Can it be sold? The rights behind AI training data.
Getting your own data out of a system is rarely the problem: most let the holder export it. The restriction sits on apps that pull the same data through the vendor's API, and on what the data holds about other people. Here is what each system's terms say, and the questions a buyer's counsel asks, each with its source.
Two separate questions. Can the holder export it? Usually. May an app that pulls it through the vendor's API use it to train AI? Increasingly not: 6 of these bar it outright and 2 more bar copying what an app pulls. So the safer route is to export with the system's own tool, run by the holder's own administrator, rather than through a connector.
Business systems: whether the holder can export, and the rights risk of their content
Export is limited. Its API: bars training ai on it. Content risk: high.
Two separate questions. (1) The admin export is a feature of the customer contract; Slack frames full (private and DM) export as a legal and regulatory compliance tool, so a holder should not assume the export button grants a licence-for-AI purpose, and the export contains every employee's and every external Slack Connect participant's messages. (2) Any pipeline that pulls Slack data through the API (including a broker's connector) is barred from LLM training and bulk export by the 2025 API terms. Hunton (28 July 2025): customers who built internal LLMs on Slack archives may need to redesign unless they negotiated limits on unilateral amendments. In practice the safer route is the JSON export run by the workspace owner rather than an API tool, and the content should be treated as personal data of employees and of other organisations (Slack Connect).
Getting your own data out
Free and Pro: public channels only (Free: file links from last 90 days only). Business+: public channels; full export of private channels and DMs only by application approved by the Primary Owner, framed as for legal and regulatory compliance. Enterprise: Org Owners/Admins, Workspace Owners and members with the Export Admin role; full exports require application.
Apps that use its API
Slack API Terms of Service (Effective: October 10, 2025), Data Usage by Third Parties: you may not "(A) use API Data to train a large language model;" or "(B) bulk export Slack message and file data except where expressly allowed by an additional agreement". Data Access API and Real-Time Search API: "you may not create persistent copies, archives, indexes, or long-term data stores of other organizations' API Data."
“Salesforce recently modified its Slack API Terms of Service (the Terms) to prohibit (i) bulk exporting of data accessible through the Slack application programming interfaces (APIs), (ii) the creation of persistent copies, archives, indexes or long-term data stores, and (iii) the usage of such data in large language models (LLMs).”
Microsoft 365 / Teams / Exchange / SharePoint
The holder can export it. Its API: bars copying or reselling it. Content risk: medium.
The customer's ownership and extraction right are strong, so the blocker is not Microsoft's contract but what is inside the mailboxes and chats (third-party and employee personal data, counterparties' confidential material). The API terms restrict a third-party app (for example a broker connector) from building a copy of the data beyond its intended scenario and from reselling data obtained through the API. The May 2026 Product Terms clause targets bot or RPA outputs of the Office apps used for AI training; it does not on its face bar a customer from licensing its own documents. Prefer customer-run eDiscovery or admin export over a third-party Graph app.
Getting your own data out
Customer can extract Customer Data during the subscription and for 90 days after it ends (DPA). Bulk mailbox and Teams chat export for an organisation is usually done through Microsoft Purview eDiscovery; Microsoft Learn says premium eDiscovery features require Office 365 E5 or Microsoft 365 E5 (or E5 add-ons); a vendor guide says standard eDiscovery content search and export is in E3 and above (not verified against Microsoft licensing docs). Teams export APIs were protected and metered; Microsoft says they are no longer metered as of 25 August 2025.
Apps that use its API
Microsoft APIs Terms of Use (Last Updated October 2025, covers Microsoft Graph, Teams, Exchange, SharePoint, OneDrive APIs): you may NOT "Scrape, build databases or otherwise create copies of any data accessed or obtained using the Microsoft APIs, except as necessary to enable an intended usage scenario for your Application" and may not "Redistribute or resell, or sublicense access to, the Microsoft APIs, any data obtained using the Microsoft APIs". Email protocols and APIs may only be used for "syncing" or "backing up" mail, calendar and contacts unless customers expressly grant other permissions. Product Terms change (1 May 2026, Office and Microsoft 365 Applications): "outputs generated through Robotic Process Automation, bots, or other similar technologies may not be used to develop, train, evaluate, improve AI or machine learning models, or to replicate product functionality."
“Redistribute or resell, or sublicense access to, the Microsoft APIs, any data obtained using the Microsoft APIs, or any other Microsoft Offering accessed through the Microsoft APIs”
“outputs generated through Robotic Process Automation, bots, or other similar technologies may not be used to develop, train, evaluate, improve AI or machine learning models, or to replicate product functionality.”
“Starting August 25, 2025, the Teams APIs are no longer metered, and no billing configuration is required to use these APIs.”
Google Workspace (Gmail, Drive, Docs, Chat)
The holder can export it. Its API: bars training ai on it. Content risk: medium.
