A Deep Dive into Amazon Quick Desktop's Data Handling Policy Using the Official Documentation
This page has been translated by machine translation. View original
Recently, the Amazon Quick desktop application was released in preview for macOS and Windows.
It adds many features not available in the browser version, such as direct access to local files, connections to MCP servers, integration with coding agents like Kiro and Claude Code, and scheduled tasks run by background agents. It looks set to become an AI assistant on your desktop that you can entrust with a variety of business tasks.
When using such AI services for work, what matters is the policy on how data is handled. Classmethod has drafted and published guidelines for the business use of AI services.
The guidelines are continuously updated, but as conditions for using an AI service for work, they require confirmation from perspectives such as the following.
That information entered by users is not used for training the service as a whole, and that the input data itself is not retained by the service provider
The article introduced at the beginning covers the overview and features of Amazon Quick Desktop, but for business use, you need to check "whether data is used for training" and "whether data is retained by the service side". In this article, I share the results of an in-depth investigation into the data handling policy, focusing mainly on the "Security, privacy, and architecture" section of the official documentation.
Summary of Findings
Here is the conclusion first. Detailed evidence is explained in the following sections.
| Guideline requirement | Result | Basis |
|---|---|---|
| Not used for training | ✅ Cleared | Explicitly denied in multiple official sources (D1, S2, F1, B2) |
| Data not retained (persistent data) | ✅ Cleared | Local-first design. Conversation history, memory, and file indexes are not uploaded to the cloud (*the knowledge graph is backed up; see below) |
| Data not retained (inference-time logs) | ⚠️ Not disclosed | The official documentation does not state whether the payloads of inference requests (via API Gateway) are logged on the AWS side |
However, there are the following points to note.
- ⚠️ When making AI inference requests, data is sent to the cloud over HTTPS (processing is not entirely local)
- ⚠️ The knowledge graph is backed up to your AWS account for cross-device continuity (but it is not used for AI model training)
- ⚠️ The backend AI service configuration (such as specific model names) is not officially disclosed
- ⚠️ Amazon Quick Desktop is in the preview stage
Scope of Investigation and Official Sources
What Is Amazon Quick Desktop?
Please refer to the article introduced at the beginning for the overview and features. The desktop version is available on all plans: Free, Plus, Professional, and Enterprise (D2). New sign-ups include a 30-day free trial of the Plus plan (F1).
Official Documents Reviewed
I distinguish between documents specific to the desktop version and documents that apply to Amazon Quick as a whole. In the body text, each document is referenced by its ID (D1, S2, etc.).
List of investigated sources (click to expand)
Desktop-specific documents (primary basis)
| ID | Document name | Overview |
|---|---|---|
| D1 | Security, privacy, and architecture | Local-first design, no use of data for training, data storage locations |
| D2 | What is Amazon Quick on desktop? | Definition of local-first architecture, communication via API Gateway |
| D3 | Enterprise setup | Inference Region (us-east-1), OIDC authentication flow |
| D4 | Working with chat (desktop) | AI model modes, Amazon Nova Canvas |
| D5 | Settings | Amazon Federate for Bedrock (authentication section), privacy settings |
| D6 | System tools | OS-level sandbox, system tool permissions |
| D7 | Knowledge graph | How to use the knowledge graph, local storage, backup to AWS account |
- D1: https://docs.aws.amazon.com/quick/latest/userguide/desktop-security.html
- D2: https://docs.aws.amazon.com/quick/latest/userguide/what-is-desktop.html
- D3: https://docs.aws.amazon.com/quick/latest/userguide/desktop-enterprise-setup.html
- D4: https://docs.aws.amazon.com/quick/latest/userguide/working-with-chat-desktop.html
- D5: https://docs.aws.amazon.com/quick/latest/userguide/desktop-settings.html
- D6: https://docs.aws.amazon.com/quick/latest/userguide/system-tools-desktop.html
- D7: https://docs.aws.amazon.com/quick/latest/userguide/knowledge-graph-desktop.html
Quick-wide / FAQ / Blog (supplementary basis)
| ID | Document name | Overview |
|---|---|---|
| S2 | Data protection in Amazon Quick | No use of customer data for training, TLS 1.2+ recommended |
| F1 | Amazon Quick FAQs | No use for training on any plan, HIPAA/FedRAMP/SOC 2 |
| B2 | About Amazon - Quick Desktop launch | "built on AWS", no use of data for training |
| B3 | AWS What's New - Desktop Preview | Preview release, us-east-1 |
- S2: https://docs.aws.amazon.com/quick/latest/userguide/sec-data-protection.html
- F1: https://aws.amazon.com/quick/faqs/
- B2: https://www.aboutamazon.com/news/aws/amazon-quick-desktop-ai-assistant
- B3: https://aws.amazon.com/about-aws/whats-new/2026/04/amazon-quick-macos-windows-preview/
Navigating the Documentation (Supplement for Readers)
Here are some useful points to know when checking the Amazon Quick documentation yourself.
