, ideally citing the source.
The advantage: when you change the price list or the terms, you just update the document – there's no need to "teach" the model from scratch.
How to stop the bot making things up
Language models can sound confident even when they're wrong. A well-designed system limits that risk:
- it answers only from the sources provided, and when they're missing it says plainly that it doesn't know and hands the conversation over to a person,
- it cites the source of its answer so it can be checked,
- it has a clearly defined scope – it doesn't chat about topics unrelated to the company,
- it's tested on a set of real questions before launch and regularly afterwards.
The best implementations start with one narrow use case – for example questions about delivery and returns. Once that works well, the scope grows.
Data, GDPR and security
- choose a model provider and configuration that don't use your data for training and allow data to be processed in the right region,
- keep personal data in conversations to the necessary minimum and set how long it's stored,
- sign a data processing agreement with the provider and include the chatbot in your privacy policy,
- tell users they're talking to an AI – the EU rules on artificial intelligence (AI Act) require this too,
- respect permissions – an internal assistant must not show an employee documents they don't have access to.
This isn't legal advice – for an implementation that processes personal data, it's worth checking the details with a data protection officer or a lawyer.
How much an AI chatbot costs
The cost has two parts:
- implementation – preparing and indexing documents, conversation logic, integration with the website, app or messenger, handing conversations over to a person, testing,
- running costs – fees for using the model (depending on the number and length of conversations), hosting the search engine and the app, and maintaining the knowledge base.
With typical traffic on a company website, the cost of the model itself is often small compared with the working time the chatbot saves. A lot depends on the choice of model and the number of conversations, though – which is why we calculate it from real data about enquiries.
How to get started
- Collect the most common questions from customers or employees – for example from the last few months of emails and tickets.
- Check where the answers are – whether there are up-to-date documents that can be used.
- Choose one channel – website, app, messenger or an internal tool.
- Run a pilot with a limited group and measure the results: how many questions the bot solved on its own and how many it passed on.
A chatbot is just one use of AI – automating document processing or reports often brings just as much. See what we do in AI and automation.
Related services: ai & automation.