Admin export of the organisation's own data is clearly allowed. The hard restriction is on apps that obtain Workspace data via Google APIs (OAuth apps, including any broker or labelling connector): they may not use it to train generalised AI models or transfer it to data brokers. So the holder should export with the native Data Export tool or Vault and transfer the files under its own licence, rather than authorise a third-party app to pull the data.
Getting your own data out
Full organisation export: all editions, by a super admin (30-day-old account, 2SV). Export for specific users or filtered exports: "If available"; advanced editions such as Enterprise Plus or Education Plus use a separate "Choose the Workspace data you want to export" flow. Organisations over 1,000 users or with FedRAMP authorisation must contact support first. Google Vault is the eDiscovery export route on editions that include it.
Apps that use its API
Google Workspace API User Data and Developer Policy (last updated 2026-09-03), prohibited: "Transferring, selling, or using user data to create, train, or improve a machine learning or artificial intelligence model beyond that specific user's personalized model for the appropriate use case or user-facing feature." Also prohibited: "Transferring or selling user data to third parties like advertising platforms, data brokers, or any information resellers."
“Transferring, selling, or using user data to create, train, or improve a machine learning or artificial intelligence model beyond that specific user's personalized model for the appropriate use case or user-facing feature.”
“These steps initiate a full export of your organization's data and are available in all Google Workspace editions.”
Atlassian (Jira, Confluence, Jira Service Management)
The holder can export it. Its API: not established. Content risk: low.
Export of own data is permitted and documented. Two points for the holder: (1) Jira and Confluence hold customer tickets (JSM), contractor and partner content and personal data, so content review is the gate, not Atlassian's terms. (2) Since 17 August 2026 Atlassian itself uses de-identified, aggregated in-app content (on Free and Standard by default) to improve its AI features, so a buyer asking for exclusivity should be told that the platform vendor may already have derived value from the same content; it is not an exclusive grant Atlassian holds over the data, but it is worth disclosing. Atlassian says data is de-identified and aggregated before use.
Getting your own data out
Jira Cloud: Backup manager (Administration, System, Import and Export) produces a site backup with issues, users and optional attachments; one backup at a time, regenerable every 24 hours (older Atlassian docs). Confluence: space and site exports. Data contribution defaults (metadata / in-app data): Free On/On, Standard On/On, Premium On/Off, Enterprise On/Off. Only an organization with an active Enterprise plan can opt out of metadata contribution ("If your Atlassian organization has an active Enterprise plan, you can opt-out of metadata contribution"), and "every customer has the option to opt out of contributing in-app data" (read 16 Oct 2026).
Apps that use its API
No clause found in the Atlassian Customer Agreement barring a customer from using its own data to train AI; section 2.2 bars using "the Products to develop a similar or competing product or service". Atlassian Developer Terms were not retrievable (404) when read, so restrictions on Marketplace or API apps are unverified. Separately, Atlassian's own use: "Changes to how Atlassian will use metadata and in-app data to improve apps and experiences for all customers went into effect on August 17, 2026."
“If your Atlassian organization has an active Enterprise plan, you can opt-out of metadata contribution. If your Atlassian organization’s highest active plan is Free, Standard, or Premium, metadata contribution is always on, and you’re not able to opt-out. ... every customer has the option to opt out of contributing in-app data.”
“To view your settings, navigate to Atlassian Administration, then Security, and then Data contribution. ... Atlassian will begin to use data according to the data contribution settings on August 17, 2026.”
GitHub
The holder can export it. Its API: no ai clause found. Content risk: low.
GitHub's terms are not the blocker; the content is. A repository mixes the holder's own code (sellable only if IP assignment from employees and contractors is clean) with vendored third-party and open-source code whose licences travel with it, plus issues, pull request comments and commit metadata containing contributors' names and emails (personal data). Public repositories already carry a licence to other users under D.5. Since 24 to 27 April 2026, GitHub may use Copilot inputs and outputs on Free, Pro and Pro+ to train AI models unless the user opts out (J.3), and this does not apply where a GitHub Customer Agreement governs; disclose this to a buyer seeking exclusivity over code written with Copilot. Commentary (Heise, March 2026) reports private repository content at rest is not used for training.
Getting your own data out
Any owner can clone repositories. Organisation owners can run migration exports. No plan gate on getting your own code out.
Apps that use its API
No clause found in the GitHub Terms of Service or Acceptable Use Policies barring a repository owner from licensing its own code for AI training. The API terms (section H) restrict abuse: "Abuse or excessively frequent requests to GitHub via the API may result in the temporary or permanent suspension" and "You may not use the API to download data or Content from GitHub for spamming purposes." Acceptable Use, section 7: "Researchers may use public, non-personal information from the Service for research purposes, only if any publications resulting from that research are open access." Scraping "does not refer to the collection of information through our API."
