AI agents that answer from your reviewed knowledge base

The short answer: Supportifier's AI agents answer only what is in your reviewed knowledge base and the customer's own data. Every answer is checked against the source before the customer sees it, and when the material does not cover the question the AI hands over to a person instead of guessing. The knowledge base is built from what you already have and you review it before it is used.

Where the answers come from

The AI uses four sources and nothing else: the knowledge base you have reviewed, the whole conversation so far, customer data from connected systems such as orders and deliveries, and your own instructions. There is no general web knowledge in the answer. If those four do not cover the question, the AI says so rather than filling in.

In the chat and the AI form the agent searches the published articles, reads the conversation and, when the customer is identified and a connection exists, fetches order status, delivery and payment, and answers with links to the articles it used. In the inbox it reads the whole ticket thread and the customer's earlier tickets and lays out a reply suggestion for the agent.

The source check

Every answer and every reply suggestion is fact-checked against the sources: dates, prices, terms and links. The answer is labelled Safe, Review or Uncertain. A claim that is not in the source stops the automatic reply. Only safe answers with support in the knowledge base are sent automatically; everything else goes to an agent as a suggestion.

Automatic replies are switched on per ticket category, and a category can first run in shadow mode, where you see what would have been sent without anything going out. In live mode each category's accuracy is followed, and a category that slips is flagged with the suggestion to go back to drafts only; nothing is switched off behind your back.

How the knowledge base is built

You import knowledge from web addresses, from documents (PDF, Word, Markdown, text, CSV, HTML) or from the answered emails in a connected inbox. Files and emails are planned into one article per topic, contradictions between sources are flagged, and everything lands in a review queue. Nothing is used until you have approved it.

The system also learns from what agents send, but carefully: sent answers are not stored raw. Names, personal identity numbers, email addresses, phone numbers and customer-specific details are removed and the question is generalised. When the same knowledge has been reinforced several times, a draft article is proposed for review, and it stays inactive until a person approves it. Learned articles can never be published in the public help centre.

When the AI cannot answer

When the AI lacks the material it notes what was missing. Those notes, together with help centre searches without a hit and unanswered chatbot questions, are collected as knowledge gaps with a Create article button on each row. A draft that an agent rewrites more than half also counts as a signal that knowledge is missing. The gaps are the to-do list for the knowledge base.

Languages

The AI answers in the customer's language regardless of the workspace language, and incoming messages can be translated. The agent app is in Swedish and English; help centre, chat and form are available in Swedish, English, Norwegian, Danish, Finnish and German.

What the AI can do, not just say

Where you switch it on, the AI can also act: cancel an unshipped order, refund, change a delivery address or open a return in Shopify, WooCommerce and Magento, and issue a credit note in Business Central. Every action is off by default, each one is set to off, suggest to an agent or automatic, within limits such as the cancellation window and the size of a discount, and everything is logged. See automated returns, refunds and credit notes.

Read more on the platform page, the AI chat page and the AI form page.