The AI assistant

The AI assistant answers your visitors' questions when nobody from your team is answering at the moment. It answers from your knowledge base: the pages of your website, uploaded files, help articles and improved answers. Whatever it cannot back up, it passes on to your team.

The assistant is switched off for new projects. While it is off, every question goes straight to your team and, if nobody is signed in, to the form for name and email.

Switching it on and setting it up

The settings are on the project page, AI & knowledge tab, Assistant section. Owners and admins may change them.

• Let the assistant answer – the main switch.
• Assistant name – up to 40 characters. If it is left empty, the default name applies. A change appears in the chat within a minute; ongoing conversations keep the old name in their history.
• Type of website – not set, Online shop, Blog / magazine, Company / service provider or Portal / application. With “Online shop” the assistant answers general questions about shipping, withdrawal and returns even when your pages say nothing about them; such answers are labelled as general information.
• How our team speaks – one or two sentences on tone, up to 400 characters. Empty means “friendly and natural”.
• Topic range – how far beyond your website the assistant may answer (the three levels are listed directly below).
• Allow work requests (code, texts, translations) – off by default. As long as it is off, the assistant writes no code, no essays and no translations for visitors, whatever topic range applies.
• Show AI status in the chat – on by default. The visitor then sees what the assistant is working on.
• Automatic translation in the desk – the default for new conversations (on paid plans). See The agent desk.

The three levels of the topic range:

• Our website only – only what is in your knowledge base.
• Website + small talk (default) – plus greetings, thanks and brief small talk.
• Website + general knowledge – plus general questions that have nothing to do with your website.

The narrower the level, the more often the assistant hands over to your team.

There is no language setting for the assistant: it answers in the language the question was asked in. It works best in German, English and Polish.

In the reactive.chat inside the chat section you can, from the Professional plan, specify that the assistant does not talk about reactive.chat, and hide the “Powered by reactive.chat” note. On the free plan both stay on.

How the assistant looks in the chat and which texts it uses is described in the Customisation part.

How an answer comes about

1. Search. The assistant searches your project's knowledge base for matching passages. For questions about the chat itself it also knows a collection of information about reactive.chat.
2. Phrase. From the passages found it writes an answer. Answers based on your content name their sources; the visitor can open them and jump to the passage.
3. Check. Before the answer goes out, a fixed set of rules checks details where mistakes do particular harm: numbers and prices, weekdays, “weekend” and “working days”, “free of charge”. If such a detail is not in the sources used or on the page the visitor is looking at, the answer is not delivered. Instead a verbatim quote from your knowledge base, a fixed sentence or the handover to your team goes out.

This check is targeted and no guarantee that every answer is correct in substance. What the assistant answered and how the check decided you can see in the AI log (below).

If you wish, answers appear while they are being written, sentence by sentence.

What the assistant remembers

Within a conversation it knows the latest messages and a summary of the older ones. In very long conversations, details from the beginning can get lost.

Across conversations it only remembers something if you switch on Memory across conversations (/app/projects/…/memory, on paid plans, off by default). It then makes at most one short note per conversation, but only for visitors with a signed identity or with consent on the same device. You see and delete the notes on the same page; they are deleted automatically after 180 days, earlier if the retention period is shorter. There you also switch Suggest follow-up questions under bot answers on or off (on by default).

Images and files

If a visitor sends an image, the assistant can look at it if this is switched on for your project: one image per message, scaled down and without location data. From attached PDF, text and CSV files it reads the beginning as well.

When the assistant hands over to your team

• The visitor wants a person. If they say so, or click Talk to the team, the handover is immediate.
• Nothing found to back it up. If the assistant finds nothing on a question about your website, it says so and hands over.
• A rule requires it, or it is outside service hours. The reasons appear in the Reports under “Why it was handed over”.
• Job applications. If the assistant recognises a job application, it answers with a fixed sentence and hands over to a person. After that no AI works in this conversation any more, except translation. This cannot be switched off.

If someone asks something that has nothing at all to do with your website, the assistant answers itself or says politely that it only gives information about your website, depending on the topic range; there is no handover then. There are no more than three such general-knowledge answers per conversation.

If nobody is online at a handover, the visitor gets the form to leave a message.

When someone from your team takes over a conversation, the assistant falls silent. With Return to AI or /ki in the desk it answers again and then knows what your team has written.

What the assistant does not do

• It does not pretend to be a person. There is a notice before its first answer, and each of its messages carries the AI badge. You may not remove this labelling.
• It does not take on other work requests unless you allow them. A multi-layered protection recognises such attempts before the answer.
• It answers the question “What happens to my data?” with a fixed, checked text instead of a freely worded one: the chat stores the conversation, closing it deletes nothing, you as the website operator set the retention period, and deletion works in the chat with a sentence such as “Delete my data”.

Still, check for yourself what is in your knowledge base: the assistant repeats what you give it, including mistakes, outdated details and instructions that appear in your texts.

Monitoring and improving

Below the Assistant section you find the tools for this.

AI log

/app/projects/…/ai-log shows every answer of the assistant with question, answer, type (AI-worded, quote, fixed sentence, proactive sentence), check result (passed; flagged, replaced; no model text; unchecked) and sources. In the detail view you jump to the passage in the source and open the conversation in the desk.

