AI Addendum
Customer data protections, provider selection, and operational controls for Stacksync AI services.
Updated November 13, 2026
This AI Addendum supplements the agreement governing Customer’s use of the Stacksync Services (the “Agreement”) with respect to its AI-enabled products and functionality (collectively, the “AI Services”), including AI Inference and AI Genies. Capitalized terms not defined herein have the meanings given in the Agreement. The Data Processing Addendum (DPA) continues to govern the processing of Personal Data and shall prevail in the event of a conflict concerning such processing.
The AI Services include:
- AI Inference: access to AI models through a unified API for Customer’s applications, with model selection, AI Inference API keys, spending limits, and request visibility.
- AI Genies: Stacksync’s AI agents, which use Customer-configured instructions, knowledge sources, and skills to perform tasks through connected applications and communication channels.
The data-use protections below apply to both services. Product-specific controls apply only to the functionality Customer uses.
1. Customer Data and Model Training
“Customer Data” includes the prompts, instructions, files, knowledge sources, conversations, run records, and outputs processed through Customer’s use of the AI Services.
Customer Data is never used to train or fine-tune any model offered on the Stacksync platform, whether by Stacksync or any of its AI providers. This applies to both AI Inference and AI Genies, regardless of the provider’s data retention settings.
A Genie uses Customer’s instructions, knowledge base, and authorized memory to carry out its tasks. This provides context to the Genie at runtime; it does not train the underlying model.
2. Inference Providers and Processing Locations
Customer can choose, per resource, which AI model and provider is used by that resource. This applies to both direct inference requests and Genie runs. Inference is processed through the providers supporting the selected model and configuration. Inclusion on the Subprocessors list does not mean that every listed AI provider processes every Customer’s data.
AI inference is regional, never global. For Genies, inference is performed in the same processing region as the Genie itself. For standalone AI Inference, Customer should confirm the processing locations applicable to its selected model and configuration. Model availability varies by region. Customer should confirm that the selected model and provider meet its residency requirements before submitting data subject to geographic restrictions. The DPA governs applicable subprocessor and international transfer obligations.
Zero data retention is available for certain providers and configurations; it is not enabled universally. Customer may request assistance selecting a configuration appropriate to its retention requirements. When Customer uses its own AI provider API key, Customer’s separately negotiated provider terms apply to its use of that provider.
3. Customer Control and Human Oversight
Workspace administrators may restrict available models in AI Settings. These restrictions apply across all AI Services for the entire workspace, including AI Inference and AI Genies.
AI Inference. Customer selects the model for each request. Customer may create separate AI Inference API keys with spending limits, control access to key resources, inspect request usage, and revoke keys to stop new requests. Customer is responsible for implementing application-level validation, tool permissions, and approval mechanisms in applications it builds using the inference API.
AI Genies. Customer determines the Genie’s instructions, knowledge sources, enabled skills, connected applications, and approval requirements. A Genie may take actions through enabled skills, such as querying information, creating or updating records, or sending messages, subject to the permissions of the configured connections and any required approval. Customer may require approval before a skill executes and may deactivate a Genie to stop new messages and scheduled tasks.
For Genies, access to conversations, execution steps, and files is governed by workspace roles and permissions. Viewers may read conversations and steps; editors and owners may modify a Genie. Raw model prompts are not exposed to workspace users.
AI outputs may contain errors or omissions. Customer remains responsible for evaluating outputs and configuring human review appropriate to the intended use, particularly before consequential decisions or external actions. Instructions and knowledge sources guide responses but do not constitute a guarantee of accuracy, completeness, or freedom from bias.
4. Retention and Memory
Genies memory is segmented in several ways, and per-user segmentation is one of the main segmentation. Users only access their partition of memory which deterministically guarantees privacy and compliance.
Learning from edge cases: A genie operates from what Customer explicitly indicates in its instructions and its knowledge base. Automatic self-improvement memory allows a genie to remember what worked and each person’s preferences, subject to the memory access controls described below.
Memory is built enterprise-first. Never sharing a memory with someone who shouldn’t see it is a strict requirement and our number one priority. Here’s how it works:
- Every piece of memory has an audience: the exact list of people allowed to see it. By default it’s the people who were in the conversation when the memory content was created. For example, if you mention a preference in a private chat with the genie the audience is set to be only you.
- A genie only uses a memory if everyone in the conversation is part of its audience. If a partner or anyone else joins a shared channel, memories they’re not allowed to see are not used.
- This decision is strict and rule-based, never left to AI’s judgement. Memories are filtered before anything is sent to the model, so the model only ever receives a context that has already been cleared for everyone present in a given thread.
- Memory never leaves your workspace. It is not possible to transfer memory across workspaces.
- Full control: admins can review, correct and permanently delete what a genie remembers, per user, and every use of memory is auditable.
Access to conversations and files follows your workspace roles: viewers can read conversations and their steps, and only editors and owners can change a genie. Raw model prompts are never exposed.
5. Security
The AI Services are subject to Stacksync’s documented security policies and secure development practices described in the Security Overview. Stacksync performs regular, at least monthly, internal penetration testing and security scans, including for its AI Services.
Common Questions
Is Customer Data used to train models?
No. Neither Stacksync nor the AI inference providers engaged by Stacksync may use Customer Data to train models offered through the platform. Providing context or authorized memory to a Genie does not train the underlying model.
Can Customer limit which providers its users use?
Yes. Workspace administrators can restrict available models in Workspace settings → AI Settings. These restrictions apply to all AI Services across the entire workspace, including AI Inference and AI Genies.
Are all providers configured for zero data retention?
No. Retention varies by provider and configuration. Stacksync can help identify an appropriate model and configuration.
What controls support reliable outputs?
For Genies, Customer can configure instructions, approved knowledge sources, enabled skills, connection permissions, and approval gates. For AI Inference, Customer implements validation and human review in its own application. These controls support oversight and reliability but do not guarantee correct outputs. Requirements for confidence thresholds, anomaly alerts, or specialized evaluations should be confirmed for the proposed deployment.
Can Customer obtain deployment-specific details?
Yes. Customer may contact hello@stacksync.com to confirm model hosting, provider routing, retention, exports, workspace controls, or security assessment information for Customer’s intended configuration. Any additional contractual commitments must be agreed in writing by the parties.