“a license to collect and use your Inputs and Outputs to develop, train and improve artificial intelligence ... unless (a) you opt out through your account settings”
“Researchers may use public, non-personal information from the Service for research purposes, only if any publications resulting from that research are open access.”
“GitHub is changing its terms of use and will use user interactions with Copilot for its AI model training starting April 24, 2026. ... The change affects the Free, Pro, and Pro+ plans, but not Business and Enterprise.”
GitLab
The holder can export it. Its API: no ai clause found. Content risk: low.
Like GitHub: the blocker is open-source licences, contributor personal data and contractor IP assignment, not GitLab's terms. Self-managed GitLab instances are the cleanest source because no SaaS vendor terms sit between holder and data. GitLab's AI Transparency Center states (per a search excerpt, not contract language) that inputs to and outputs from GitLab Duo are not currently used to train GitLab models.
Getting your own data out
Self-managed: the customer holds the data. GitLab.com and Dedicated: project and group exports by maintainers and owners; the customer must migrate content off before termination.
Apps that use its API
No clause found restricting a customer's use of its own content for AI training. Use restrictions target the software: customer may not "disassemble, reverse engineer, or decompile the Software or access it to: (a) build a competitive product or service". Section 12 (U.S. Government Matters): "Customer shall not export, reexport, or transfer the Controlled Subject Matter except as authorized by these laws and regulations."
“both Input and Output constitute additional Customer Content as that term is defined in the Subscription Agreement. Customer will retain all ownership of Customer Content, as permitted under applicable law.”
Notion
The holder can export it. Its API: not established. Content risk: low.
Clean contractually. Watch two things: a full export by one admin can miss other users' private pages, so the holder must state what is and is not in the export; and wikis often embed customer documents, contracts and HR pages. The April 2026 MSA update also notes that enabling integrations (APIs, MCP connectors) may require sharing Customer Data with third parties, so a broker connector is a disclosure the holder authorises.
Getting your own data out
Help Center: "You can export all your pages as HTML, Markdown, or CSV (for databases)" from Settings, General; "Pages that the exporter doesn't have access to, such as private pages of other users, will not be included in the export." Workspace owners can disable export (Settings, Security). Admin API for workspace-level exports is in beta on Enterprise. Download link "will expire after 7 days."
Apps that use its API
No clause in the MSA bars the customer from using its own exported data to train AI. Use restriction 2.1(8): customer will not "access the Services for the purpose of building a competitive product or service or copying its features or user interface". Notion's Developer Terms were not reviewed when read (unverified).
“At any time during the Subscription Term, Customer may export Customer Data from the Services using Notion's self-service export functionality in an industry-standard format then-supported by the Services”
“Customers continue to own their Customer Data. The updated language makes that clearer and confirms Customer Data is treated as Customer Confidential Information under the MSA.”
Salesforce (Sales Cloud, Service Cloud, platform)
The holder can export it. Its API: no ai clause found. Content risk: high.
The contract lets the customer take its data out. What blocks a sale is what the CRM holds: customers' and prospects' personal data (names, emails, call notes), customer confidential deal terms, and data bought from list vendors whose licence usually forbids resale. Treat CRM records as presumptively non-sellable except aggregated or properly de-identified, and check any third-party enrichment data (ZoomInfo and similar) licence before including it.
Getting your own data out
Data Export Service in Setup: weekly exports on Enterprise, Performance and Unlimited; every 29 days on Professional and Developer (Salesforce Trailhead); download links reportedly expire after 48 hours (third-party guides). Data Loader and Bulk API for full object exports.
Apps that use its API
The MSA has no clause barring a customer from using its own Customer Data to train AI. Customer must use Services in accordance with "the Acceptable Use and External Facing Services Policy and the Artificial Intelligence Acceptable Use Policy" and may not "access it to (1) build a competitive product or service" or "use any Services to access, copy or use any of SFDC intellectual property except as permitted". The restrictive API-level AI terms Salesforce adopted in 2025 apply to Slack (see slack), not found for core CRM.
“Upon request by Customer made within 30 days after the effective date of termination or expiration of this Agreement, SFDC will make Customer Data available to Customer for export or download as provided in the Documentation. After such 30-day period, SFDC will have no obligation to maintain or provide any Customer Data”
“It allows you to export data manually once every 7 days (for weekly export) or 29 days (for monthly export). You can also export data automatically at weekly or monthly intervals. Weekly exports are available in Enterprise, Performance, and Unlimited Editions. In Professional Edition and Developer Edition, you can generate backup files only every 29 days, or automatically at monthly intervals only.”
HubSpot
The holder can export it. Its API: bars training ai on it. Content risk: high.
