Guides › AI inside SaaS tools
Slack AI, Notion AI, Atlassian Rovo, Agentforce and Glean: What They Do With Your Data
None of the five lets outside model providers train on your content, and all five answer from content the user is already permitted to see. They differ on where the model runs, what gets copied and whether the vendor improves its own AI with your data. Slack runs its models inside its own infrastructure. Notion, Atlassian, Salesforce and Glean send prompts to outside model providers under zero data retention agreements, and Notion offers zero retention only to Enterprise workspaces. Since August 17, 2026, Atlassian also uses de-identified customer data to improve its own apps and AI, with defaults that depend on your plan. Each product builds an index of your content, and connectors extend that index to other systems.
- Training. None of the five lets third-party model providers train on customer content. Slack, Notion and Glean state that customer data is not used to train large language models. Atlassian now uses de-identified customer data to improve its own apps and AI unless the setting is off.
- Model providers. Slack hosts its models itself. The other four send prompts to third-party providers under zero data retention agreements. On Notion, zero retention applies to Enterprise workspaces, and other plans allow up to 30 days.
- The index. Each product keeps a searchable copy or embedding of your content, and connectors pull other systems into it.
- Permissions. All five answer only from what the user can already see, which makes existing oversharing the main exposure.
Product by Product
Each row paraphrases the vendor's own documentation as of November 2026.
| Product | Training | Where prompts go | Permissions and index | Worth knowing |
|---|---|---|---|---|
| Slack AI | Customer data is never used to train LLMs. | The LLMs run inside Slack's own cloud environment, so model providers have no access to the data. | Answers use only what the user can already see, the same results a standard search would return. | Separate non-generative models for recommendations and search ranking can use customer data unless the organization opts out. Admins can turn AI features on or off. |
| Notion AI | Notion and its AI subprocessors do not use customer data to train models. | Prompts go to AI subprocessors. Enterprise workspaces get zero data retention at the LLM providers. Other plans allow up to 30 days. | Respects page permissions. AI Connectors check permissions at query time. Embeddings sit in a vector database and are deleted within 60 days of a page or workspace deletion. | Disconnecting a connector deletes its data within 24 hours. |
| Atlassian Rovo and Atlassian Intelligence | Third-party LLM providers do not train on customer data. Since August 17, 2026, Atlassian uses de-identified, aggregated metadata and in-app data to improve its own apps and AI where data contribution is on. | Prompts leave your site for third-party LLM providers that operate under zero data retention agreements. | Rovo stores the content and permission settings of files in connected third-party apps and shows only what the user can access. | Data contribution defaults follow your plan. Metadata is on for every plan, and only Enterprise can turn it off. In-app data is on by default for Free and Standard, off by default for Premium and Enterprise, and can be turned off on any plan. |
| Salesforce Agentforce | Customer data is not used to train third-party models. | Third-party LLM providers operate under zero data retention agreements through the Trust Layer. | Grounding retrieves data under the user's existing permissions, and the Trust Layer keeps an audit trail. | Salesforce's Trust Layer documentation marks data masking as not included with agents. Confirm what applies in your org. |
| Glean | No training of LLMs on enterprise data. | Zero LLM data retention. The query and snippets from the results go to the model provider. | Single-tenant deployment, hosted by Glean or in the customer's cloud, with source permissions enforced. | Continuous sensitive content scanning is part of the Protect+ add-on. |
Vendor Improvement Is a Separate Question From Training
A vendor can keep your content away from OpenAI, Anthropic and Google and still use it to improve its own product. Atlassian is the current example. Its data contribution settings took effect on August 17, 2026 and cover two categories. Metadata means content characteristics and common patterns, including frequent phrases from search queries and Rovo Chat prompts. In-app data means the content itself, such as Confluence page text and Jira descriptions and comments. Atlassian states that both are de-identified and aggregated before use. An admin sets this per organization under Atlassian Administration, and the defaults depend on the highest plan the organization holds. Slack draws a similar line between generative AI, which never trains on customer data, and its older recommendation and ranking models, which can unless you opt out.