- The Amazon Quick documentation shows remnants of the former QuickSight foundation. Pages that still have
hidden_service_name=QuickSightin the footer or metadata, or whose Related resources links point toaws.amazon.com/quicksight/, are likely derived from the old QuickSight documentation - Desktop-specific information is consolidated under "Amazon Quick on desktop" in the sidebar (pages whose URLs start with
desktop-). For security and privacy,desktop-security.htmlis the most reliable source data-encryption.htmlunder the general sectionsec-data-protection.html(Data protection) mainly covers encryption related to SPICE (QuickSight's in-memory engine) and dashboards, so it does not necessarily apply directly to AI inference data in the desktop version- It is important to distinguish between descriptions that apply to Quick as a whole (e.g., S2) and those specific to the desktop version (e.g., D1)
Finding 1: Use of Data for Training
Regarding the guideline requirement "Information entered by users is not used for training the service as a whole," I checked multiple official sources.
Desktop-Specific Documentation (D1)
At the beginning of the Security, privacy, and architecture page, the Important box contains the following.
Your data is never used for AI model training. AWS does not use your conversations, files, or personal context to train or improve AI models.
Data Protection for Quick as a Whole (S2)
Amazon Quick does not use customer data for training or improving underlying LLMs.
This statement applies to Amazon Quick as a whole, not just the desktop version.
FAQ (F1)
The "Plans and signup" section contains the following.
Is my data used to train AI models?
No. AWS does not use your data to train models in any Amazon Quick plan—free or paid.
This applies to both free and paid plans.
About Amazon Blog (B2)
Quick never uses your data to train someone else's model.
*Because B2 is a marketing blog, I prioritize D1/S2/F1 as formal evidence.
Summary
| Source | Scope | Original text (excerpt) |
|---|---|---|
| D1 | Desktop version | "Your data is never used for AI model training." |
| S2 | Quick as a whole | "does not use customer data for training or improving underlying LLMs." |
| F1 | Quick as a whole (all plans) | "AWS does not use your data to train models in any Amazon Quick plan—free or paid." |
| B2 | Quick as a whole | "Quick never uses your data to train someone else's model." |
Four official sources consistently confirm that data is not used for AI model training.
Finding 2: Data Storage and Retention
I checked the guideline requirement "The input data itself is not retained by the service provider."
Local-First Architecture
It is defined in D2 (What is Amazon Quick on desktop?).
The desktop application uses a local-first architecture. The AI backend runs locally on your machine, and your files stay on your computer. The only network calls are to AI models through API Gateway and to your connected services (such as Slack, Outlook, or Gmail).
Network communication occurs only for inference requests to AI models (via API Gateway) and for communication with external services (Slack, Outlook, etc.).
Data Stored Locally
D1 (Security, privacy, and architecture) lists the components that run locally.
| Component | Description |
|---|---|
| AI agent backend | Tool execution, skill loading, task orchestration |
| Conversation history | Chat messages, threads, metadata |
| Knowledge graph | People, projects, customers, channels, events, and their relationships |
| Memory | Learned facts, procedures, preferences |
| File index | Indexes for keyword search and semantic search |
| Scheduled agents | Agents that run periodically in the background |
| Artifacts and outputs | Generated documents, images, visualizations |
All of these are stored in ~/.quickwork/ (macOS) or %USERPROFILE%\.quickwork\ (Windows).