Export by the account owner is fine. Any app or connector that pulls HubSpot data through the API to build a multi-customer training set (a broker's pipeline, a data marketplace connector) falls within the May 2026 restriction. Content is mostly contacts' personal data and marketing email history, so the privacy layer is the real blocker. HubSpot itself trains on Customer Data unless opted out.
Getting your own data out
Record exports from the CRM by users with export permission; Super Admins see an audit log of all exports and can set large-export alerts (up to 275,000 records). Plan-level detail for a full account export not verified.
Apps that use its API
HubSpot Developer Terms update (announced and effective May 4, 2026) restricts "using data accessed through the APIs to train, fine-tune, or improve AI or machine learning models" "with a clear carve-out for legitimate single-customer use cases"; the Technology Partner Program Agreement will "explicitly exempt all technology partners in good standing from the new restriction". HubSpot's own use (Customer Terms 5.3.1): "We may also use Customer Data to train our AI models in compliance with our obligations under the Agreement." "You may opt out of having your Customer Data used for this purpose by updating your settings in your HubSpot account."
“We may also use Customer Data to train our AI models in compliance with our obligations under the Agreement.”
Zendesk
Export is limited. Its API: bars training ai on it. Content risk: high.
Its Developer Terms are among the most restrictive we read. A connector or script that pulls tickets through the API to build a training set, or to sanitise them for that purpose, appears to fall within clause 3.6; that would include a Team-plan customer whose only route out is the API. The native export, enabled by the account owner, is the route least exposed to those terms. Content is customer conversations: the end customers' personal data and often the holder's own customers' confidential information, usually covered by a DPA. Treat it as hard to sell unless properly de-identified and permitted by customer contracts.
Getting your own data out
Native JSON, CSV or XML export after the account owner asks Support to enable it; exports can be restricted to certain admins by email domain. AI agent tickets cannot be exported. Team plans: REST API only, which is bound by the Developer Terms above.
Apps that use its API
Zendesk Developer Terms (Effective Date: August 12, 2025), 3.6(iii): developer must "not use API Data for any purpose other than to provide its Application functionality", "including creating, training, or improving any general AI model, such as, but not limited to, a large language model", "or sanitizing API Data for such purposes;" 3.6(iv): "not directly or indirectly use the APIs to develop Applications that allow bulk export of API Data;" 3.6(v): "not create persistent copies, archives, indexes, or long-term data stores using API Data". 3.3(iv): may not "repackage or resell the Services, APIs, or API Data;"
“To protect the data in your Zendesk account, data exports are not enabled by default in Zendesk accounts.”
ServiceNow
The holder can export it. Its API: no ai clause found. Content risk: medium.
Contract allows retrieval. Content is IT and HR tickets: employee personal data, often sensitive HR cases (health, grievances, investigations), security incident details and vendor confidential data. HR Service Delivery and security incident tables should be excluded by default. Disclose whether the customer was opted in to Now Assist data sharing.
Getting your own data out
During the subscription: instance admins export tables (XML, CSV, Excel) and use the REST/Table API. After termination: 45-day window to request a database export. Now Assist data sharing opt-out is per instance from the Now Assist Admin console and needs a data steward role (search excerpt; reportedly opted in by default and not available for GCC or self-hosted instances).
Apps that use its API
No clause found restricting a customer's use of its own Customer Data for AI training; restrictions in the agreement target the ServiceNow Core Technology (for example no use in "competition with ServiceNow Core Technology", no reverse engineering). ServiceNow's own use: "ServiceNow uses customer data sharing to improve its Now LLMs continuously." "Customers have the option to opt out of data sharing to prevent their usage data from contributing to model improvements."
“ServiceNow will provide any Customer Data in the Subscription Service to Customer in ServiceNow's standard database export format at no additional charge.”
“You must opt out of data sharing to turn off ServiceNow ability to gather Customer data for AI and data products.”
Intercom
Unclear from its terms. Its API: bars copying or reselling it. Content risk: high.
The developer terms bar a third-party app from scraping or copying data obtained through the Intercom Service, which catches a broker connector even without an explicit AI clause. Content is the holder's customers' support chats: their personal data and the holder's customers' confidential information, usually under a DPA. Treat as hard to sell unless properly de-identified and permitted by customer contracts.
Getting your own data out
Unverified.
Apps that use its API
Intercom Developer Terms (no date shown): no AI-training clause found. 3.2.1(g): "You will not scrape, or otherwise create copies of any data obtained using the Intercom Service"; 1.5.2: "Developer shall not collect or store any Developer Data without end-user's authorization." Intercom contractually restricts "Third Party LLM Providers from using Customer Data for training or otherwise improving Third Party LLM Provider's services" (Additional Product Terms, from a search excerpt).
“Intercom contractually restricts Third Party AI Providers from using Customer Data for training or otherwise improving Third Party AI Provider's services.”
