Customers expect fast answers. A chatbot can help. Whether it helps depends on data flow, goals and operations — not on the plugin alone.
Chapter 1: what to check when selecting a plugin
Many plugins send questions to external services. Without contract, logging and a data-flow review, that is a risk for SMEs. Whether the EU AI Act applies depends on role, use case, system category, market placement and other facts; a concrete case must be assessed and may require legal review [1].
Privacy (FADP)
The revised Swiss Data Protection Act (FADP) has been in force since 1 September 2023 [2]. It requires proportionality, transparency and security, among other things. A plugin without a documented data flow does not automatically meet that.
Hallucinations and trust
Language models can produce false statements. For SMEs that is a reputation issue. Answers must come from approved sources, not from model memory.
Chapter 2: Hybrid Mode — architecture, not a promise
Hybrid Mode separates sensitive content from external inference. Whether it works depends on rules, tests and operations.
Data abstraction (DAL)
A local pre-check can mark or remove PII — only if that module is active.
Placeholders replace identifiers; originals stay in a project-defined store.
Only approved context is sent to the external model.
Answers are reassembled locally.
What a chatbot can do — and what it cannot
A chatbot can answer from approved content. It does not replace binding advice.
A chatbot can capture and structure requests. It does not replace human confirmation.
For complex cases, escalation to people is sensible. The chatbot is a tool, not a substitute for responsibility.
Chapter 3: Public prices
Public prices are fixed and limited: Short Check, SOURCE/01 Full and mandate. RAG has no public numeric price.
First alignment, one domain, limited scope. CHF 0.
Separately bookable in-depth audit with a documentation package. CHF 2'450 one-time.
Ongoing visibility work after the appropriate audit or baseline. From CHF 1'500 / month.
Internal company knowledge, FADP-aware, project-specific. No public figure — scope and price are set in an introductory call.
Current public prices and next step
Chapter 4: Define goals — without borrowed KPIs
Define goals from your own baseline and measure them during a controlled pilot. A chatbot is not an end in itself.
Which requests come in today?
Which of them are repetitive and approved?
Which answers may be automated?
How is handover to people handled?
Visible FAQ, plugin or a dedicated RAG system?
Choose the technical form for the task. A small, rarely changing set of approved answers may be clearer directly on the website. A plugin can fit when data flows, maintenance and answer limits are documented. A dedicated RAG system becomes relevant for multiple sources, differentiated permissions or integrations. No option is superior merely because of its category.
| Option | Suitable starting case | Check before use |
|---|---|---|
| Visible FAQ | A few recurring questions with stable answers | Specialist approval and contact path. |
| Chat plugin | Bounded task with a documented service chain | Recipients, retention, cost and fallback. |
| Dedicated RAG | Distributed knowledge and role-based answers | Source permissions, evaluation questions and operation. |
Fictional pilot: a team selects 20 common support questions, including five without an approved answer. It measures source accuracy, correct escalation and handling time. This sample is a project template, not a universal quality standard. Comparison with the existing workflow shows whether the additional operational work is justified.
Conclusion
A chatbot can help when data flow, goals and operations are right. GlasBox reviews that project by project.
Kacper Ruta · owner of Ruta Group · GLASBOX Studio
Sources
EU AI Act — applicability
FADP / nDSG, in force 1 September 2023
Review the chatbot setup?
We review data flow, goals and operations. No prices beyond the public list.
Request Consultation