No Upload to the Cloud
The Desktop section of F1 (FAQ) contains the following.
Your data stays on your device—conversation history, memory, knowledge graph, and file indexes are stored locally and never uploaded to the cloud or used for model training.
However, this statement concerns data stored locally (conversation history, memory, etc.). When making AI inference requests, data is sent to the cloud over HTTPS. We will look at this in detail in the next section.
Cloud Backup of the Knowledge Graph
F1 says "never uploaded to the cloud," but D7 (Knowledge graph) contains the following.
The knowledge graph is stored locally on your machine in the
~/.quickwork/directory. Your knowledge graph data is backed up to your AWS account for cross-device continuity, but is never used for AI model training.
In other words, the knowledge graph is backed up to your AWS account for cross-device continuity. F1's "never uploaded to the cloud" should be read as not including the knowledge graph backup. It is consistent that the data is not used for AI model training, but you need to be aware that knowledge graph data also exists in the cloud.
Governance Trade-offs of the Local-First Design
While the local-first design makes it easier to satisfy privacy requirements in that "customer data is not stored in the cloud," it creates trade-offs from a governance perspective. I could not find in the current official documentation any mechanism for administrators to centrally collect and monitor audit logs of what prompts employees enter and what confidential information they handle. When considering enterprise adoption, I recommend confirming this point with AWS Support or your sales representative.
Summary
| Aspect | Result | Basis |
|---|---|---|
| Architecture | Local-first | D1, D2 |
| Data storage location | Stored locally in ~/.quickwork/ |
D1 |
| Upload to cloud | None for conversation history, memory, and file indexes | F1 |
| Knowledge graph backup | Backed up to AWS account (not used for AI training) | D7 |
| Network communication | Only AI inference (via API Gateway) and connections to external services | D2 |
Finding 3: How AI Inference Is Processed
Even with a local-first design, AI inference is processed in the cloud. I organize what can be read from the official documentation and what is not disclosed.
What Is Officially Disclosed
The Desktop section of F1 (FAQ) contains the following.
AI inference is processed securely through your Amazon Quick account via HTTPS
According to D2, there are two types of network destinations.
- AI models through API Gateway — AI inference requests
- Connected services (Slack, Outlook, Gmail, etc.) — Integration with external services
Regarding the inference Region, D3 (Enterprise setup) contains the following.
All inference for the desktop application also uses this Region. While Amazon Quick on the web can be used in other Regions, the desktop application connects to us-east-1 for both authentication and inference.
This is not limited to the Enterprise edition; it is a constraint of the desktop version as a whole. B3 (What's New) also states "available in preview to all Quick subscribers on MacOS and Windows in all US East (N. Virginia)".
Conceptual Diagram of the Inference Architecture
Based on the information that can be read from the official documentation, I created a conceptual diagram of the inference architecture.
*Solid lines = officially confirmed, dotted lines = partially confirmed or unconfirmed
What Is Not Officially Disclosed
The following information could not be confirmed in the official documentation as of May 2026.
Backend service name — The authentication section of D5 (Settings) mentions "Amazon Federate for Bedrock," but this is a description related to authentication, not a description of the inference backend.
Quick uses Amazon Federate for Bedrock and Quick Web access. You must be signed in to use AI models and access Quick Web features.
Specific LLM model names — D4 describes four model modes (Fast, Balanced, Smart, and Auto), but specific model names are not disclosed. Only Amazon Nova Canvas for image generation (D4, D6) is disclosed.
Whether inference payloads are logged — Not stated in the official documentation.
Summary
| Item | Disclosure status | Details |
|---|---|---|
| Inference communication method | ✅ Disclosed | Via HTTPS (F1) |
| Communication path | ✅ Disclosed | Via API Gateway (D2) |
| Inference Region | ✅ Disclosed | us-east-1 (D3) *applies to the whole desktop version |
| Image generation model | ✅ Disclosed | Amazon Nova Canvas (D4, D6) |
| Backend service name | ❌ Not disclosed | "Amazon Federate for Bedrock" is a description of authentication (D5) |
| LLM model name | ❌ Not disclosed | Only the mode names Fast/Balanced/Smart/Auto (D4) |
| Inference payload logs | ❌ Not disclosed | Not stated officially |
Other Security and Privacy Information
Beyond the main guideline requirements (use for training and data retention), I organize security and privacy information from the official documentation that can inform the decision on business use.