Zoom (meetings, recordings, transcripts, chat)
The holder can export it. Its API: bars training ai on it. Content risk: high.
Zoom does not stop the account owner downloading recordings, but any API-based pipeline (including a broker's or labeller's app) is barred from using the content for multi-customer AI training without Zoom's written permission. The heavier blockers are outside Zoom: every participant's voice and face (BIPA voiceprint and face-geometry exposure for Illinois participants), all-party consent wiretap laws for recordings (California, Illinois, Pennsylvania, Washington, Florida and others), and external participants (clients, candidates, patients) who never agreed to anything beyond the meeting. Treat recordings as audio-speech and video with the highest consent bar.
Getting your own data out
Cloud recordings and transcripts are downloadable by hosts and account admins from the web portal; local recordings sit on hosts' computers (plan-level detail not verified).
Apps that use its API
Zoom API License and Terms of Use (Last Updated: July 16, 2025): "You may not use any Customer Content (as defined in the Zoom Terms of Service) accessed or obtained using the Zoom APIs or your Application to train, develop, evaluate, test, improve, or modify artificial intelligence or machine learning models (including language models) ("AI Training"), nor allow such data to be used for those purposes, except with Zoom's express written permission. Notwithstanding the foregoing, if you obtain explicit consent from a Zoom Account Owner to use their Customer Content for AI Training, you may use such Customer Content solely for AI Training specifically for their use. For clarity, the resulting customized model may only be used by the Customer who consented to the use of their Customer Content and not by any other party."
“You may not use any Customer Content (as defined in the Zoom Terms of Service) accessed or obtained using the Zoom APIs or your Application to train, develop, evaluate, test, improve, or modify artificial intelligence or machine learning models (including language models)”
“You retain all ownership rights in your Customer Content subject to any license or other rights granted herein.”
QuickBooks Online / Intuit
The holder can export it. Its API: not established. Content risk: medium.
Ambiguous clause: the prohibition is on training AI "using information derived from the Platform without authorization". It is unclear whether a customer's own exported books count as "information derived from the Platform"; the safer reading is that exported ledger data is the customer's Content, but derived reports, Intuit categorisation and benchmark fields may be Platform-derived. Accounting data also contains customers' and vendors' names, bank details and employee payroll (sensitive), so heavy de-identification is needed. Have a lawyer read the clause before selling QuickBooks-derived data.
Getting your own data out
Export to Excel (Settings, Tools, Export Data) during the subscription and for one year read-only after cancellation for paid plans (90 days for trials), per Intuit help and community answers.
Apps that use its API
Intuit Terms of Service, Prohibited Uses: "Engage in unauthorized access, monitoring, interference with, or use of the Platform or third party accounts, information (including personal information), computers, systems or networks, including scraping, accessing, or downloading content that doesn't belong to you, or training or developing artificial intelligence or machine learning models or related technology using information derived from the Platform without authorization;". The Intuit Developer Terms (QuickBooks API) could not be read (client-rendered page); their AI clause is unverified.
“or training or developing artificial intelligence or machine learning models or related technology using information derived from the Platform without authorization;”
“your data will be available in read-only access for 1 year from the date of cancellation”
Xero
Export is limited. Its API: bars training ai on it. Content risk: high.
Of the accounting platforms we read, Xero's terms are the most restrictive: data pulled through the Xero API may not be used to train or fine-tune any AI model at all, with no single-customer carve-out stated in the quoted text. A holder should use native exports rather than an API connector, and should still expect a buyer's counsel to ask how the data left Xero. Accounting data carries counterparties' names, bank and payroll data.
Getting your own data out
Unverified. Terms apply immediately to developers registered on or after 4 December 2025 and from 2 March 2026 to earlier registrants.
Apps that use its API
Xero Developer Platform Terms and Conditions (Last updated: 4 December 2025): "You must not use API Data sourced from the Xero developer platform to train or fine tune any artificial intelligence models"; clause 7(b)(i): "API Data cannot be used to train, fine tune, adapt, or enhance any AI Models." 7(b)(ii): "API Data shall not be passed to another third party without consent from the associated user." Clause 29 bars developers from: "Sell any user data (but you can charge users a fee for your app)."
“apps must not use bots or browser extensions to undermine our security controls or simulate user actions.”
Oracle NetSuite
Unclear from its terms. Its API: not established. Content risk: medium.
Treat as unverified. ERP data includes customers, vendors, pricing, payroll and employee records, plus counterparties' confidential terms. The EU Data Act addendum shows Oracle documents structured export for switching, which supports the view that the customer can get its own data out; whether NetSuite terms restrict AI use was not established.
Getting your own data out
Saved searches and CSV export, SuiteTalk and REST APIs during the subscription; post-termination access reportedly ends unless a paid read-only arrangement is bought (consultancy blog, unverified).