You can filter by period, “Flagged only”, “Without source only” and “Include fixed sentences”. The CSV or JSON export uses the same filters, up to 10,000 rows. The log is deleted together with the conversation. Anyone who may read conversations may read it.

Knowledge gaps

/app/projects/…/knowledge-gaps collects the questions of the last 90 days for which the assistant answered without a source, found nothing, handed over or got a “thumbs down”, and groups them into topics. For each topic you can:

• Improve answer – store a correct answer (see below),
• Create article – write a help article, with an AI draft if you like (at most 5 per day and project; it stays a draft until you publish it),
• Done – hide it until the question comes back,
• Ignore – hide it permanently.

Analyse now refreshes the list.

Improved answers

Under /app/projects/…/answers you store the Correct answer (up to 3000 characters) for a question (up to 300 characters). It becomes a “Q&A” source that the assistant finds with priority, and takes effect within a few minutes. Remove names and order numbers from the question, because the source can be shown to visitors.

Test runs

/app/projects/…/simulation (owners and admins) runs test questions through the same answer path as in the chat, but without a real conversation, without notifications and without entries in reports and log.

• For each question you set a key statement that should appear in the answer, or that a handover is expected.
• Test against the current settings, a saved version or a draft, which you put live with Apply draft.
• Visitor types – normal, impatient, vague or in Polish.
• Suggestions for test questions come from knowledge gaps, improved answers and your knowledge base.

A question passes if the answer has a source and contains the key statement (every number must appear), or if it was handed over as expected. At most 100 test questions per project; the page shows how many runs are still possible today. A test run always leaves enough AI allowance for your visitors.

Versions

/app/projects/…/bot-versions records every saved change to the assistant settings with time and person. You can Compare two versions, start a Test run against this version or Roll back to this version. Rolling back applies in the chat immediately and is itself a new version.

Assistant actions

From the Professional plan the assistant can do more than answer (/app/projects/…/actions, owners and admins). On the free plan you can look at everything but not save.

Built-in actions:

• Order status – with Customer data in the chat, only for signed-in, verified visitors.
• Available appointments and Book an appointment – with the Cal.com integration.
• Subscription and invoices – with the Stripe integration; the search uses the verified customer number.

On top of that there are up to 20 own HTTP actions: name, title, description, method and an https:// address with placeholders, with a secret or signature if you like, and, if you like, only for verified visitors. The assistant only sees the response fields you release.

Actions that change something (such as a booking) are only carried out by the assistant once the visitor explicitly agrees. Take care with the Test button: a changing action is really carried out. The technical details are in the Technical integration part.

Customer data and shopping assistant

• Customer data in the chat (on paid plans) – the assistant answers signed-in customers' questions about order, shipping and invoice, only about their own data. Everything about it: Customer data in the chat.
• Shopping assistant (/app/projects/…/shop, on paid plans) – after its answer the assistant shows matching product cards with image, price, availability and link. Prices and availability come only from your product catalogue, never from the AI. On your shop domain the visitor can add items to the cart directly (WooCommerce, Shopify, Shopware, or without a shop just cards with a link). There is also a discount code on checkout abandonment and the “Revenue with chat” report. The report shows a connection in time, not proof that the chat triggered the purchase.

Smart Answer in your website's search

Under /app/projects/…/suche you switch on Smart Answer for your website's search field. When a visitor searches there, a short answer from your knowledge base appears above the results, with sources. Smart Answer only answers with backing, never from general knowledge. A follow-up question can lead into the chat if you wish; such conversations carry the “Site search” label in the desk.

Among other things you can set whether search queries are stored as conversations, the answer style, alternative terms, when the answer is triggered, which search field it appears at and what it looks like. The answer box stays invisible as long as the assistant is switched off or the project is not active. If the service cannot be reached at some point, the page says so, and you can save the settings later.

Own AI node

On paid plans you can have your workspace's AI jobs run on a machine of your own under Own AI node (/app/ai-node, owners and admins):

• reactive.chat nodes – the default.
• Own node preferred – your node gets every job first; if it does not pick it up within a few seconds, reactive.chat does the work (“fallback”).
• Own node only – if your node cannot be reached, there is no AI answer, but a handover or the offline form.

Up to 5 nodes per workspace. The software for the node is open source and only needs outgoing HTTPS. Conversations are still stored at reactive.chat, and the check, allowances and abuse protection still apply. The setup is described in the Technical integration part.

Limits and allowance

• Per conversation and minute and per website and minute there is a maximum number of AI requests, which rises with the plan (see Plans and limits).
• If the allowance per website is used up, visitors get a fixed sentence until the end of the minute (“A lot of questions are arriving here at once right now …”). Your team can keep answering.
• If the allowance per conversation is used up, the conversation is handed over to your team.
• Copilot, translation, AI tags and the AI rating in AutoQA count towards the same allowance.
• There is no fixed number of AI answers per month.
• It can happen that AI is temporarily switched off for a project. Your team then takes over, and the chat keeps running.

How reactive.chat uses AI and what that means legally is described in detail in AI Act: how reactive.chat uses artificial intelligence.