Our Approach to Data Protection: A Transparent Security Guide

Piotr Grudzień
Published Updated 6 min read
Background image for Our Approach to Data Protection: A Transparent Security Guide

Quickchat AI uses transport encryption, storage encryption and access controls to protect data. PII scrubbing is available on request, rather than automatically applied to every conversation. Model processing, model training and data retention are separate activities: a commitment not to use customer content for training does not mean that no provider receives or stores data.

Our Legal and Security FAQ describes the published controls and enterprise options. This guide, checked against current documentation on November 5, 2026, explains their scope through the data flow. It does not certify a deployment or promise complete prevention of data disclosure, model errors or unauthorized access.

Start with the information your AI Agent needs for its task. A useful Knowledge Base contains relevant, maintained material; adding more confidential data is not automatically necessary or a guarantee of better answers. Review the channels, connected systems and people who can access that information.

StageData involvedControl or decision to confirm
Knowledge Base importThe documents and pages you choose to provideAuthority to use the content, relevance and access permissions
Conversation inputVisitor messages and data supplied through the chosen channelNecessary fields, notices and whether PII scrubbing is configured
Retrieval and ActionsSelected knowledge and data returned by connected servicesEndpoint permissions, visitor authorization and the fields returned
Model processingThe input needed to generate the responseProvider, processing locations, training terms and retention
Response and storageGenerated replies, conversation records and other operational dataRecipient access, review, retention and the appropriate deletion route

Knowledge Base upload and initial processing

Knowledge Base upload

The Knowledge Base accepts supported sources such as written articles, imported web pages and uploaded documents. Choose the material the Agent should use, remove unnecessary personal information, and review the imported content before making it available in customer conversations.

Transport protection: TLS encrypts a connection while data travels between its endpoints. The receiving service decrypts the data to process it. Google’s encryption-in-transit documentation explains this distinction. Transport encryption does not mean that content stays unreadable to the services performing retrieval or generating a response.

Knowledge Base storage and access controls

Knowledge Base storage

Quickchat AI’s published FAQ describes Google Cloud storage in Belgium, logical separation between customers and cloud-provider encryption at rest. Logical separation is not the same promise as a dedicated database, storage bucket or encryption key for every account.

Use the documented team roles to control what signed-in colleagues can reach. Review who can access the Knowledge Base, Inbox, Actions and configuration. Team permissions and visitor access to answers are different boundaries; a colleague’s role does not establish which customer records a chat visitor is entitled to receive.

Integration with user messages

Relevant Knowledge Base content can become part of the context used to answer a visitor. Treat information placed there as material the Agent may use in replies, and review whether it is appropriate for that audience.

Encryption protects transport and storage; it does not decide whether a particular answer should disclose a particular fact. Limit the source material and connected data to the intended task, and test whether the configured Agent respects that boundary.

User message collection

When a visitor sends a message, its contents enter the conversation’s processing flow. A visitor can include personal information even when the Agent did not ask for it. Define the information you need, explain its use and discourage unnecessary sensitive submissions.

Review the selected channel as well as Quickchat AI. A messaging service, helpdesk or connected application may process its own copy under its own terms. A transport-encrypted connection does not remove those parties from the data flow.

Personally Identifiable Information (PII) detection and processing

PII scrubbing is an optional, requested control. Confirm whether it is enabled for your deployment and which data paths it covers. Do not assume that every message is scrubbed or that a setting for message text also covers Knowledge Base content, attachments, metadata, connected systems and every operational record.

Detection can miss personal information, particularly when identity is conveyed through context. Removing an email address or replacing a name does not prove that the remaining text is anonymous. The EDPB’s explanation of anonymisation and pseudonymisation distinguishes these concepts.

Point in the workflowWhat to establishLimit to keep visible
Before processingWhich inputs are covered by the requested scrubbing configurationAn optional control is not a statement about all conversations
During detection and maskingHow the configuration handles representative personal informationA successful example does not prove complete detection
After maskingWhether the remaining content or related records can identify someoneMasked text is not automatically anonymous
Storage and onward useWhich records and providers still receive or retain informationScrubbing is not a replacement for retention and deletion arrangements

Use synthetic examples when evaluating the behavior, including cases where several ordinary details together identify a person. This is evaluation guidance, not a report of a test or a promise about a particular customer’s configuration.