Apps that use its API
Not found. The current NetSuite terms of service and SuiteTalk/REST API terms could not be retrieved when read.
“The available self-service tools and processes to retrieve exportable data in a structured, machine-readable format from an eligible Cloud Service are described in its applicable User Guides”
Figma
The holder can export it. Its API: no ai clause found. Content risk: low.
Design files are usually the company's own work product and low personal-data risk, but often embed client brand assets, licensed fonts, stock imagery, and third-party UI kits whose licences forbid redistribution. If the team was on Starter or Professional with content training on, Figma may already have trained on the files: disclose to a buyer seeking exclusivity.
Getting your own data out
Files can be saved locally as .fig and assets exported (PNG, SVG, PDF) by editors; bulk organisation export not verified.
Apps that use its API
Figma Developer Terms (Effective Date May 5, 2026): no clause found on using API or MCP data to train AI. Section 8(a): "Figma owns all right, title, and interest in the Developer Resources, the Services, and any data derived from either". Figma's own use: content training is an admin setting; per Figma help (via search) Starter and Professional teams default to content training on and Organization and Enterprise default to off; training began 15 August 2024.
“As between Customer and Figma, Customer retains all right, title, and interest in Customer Content, and Figma owns all right, title, and interest in the Services.”
“content training does not affect access to AI features”
Linear
The holder can export it. Its API: no ai clause found. Content risk: low.
Contractually clean. Content is engineering tickets: product roadmap, bug reports that may quote customers, and employee names. The contract has no post-termination retrieval window, so export before cancelling. Linear owns aggregated and anonymised Service Data, which is not part of what the customer can license.
Getting your own data out
Not verified (Linear offers CSV export of issues and a public GraphQL API).
Apps that use its API
No clause found barring a customer from using its own User Submissions for AI training. Restriction (g): customer may not "use or access the Service to build or support and/or assist a third party in building or supporting products or services competitive to the Service". AI features are governed by a separate AI Services Addendum (not read).
“Customer will own all right, title and interest in and to the User Submissions and there are no implied licenses under this Agreement.”
The questions a buyer's counsel asks
Each with what a good answer and a red flag look like, and the kinds of data it applies to. The rights check walks you through the ones that apply to what you hold.
What blocks a sale (16)
An open answer here stops a buyer's legal review. Settle each before any buyer sees the datasheet.
Does the company selling the data own it, and has someone with authority approved the sale?
A buyer's first warranty request is that the licensor owns or controls the rights granted, and a sale signed by the wrong person is void or challengeable.
A good answer
The legal entity that collected the data is the licensor; the board has approved the licence in writing; if the data is all or substantially all of the company's assets, shareholders have approved too.
A red flag
A founder selling personally after a shutdown, a former employee holding an export, or no board resolution.
“Every corporation may at any meeting of its board of directors or governing body sell, lease or exchange all or substantially all of its property and assets ... when and as authorized by a resolution adopted by the holders of a majority of the outstanding stock of the corporation entitled to vote thereon”
How did the data leave each software system: a built-in export run by your own admin, or an app or script using the vendor's API?
Several vendors now forbid using API-obtained data for AI training or bulk export, so data pulled the wrong way can be tainted even if the company owns it.
A good answer
Each system was exported by its own admin with the vendor's native export tool, the method is recorded per system, and no third-party connector touched it.
A red flag
Data pulled through a connector or scraper from Slack, Zendesk, Xero, Zoom, HubSpot or Google Workspace APIs; Slack DMs exported without the Primary Owner's approved application.
“API Data cannot be used to train, fine tune, adapt, or enhance any AI Models.”
Does any of the data belong to or describe your customers, and do your customer contracts and data processing agreements allow using it for anything other than serving them?
Where you hold customer data as their processor you may use it only on their instructions, and a processor that decides its own purposes becomes a controller in breach.
A good answer
Customer content (support tickets, customer files, CRM notes, customer email) is excluded, or each customer contract expressly permits this use, or the data is anonymised to the standard in the anonymisation item and the contract does not forbid that.
A red flag
Support tickets, customer success call recordings or customer-uploaded files in the dataset; contracts that restrict use to providing the service.
“if a processor infringes this Regulation by determining the purposes and means of processing, the processor shall be considered to be a controller in respect of that processing.”
“Review customer, partner, vendor, and data license agreements, identify data held as a service provider or processor, and review platform terms for the systems the data will be extracted from.”
“training data can be "deidentified" and still contain confidential or valuable information.”
Does the data contain other companies' confidential information, such as partner documents, pricing, roadmaps or anything shared under an NDA?
NDAs usually limit use to a stated purpose and survive termination, so passing counterparties' material to an AI lab invites breach of contract and trade secret claims.
A good answer
Shared channels, external email threads, partner folders and attachments from third parties are excluded or reviewed, and you can list which counterparties appear.
