Sunday, 2 August 2026 — just a few days away — the transparency obligations of the AI Act come into force, including the one that concerns chatbots directly: the user must know they are talking to a program, not to a human. If your company already has a chatbot on its website or you receive weekly offers for one, this is the right moment to sort out three questions: what a chatbot can realistically do in 2026, what its cost is made of and what the law asks of you starting this week.
This article is written for owners and directors of companies with 10–100 employees, with no technical background. You will not find promises about “the AI revolution” here, but an evaluation method: what to request at quotation, what to ask before signing and how to compare offers without going by the numbers in the brochures.
What a chatbot can realistically do in 2026
A chatbot built on today's language models no longer resembles the rigid menus of a few years ago, where you pressed “1” for opening hours and “2” for the address. Today it can:
- answer the repetitive questions — opening hours, delivery times, return conditions, required documents — based on your company's documents;
- qualify enquiries: it gathers the essential details (name, company, what the client is looking for, in what timeframe) before a colleague takes over the conversation;
- book meetings directly in the team's calendar;
- check an order's status, if it is connected to the internal system;
- work across several channels — website, WhatsApp, Facebook Messenger — with the same knowledge base.
Just as important is what it cannot do. A chatbot left unsupervised can invent plausible but false answers — the phenomenon is called hallucination. Any serious implementation includes two safety nets: restricting the answers to the documents approved by the company and handing the conversation to a human when the question falls outside the covered area. If the offer on your table mentions neither, you have a warning sign.
What the cost is made of — and why there is no list price
Nobody can give a fair price for a chatbot without knowing what volume of conversations you have, how many channels you want, which systems must be connected and who maintains the content. You can, however, understand the structure of the cost, so you compare offers component by component, not as a block:
- Building the knowledge base. Someone has to gather, verify and structure the company's information: prices, procedures, commercial terms. It is the most often underestimated part — and the one that decides whether the bot answers correctly.
- Implementation and integrations. A simple widget on the website is one end of the spectrum; connecting to the CRM, the ordering system or the calendar is the other. Every integration adds days of work and points to test.
- The running cost. Language models are usually paid per use, so the monthly cost grows with the conversation volume. Ask for it separately from the implementation cost and ask for a simulation at your realistic volume.
- Content maintenance. Prices and terms change; someone has to update the knowledge base, otherwise the bot becomes a source of outdated information.
- Quality supervision. Someone periodically reads samples of conversations and corrects the slips before clients see them.
The practical rule: request each component as a separate line in the offer. Any amount you hear before the discussion of these five points is an indicative estimate, not a price — it is confirmed only at quotation, after the vendor understands your situation. A company that refuses the breakdown is telling you, in fact, that it does not want to be compared.
Chatbot or agent? Clarify what you are buying
In offers, the words “chatbot” and “AI agent” are often mixed together, but the difference matters: a chatbot answers and directs, an agent executes actions — it issues documents, modifies an order, sends emails in the company's name. Complexity, cost and risks rise significantly with agents; we detailed the difference in the comparison between AI agents and classic automation. For most small companies, the healthy first step is a chatbot that answers and qualifies — not an agent that executes.
What the AI Act asks of you from 2 August
From Sunday, the transparency obligations of Article 50 of the AI Act, the enforcement powers for general-purpose models and the entire penalty regime apply. For a company chatbot, concretely:
- the user must be clearly informed that they are interacting with an AI — not hidden in terms and conditions, but visible in the conversation;
- AI-generated content must carry a machine-readable marking, so it can be identified as such;
- deepfakes must be labelled — less relevant for a support bot, but important if you use AI in video or audio materials.
A useful clarification, because a lot of confusion is circulating: the “Digital Omnibus” amendments, approved by the European Parliament on 16 June 2026, postponed the obligations for high-risk systems — the standalone ones from 2 August 2026 to 2 December 2027, and those embedded in products to 2 August 2028. Transparency, however, was not postponed. A support or sales chatbot is not, as a rule, a high-risk system; its main obligation remains transparency — and it applies from now on. In practice: ask the vendor for written confirmation that the bot presents itself as AI and that content marking is active. If the bot is already live, check for yourself: open a conversation and see whether you learn, without asking, that you are talking to a program.
When a chatbot makes no sense
A chatbot is a good investment only if it has something to solve. Signs that it is not the moment:
- a small volume of repetitive questions — if you receive a few enquiries a day, a person handles them better and cheaper;
- unstable content — if the company's offer changes weekly and nobody has time to update the knowledge base, the bot will mislead clients;
- consultative selling — where the client is buying trust in a person, putting a bot in between can cost more than it saves;
- budget only for implementation — without a monthly budget for running and maintenance, the project dies within a few months.
The cheap and often ignored alternative: a well-written frequently-asked-questions page and a form asking the same qualifying questions the bot would ask. Many companies discover this solves half the problem almost for free — and that only the remainder justifies a chatbot.
The next step
Before requesting offers, do a one-month inventory: put on a list the questions that repeat by email, on the phone and on WhatsApp, with the frequency of each. That list is your brief — with it, the offers become comparable and broken down into the five components above. Answers to the uncertainties that usually appear in such projects can be found in the frequently asked questions about AI and digital marketing, and if you want an assessment applied to your company's situation, Neoxis does this type of analysis as part of its AI and digital marketing services.