Integration of the Knowledge Base and user message data

Data retrieval from the Knowledge Base

The Agent can use relevant Knowledge Base content and configured Actions to obtain information for a reply. Access to order history, shipping status or another private system requires the appropriate integration and permissions; it is not established merely by uploading a general help article.

Give an integration only the access its task requires. For customer-specific records, the connected service should enforce the visitor’s authorization before returning data. A prompt telling the model not to reveal another customer’s order is not an authorization check. Keep credentials in the appropriate connection configuration, rather than in Knowledge Base text that could be used in an answer.

If your requirements include private networking, dedicated storage or customer-managed encryption keys, confirm the exact architecture and responsibilities with the team. Do not infer that these are standard settings or included features for every account. Private cloud, self-hosted model and on-premises options described in our FAQ require deployment-specific agreement.

LLM processing and AI Guidelines

Data transmission to LLM

Generating a response can require sending conversation context and relevant retrieved information to a model provider. Quickchat AI’s FAQ states that customer content is not used to train LLMs and identifies the use of third-party providers for inference. No training does not mean no transmission or retention. Confirm the provider terms, processing locations and retention applicable to your deployment.

TLS protects transmission to the receiving service. It does not make the model input unreadable to that service, so it should not be described as end-to-end encryption that excludes the processing provider. Optional scrubbing also does not justify a blanket promise that no personal information reaches a provider.

Our Privacy Policy discusses retention and international transfers, including US transfers. Belgian storage alone does not establish that all processing, support access or onward transfers occur exclusively in the EU.

AI Guidelines and filtering

Instructions and configured response rules help define the Agent’s intended scope. They do not guarantee that every restricted topic, sensitive detail or inaccurate answer will be blocked. Use them alongside limited data access, appropriate permissions, evaluation and a human support route.

ControlWhat it addressesWhat it does not establish
TLS and encryption at restProtection during transmission and storageThat authorized processing services cannot read the data
Access roles and integration permissionsWhich people and systems can access defined resourcesThat every generated answer is appropriate for its recipient
Requested PII scrubbingDetection and masking within the configured scopeComplete detection, anonymity or no stored personal data
Guidelines and grounded answersIntended behavior and use of relevant sourcesError-free answers or guaranteed prevention of disclosure
Contract and deployment reviewApplicable responsibilities, providers and termsAutomatic compliance for every customer use case

Data output and response generation

A generated reply may contain information drawn from the conversation, retrieved sources or an Action result. Assess both whether the answer is correct and whether the recipient is entitled to receive it. Masking an input field does not prove that the output contains no personal information.

Use the Inbox to review conversations and available source or execution evidence. Investigate unexpected disclosures and incorrect answers, then adjust the relevant content, permissions or configuration. Grounding and instructions can support this process, but they are not a guarantee against hallucinations or prompt injection.

Configure an appropriate Human Handoff route for questions the Agent should not handle. Verify the actual recipient, notification and service process before relying on it for sensitive or consequential requests.

Data storage for the end output

Data storage (conversations and Knowledge Base)

Conversation records, Knowledge Base content and other operational records may include personal information. Restrict access and establish the retention and deletion arrangements for the records your workflow creates. Encryption at rest does not remove the need to manage those records.

Review the applicable Data Processing Agreement, Privacy Policy and data-deletion information. A no-training commitment is not a zero-retention commitment, and ending a channel connection is not the same operation as erasing conversation history.

The deletion route matters. Deleting an AI Agent makes its conversations inaccessible without erasing them. Account deletion erases the conversations of current Agents in the background, with documented exclusions such as previously deleted Agents’ conversations, attachments, exports, search indexes and external records. Contact support for the applicable erasure process, including retained copies and any legal exceptions; do not assume one button removes every copy immediately.

Summary

Evaluate security against your actual data flow: the sources you upload, the visitors and colleagues who can access them, the services that process them and the records that remain. Confirm optional controls and enterprise requirements in the applicable configuration and agreement.

For a deployment review, bring that data-flow inventory and your requirements to contact@quickchat.ai. Ask which controls apply to the proposed setup, what evidence is available and how retention, access and deletion will work in practice.

Quickchat AI Quickchat AI Ask Quickchat AI a question