A red flag
Slack Connect channels, vendor negotiation threads, M&A data rooms, or client deliverables in the archive.
“Consider privileged material, HR and personnel files, payroll and tax records, health and benefits information, union-related communications, customer and partner confidential information, licensed data, and data from regulated systems.”
Does the data include personal information of employees, contractors or job applicants in California, and have they been told about the sale and given a way to opt out?
The CCPA exemption for employee data ended on 1 January 2023, so licensing staff data for money is a CCPA sale with notice and opt-out duties.
A good answer
Either employee personal information is removed to the CCPA deidentified standard, or current and former staff received notice and an opt-out before the sale.
A red flag
Slack, email or HR records with staff names and messages sold as is; no notice to former employees.
“"Sell," "selling," "sale," or "sold," means selling, renting, releasing, disclosing, disseminating, making available, transferring, or otherwise communicating ... a consumer's personal information by the business to a third party for monetary or other valuable consideration.”
For staff in the EU or UK, what legal basis lets you reuse the personal data in their work (messages, documents, code history, recordings) for a sale, and is that new purpose compatible with why you collected it?
Employee consent is rarely valid because of the power imbalance, and a new purpose must pass the GDPR compatibility test or have its own basis.
A good answer
The data is anonymised to the EU standard before it leaves (only truly anonymous data is outside GDPR; de-identified or pseudonymised data stays personal data), with a documented legitimate-interests and compatibility assessment covering the anonymisation step itself, and notice to staff.
A red flag
Relying on a consent tick-box from employees, or treating name-replaced (pseudonymised) chat logs as outside GDPR.
“Given the dependency that results from the employer/employee relationship, it is unlikely that the data subject is able to deny his/her employer consent to data processing without experiencing the fear or real risk of detrimental effects as a result of a refusal.”
“the controller shall, in order to ascertain whether processing for another purpose is compatible with the purpose for which the personal data are initially collected, take into account, inter alia: any link between the purposes ... the context in which the personal data have been collected, in particular regarding the relationship between data subjects and the controller”
For customers or staff in Canada, has each person consented to this new use, or will the data be anonymised to the Canadian standard before it leaves? For Quebec, is the privacy impact assessment done and the written agreement ready?
Canada's federal privacy law has no legitimate interest ground: a new use or disclosure needs the person's consent, and only data with no serious possibility of re-identification is outside it. Quebec adds an impact assessment and a written agreement before personal information leaves the province.
A good answer
The data is anonymised to the regulator's standard (no serious possibility of re-identification, alone or with other available information) with the test written down, or each person consented to this licence; for Quebec, the impact assessment is done and the agreement with the buyer reflects it.
A red flag
Names, phone numbers and emails stripped while the free text still points to people (the regulator found that alone insufficient in 2026), or relying on the business transaction exception, which excludes a deal whose purpose is selling personal information.
“When personal information that has been collected is to be used for a purpose not previously identified, the new purpose shall be identified prior to use. Unless the new purpose is required by law, the consent of the individual is required before information can be used for that purpose.”
“The term "made anonymous" is not defined in PIPEDA. ... We therefore find that for information to be considered anonymized for the purpose of Principle 4.5.3, the organization must take steps to ensure that there is no serious possibility that the information retained may be reidentified, either alone or in combination with other available information. ... Based on the data points that Loblaw described, we find that stripping accounts of names, phone numbers and email addresses alone is insufficient for Loblaw to demonstrate that the data it retains is anonymized.”
“Before communicating personal information outside Québec, a person carrying on an enterprise must conduct a privacy impact assessment. ... The information may be communicated if the assessment establishes that it would receive adequate protection, in particular in light of generally recognized principles regarding the protection of personal information. The communication of the information must be the subject of a written agreement that takes into account, in particular, the results of the assessment and, if applicable, the terms agreed on to mitigate the risks identified in the assessment.”
Has personal data been removed to the EU anonymisation standard, and who tested it?
Under the EDPB's draft 2026 guidelines data is anonymous only if no one can be singled out, linked or inferred, and the anonymisation process itself needs a legal basis; pseudonymised data stays personal data.
A good answer
An independent de-identification provider ran a documented test against the three criteria, certified the result before transfer, and the buyer commits not to re-identify and flows that down (the Spirit Airlines ombudsman model).
A red flag
Names swapped for tokens in workplace chat, small teams where writing style or role identifies people, no written test, or you keep the key.
“the process of anonymisation itself remains subject to the GDPR and must comply with applicable legal bases, transparency obligations and accountability requirements.”
“Spirit has hired Tonic.ai to deidentify that unstructured data, according to the report. The agent must file a certification with the court before every transfer. ... The buyer must publicly commit not to try to reidentify the data. It also may not pass it on unless the recipient makes the same commitment. ... The potential privacy losses or costs to consumers have been eliminated or mitigated.”