Compliance Certifications
Amazon Quick as a whole has obtained HIPAA Eligible, FedRAMP Authorized, and SOC 2 Audited certifications (F1). A foundation suitable for enterprise use has been put in place.
Encryption in Transit
TLS 1.2 is required and TLS 1.3 is recommended. FIPS 140-3 compliant endpoints are also available (S2). However, S2 is a description that applies to Quick as a whole, and there is no additional description of encryption specifications specific to the desktop version.
Local Security Controls
| Category | Details | Source |
|---|---|---|
| Connection security | OAuth 2.0 authentication. Passwords are not stored. Least privilege | D1 |
| File access control | Controlled by an OS-level sandbox. Only explicitly permitted folders are accessible | D1 |
| Code Execution | Python execution environment isolated by macOS Seatbelt / Windows AppContainer | D6 |
| System tool permissions | Each tool can be individually enabled/disabled. Three permission levels (Full Access / Read Only / Ask Each Time) | D6 |
| Write operations | Write operations such as sending email require explicit approval | F1 |
Privacy Controls
You can enable/disable memory, allow conversation history search, and edit or delete individual memories (D1, D5). You can also delete all data at once from Settings > Danger zone (irreversible, D5).
Conclusion
I conducted an in-depth investigation of the Amazon Quick Desktop data handling policy using the official documentation.
| Requirement | Result | Basis |
|---|---|---|
| Data is not used for training | ✅ Cleared | Consistently denied in four sources: D1, S2, F1, and B2 |
| Data is not retained by the service (persistent data) | ✅ Cleared | Local-first design. Conversation history etc. are not uploaded to the cloud |
| Data is not retained by the service (inference logs) | ⚠️ Not disclosed | Whether inference payloads are logged is not stated officially |
There are also points to keep in mind when considering business use.
- Preview stage: As of May 2026, Amazon Quick Desktop is in preview (D2, B3). Specifications and documentation may change
- Details of the AI inference backend are not disclosed: The backend service name and LLM model names are not officially disclosed. "Amazon Federate for Bedrock" in D5 is a description of authentication and does not confirm the inference backend
- Cloud backup of the knowledge graph: The knowledge graph is backed up to your AWS account for cross-device continuity (D7). Although it is not used for AI model training, you need to be aware that it does not stay only locally
- Governance under a local-first design: No mechanism for administrators to centrally collect and monitor audit logs is confirmed in the official documentation (see Finding 2)
- Inference Region: Because it is fixed to us-east-1, latency and data residency may need to be considered when using it from Japan
- Mixed documentation structure: Because it is integrated on the former QuickSight foundation, you need to pay attention to URLs and sidebar position when looking for desktop version information
Confirming the data handling policy is one of the steps required for business use. When actually starting to use it, please also check the other requirements of the internal guidelines, such as obtaining your supervisor's permission (see Appendix).
Since it is in the preview stage, the documentation is expected to become more comprehensive in the future. Please check the official documentation for the latest information.
*The content of this article is based on the official documentation as of May 2026.
Appendix: Other Items to Check for Business Use
This article investigated the data handling policy (use for training and data retention), but the internal guidelines require the following perspectives as well. I organize them together with considerations specific to Quick Desktop.
| Perspective | Guideline requirement | Considerations for Quick Desktop |
|---|---|---|
| Supervisor's permission | In principle, business use of AI services requires the supervisor's permission | — |
| Highly confidential / personal information | Even if non-use for training is confirmed, avoid entering highly confidential or personal information as much as possible | Because data is sent to the cloud over HTTPS during AI inference, continued caution about the input content is required |
| MCP / external service connections | For usage patterns that transfer code information to the service side, the scope of use must be agreed upon within the team. For highly confidential business processing, conduct an appropriate risk assessment | Quick Desktop can connect to external services such as Slack and Outlook, as well as MCP servers. You need to check where data is sent for each connection destination |
| Code generation features | The scope of use must be agreed upon within the team | Has a Code Execution (Python execution) feature. It runs inside a sandbox (D6) |