The Index Is the Part to Review
The training question gets asked first and is the easiest to answer. The harder questions are about the index. To answer from your content, each product stores some form of it: embeddings in a vector database, a search index, or the content and permissions of files from connected apps. That store is a second copy of company data, and it has its own lifecycle. Notion documents that embeddings are deleted within 60 days of the page being deleted, and that a disconnected connector's data is removed within 24 hours. Ask every vendor for the same two numbers.
Connectors widen the index. Turning on a connector from Notion to Slack, or from Rovo to Google Drive, moves content from one vendor's boundary into another's. Each connection deserves the review a new vendor would get. The guides on AI knowledge base security and vector database security cover what to look for.
Permission-Aware Is Only as Good as Your Permissions
Every product here enforces the source system's permissions. That is the right design, and it means the assistant inherits every permission mistake already in place: the public channel with payroll questions in it, the Confluence space nobody restricted, the shared drive that predates the last three reorganizations. Search made those hard to find. An assistant summarizes them on request. Clean up the widest exposures before turning the feature on. The guides on permissions at query time and oversharing describe the method.
Questions to Put to Any SaaS AI Feature
- Which model providers receive prompts?Get the subprocessor list and the retention terms for each one.
- Is zero retention on our plan?Some vendors reserve it for the top tier.
- What does the index store, and where?Raw text, embeddings, permissions, and the vendor that hosts them.
- How fast do deletions and permission changes reach it?Ask for the documented interval.
- Are permissions checked at query time?A check at indexing time alone goes stale.
- What is logged, and can we export it?You need a record of what the assistant retrieved and for whom.
- Is any data used to improve the vendor's own AI?Ask what the default is on your plan and who can change it.
Where Hardshell Sits
Hardshell is the control point for sensitive data that these assistants should use but should not see in the clear. It connects enterprise data to assistants without migrating it, enforces each user's existing permissions at query time, substitutes stand-ins for real identifiers before content reaches the model, and records every retrieval. The assistant still answers the question, and the provider's index and logs hold stand-ins in place of customer, patient or account identifiers. See how Hardshell works.
Frequently Asked Questions
Does Slack AI train on your data?
No. Slack states that customer data such as messages and files is never used to train LLMs, and that its LLMs run inside Slack's own cloud environment where model providers have no access to the data.
Does Notion AI train on your data?
No. Notion states that it and its AI subprocessors do not use customer data to train models. LLM providers retain nothing for Enterprise workspaces and up to 30 days for other plans.
Does Atlassian train AI on Jira and Confluence data?
Since August 17, 2026, Atlassian uses de-identified, aggregated metadata to improve its apps and AI, and in-app content where that setting is on. In-app data is on by default for Free and Standard plans and can be turned off on any plan. Only Enterprise can turn off metadata. The third-party model providers behind Rovo run under zero data retention and do not train on the data.
Is Agentforce data stored by the model provider?
Salesforce states that it has zero data retention agreements with its third-party LLM providers, so prompts and responses are not retained by them or used for training.
Does Glean share your data with LLM providers?
Glean sends the query and snippets from the matching results to the model provider, under zero LLM data retention and with no training on enterprise data. The index itself stays in a single-tenant deployment.
Are AI features inside SaaS tools safe for sensitive data?
The commitments about third-party model providers are strong. The remaining risk is in the index, the connectors, the vendor's own improvement settings and your permissions, because each assistant will surface whatever a user is already allowed to open.
Related guides
- AI connector and MCP security · What to ask before connecting an assistant to a system of record.
- Knowledge features in ChatGPT, Claude, Copilot and Gemini · What each assistant does with files, connectors and memory.
- What AI providers do with your data · The model providers behind these products, compared.
Sources (accessed November 2026)
Slack, AI principles; Privacy principles: search, learning and artificial intelligence. Notion, Notion AI security and privacy practices; Enterprise Search security and privacy practices; Trust Center. Atlassian, AI trust; Data contribution; Rovo data, privacy, and usage guidelines. Salesforce, Agentforce Trust Layer; Agentforce Privacy FAQ. Glean, About Glean Protect and Protect+.