Answer faster, with an AI that says it is one.

Three concrete uses

1. 01A chatbot for frequently asked questions

   AI Act level · Transparency obligation

   1. 01Validate the articlesOne owner per topic
   ↓→
   2. 02Announce the AIFrom the very first message
   ↓→
   3. 03Answer and citeOnly from validated articles
   ↓→
   4. CheckTakeover by an adviser
   ↓→
   5. 04Hand overThe adviser gets the full history

   The problem

   Customers mostly ask simple questions: order tracking, opening hours, procedures. Advisers spend time on them that is missing for difficult cases.

   What the AI does

   The chatbot answers from your validated knowledge articles, introduces itself as AI from the first message and hands over to an adviser as soon as it leaves its scope. It promises no goodwill gesture, refund or deadline that your rules do not provide for.

   Data needed

   An up-to-date knowledge base with an owner for each article, and order status in read-only mode.

   Who signs off

   An adviser takes over when the customer asks or when the chatbot reaches its limit; the service manager approves every new article.

   A system that talks directly to the public must say it is AI, at the latest at the first exchange, unless that is obvious (EU · AI Act, Art. 50, applicable since 2 August 2026).

   • AI agent
   • RAG
   • Guardrails

2. 02Suggesting a reply and a summary to the adviser

   AI Act level · Minimal risk

   1. 01Identify the customerHistory and active contracts
   ↓→
   2. 02Find the procedureCurrent version only
   ↓→
   3. 03Suggest a draftSourced reply and summary
   ↓→
   4. CheckSign-off by the adviser
   ↓→
   5. 04Send and saveAfter the adviser’s review

   The problem

   During a call or a written exchange, the adviser searches several tools for the customer’s history and the applicable procedure. The end-of-contact summary is rushed, or skipped.

   What the AI does

   The assistant shows the relevant history, suggests a sourced draft reply and writes the summary into the CRM. The adviser rereads, corrects and sends; the assistant sends nothing on its own and does not assess the adviser.

   Data needed

   Customer history in the CRM, current procedures and coded contact reasons.

   Who signs off

   The adviser approves every reply before it is sent and every summary before it is saved.

   Minimal risk while the tool assists the adviser; used to evaluate advisers’ performance or behaviour, it would become high-risk (EU · AI Act, Annex III), and inferring their emotions at work is prohibited (Art. 5).

   • RAG
   • Prompt
   • Human oversight

3. 03Sorting complaints and spotting recurring causes

   AI Act level · Minimal risk

   1. 01Collect the messagesAll channels, one stream
   ↓→
   2. 02Sort by reasonAgainst your validated list
   ↓→
   3. 03Group the causesMonthly table by reason
   ↓→
   4. CheckSign-off by the quality manager
   ↓→
   5. 04Decide on actionsQuality manager, with evidence

   The problem

   Complaints arrive by email, phone, web form and social media, each coded its own way. Management sees a volume, not causes.

   What the AI does

   AI sorts each message against your list of complaint reasons, flags urgent ones and groups the causes that keep coming back. It does not reply to the customer and decides on no compensation.

   Data needed

   Incoming messages, the list of reasons and the rule that separates a complaint from a simple request.

   Who signs off

   A supervisor checks a sample of classifications each week and handles urgent cases; the quality manager decides on corrective action.

   Minimal risk (EU · AI Act); messages contain personal data, sometimes sensitive without the customer meaning it to be: limit what the tool keeps and set a retention period (EU · GDPR, Art. 5).

   • Embeddings
   • Evaluation
   • Semantic layer