GreetScribe AI Use Policy
Version: 2.0
Updated: 2026-09-20
Operator: Qevinto Oy
GreetScribe combines structured conversation paths, knowledge retrieval and AI-generated answers. This policy explains their boundaries and supplements the Customer Terms, Privacy Policy and Data Protection Addendum.
1. What the service does
Buttons, forms, calendars, deterministic routing and booking handoffs do not all require a language-model call. The knowledge-answer path can ask a model to formulate an answer using selected knowledge content. A failed or insufficiently grounded answer can lead to a fallback or contact path instead of an invented answer.
The guest widget displays an AI-assistant notice above the conversation. The notice is available in Finnish and English; its wording can be configured but an empty override falls back to the built-in text. The customer must keep a clear, accessible disclosure and privacy information available to guests.
This visible notice is not a claim that all legal duties concerning machine-readable marking of generated content have been satisfied.
2. AI provider and request contents
The standard AI service is Scaleway Generative APIs. Compatibility with another provider's API format or client library does not mean requests are sent to that provider.
| Function | Information sent |
|---|---|
| Guest knowledge answer | System instructions, current guest question, titles and text of retrieved knowledge entries. |
| Embedding retrieval, where enabled | Query or knowledge texts supplied for vectorisation, with the configured embedding model. |
| Staff AI authoring, where enabled | Staff instructions and relevant authoring context selected by that function. |
The inspected guest request does not automatically attach the full transcript, session/booking record or a separate name/email field. This does not guarantee the absence of personal information: free text and knowledge excerpts may contain it. The inference path does not establish that all input has been anonymised.
The service also contains configuration support for alternative model endpoints. Such support is not evidence of their use. A different recipient requires an approved customer arrangement and updated processing/transfer information before customer data is routed to it.
3. Scaleway handling
Scaleway's published Generative APIs policy provides zero retention of prompt content by default, with anonymised API-usage information retained for up to six months. It allows temporary access/storage of full HTTP requests in specified service-harming error or misuse investigations. For the HTTP 500 error example, the policy states a maximum of two weeks for reproducing, investigating and fixing the issue; this should not be read as a universal deletion guarantee for every misuse scenario.
Scaleway states that inputs/outputs are not used to train, retrain or improve its underlying base models and are not available to their creators. This provider rule is separate from Qevinto's own conversation storage and product-development processing. See Scaleway's policy.
4. Product development and model training
Qevinto's product-development purpose is limited to improving GreetScribe's quality, software, prompts and knowledge content. It includes selected original conversations and anonymised findings subject to the restrictions, lawful-basis requirements, access limits and development-copy deadline in Privacy Policy section 4.
Guest data is not used to train or fine-tune model weights. Changing prompts or knowledge content is different from training a model. Personal or customer-confidential information must not be introduced into a shared knowledge resource accessible by other customers; only appropriately cleared content and genuinely anonymous findings may be reused there.
Independent development is not authorised simply by the Customer's processor instructions or by continued use of the widget. Identifiable conversations must not be copied into unapproved external AI or development tools.
5. Accuracy and human choice
Generated answers can be wrong, incomplete or unsuitable. Retrieval and output checks reduce particular risks; they do not guarantee correctness. Customers must keep supplied knowledge accurate, test configured flows and provide an appropriate contact route.
GreetScribe does not use the LLM to accept or reject bookings or payments. Guests choose whether to follow an external booking link. The booking/payment provider confirms availability, applicable terms and completion.
Do not rely on GreetScribe for emergency, medical, legal, financial or other high-stakes decisions. Do not submit secrets, payment-card information or unnecessary sensitive personal data.
6. Security and usage controls
The implementation includes tenant-scoped retrieval, target validation, bounded requests, retry limits and output validation. Availability and cost controls depend on configuration and the relevant runtime path; they are not an unlimited availability guarantee.
A billable response is a successfully delivered AI-generated answer. Deterministic menu/booking steps, failed answers and internal messages are not separately billed as AI responses. Additional response capacity requires advance written authorisation under the Billing Policy. There is no standard guest-token or model-token overage tariff.
7. Customer duties and changes
Customers must use the service lawfully, provide appropriate guest notices, supply content they may lawfully use, and avoid configuring collection of unnecessary sensitive information. They must not use the assistant to deceive guests, impersonate a human or bypass safety restrictions.
Material changes to recipients, purposes or incorporated contractual terms follow the relevant privacy, DPA and contract procedures. Publishing new wording does not itself authorise a new use of personal data.
Qevinto Oy, Business ID 3583823-3, Sarkakatu 3 F23, 96460 Rovaniemi, Finland. contact@qevinto.com. Previous published version.