Does any data come from health care, health insurance, or work you did for a health client under a business associate agreement?
Protected health information can be sold only if de-identified by HIPAA safe harbor (18 identifier types removed, no actual knowledge) or expert determination (very small risk), and a business associate agreement may forbid it entirely.
A good answer
Expert determination by a qualified statistician with written methods, or safe harbor with all 18 identifiers removed, plus confirmation the BAA permits de-identified use; data use agreement banning re-identification.
A red flag
Clinical notes or patient support chats with free text; data held as a business associate; voice or face recordings of patients (biometric identifiers are on the safe harbor list).
“Health information that does not identify an individual and with respect to which there is no reasonable basis to believe that the information can be used to identify an individual is not individually identifiable health information.”
“the BAA likely governs as a result of the typical conflict language in the BAA favoring the BAA when PHI is at issue.”
Does the data contain voices, faces, fingerprints or other biometrics of people who may be in Illinois, Texas or Washington?
Illinois BIPA forbids anyone holding biometric identifiers (voiceprints, face geometry) from selling or profiting from them, with $1,000 to $5,000 statutory damages per violation, and nine voice-training class actions were filed in May 2026.
A good answer
Recordings of your own consenting contributors with written releases naming AI training and sale; no biometric templates extracted by you; Illinois residents excluded or separately released; buyer warrants it will not extract identifiers without its own consent basis.
A red flag
Call-centre recordings, meeting videos, CCTV or customer selfies; any dataset marketed for voice cloning or face recognition.
Blocks a sale until settled. Applies to: audio and speech, video and footage, images, physical-world data. Reasoned, not stated by a source: Whether raw audio or video (as opposed to an extracted voiceprint or face template) is a 'biometric identifier' is contested in the pending May 2026 cases; treat voice and face datasets as exposed until courts rule.
“(c) No private entity in possession of a biometric identifier or biometric information may sell, lease, trade, or otherwise profit from a person's or a customer's biometric identifier or biometric information.”
“"Biometric identifier" means a retina or iris scan, fingerprint, voiceprint, or scan of hand or face geometry.”
Were calls and meetings recorded with every participant's consent, and did that consent cover use beyond the meeting?
California and other all-party-consent states make recording a confidential conversation without everyone's consent a crime, and Zoom's API terms forbid multi-customer AI training on meeting content.
A good answer
Recording notices on every call, written consent from internal participants for AI licensing, external participants (clients, candidates) excluded or separately consented.
A red flag
Sales or support call libraries recorded under 'for quality and training purposes' notices only; meetings with external guests.
“intentionally and without the consent of all parties to a confidential communication, uses an electronic amplifying or recording device to eavesdrop upon or record the confidential communication”
“the resulting customized model may only be used by the Customer who consented to the use of their Customer Content and not by any other party.”
Does the data include text, images, video or files created by others that you do not own (articles, reports, books, customer uploads, screenshots of other products)?
In September 2026 the Third Circuit held that copying Westlaw headnotes to train a competing AI tool was not fair use, so third-party content inside an archive carries real infringement risk for the buyer and the holder's warranty.
A good answer
Third-party content is identified and removed or licensed; what remains is the company's own original work.
A red flag
Shared drives full of downloaded PDFs and reports, wiki pages pasted from vendors' docs, image folders scraped from the web.
“That the use of the data for training as specified in the agreement does not infringe a third party's copyright”
What did your privacy policy, terms of service and employee handbook say about sharing or selling data at the time it was collected?
The FTC's position is that a company cannot quietly change its terms to allow AI training on data collected under earlier promises, and in bankruptcy a policy against transfer blocks a sale without court approval.
A good answer
The policies in force when the data was collected allowed this kind of transfer, or affected people gave fresh affirmative consent, or the data is fully anonymised.
A red flag
'We never sell your data' in an old privacy policy; a policy changed recently to add AI training; no record of past versions.
“A business that collects user data based on one set of privacy commitments cannot then unilaterally renege ... It may be unfair or deceptive for a company to adopt more permissive data practices ... through a surreptitious, retroactive amendment to its terms of service or privacy policy.”
“if the debtor in connection with offering a product or a service discloses to an individual a policy prohibiting the transfer of personally identifiable information about individuals to persons that are not affiliated with the debtor ... then the trustee may not sell or lease personally identifiable information to any person unless– (A) such sale or such lease is consistent with such policy; or (B) after appointment of a consumer privacy ombudsman”
If the company is closing, insolvent or in bankruptcy, who now has legal authority to sell: the board, an assignee for the benefit of creditors, a bankruptcy trustee, or a Dutch curator?
Once a company is in formal insolvency the founders lose authority: in US bankruptcy a privacy-policy conflict requires a consumer privacy ombudsman and court approval, and a Dutch curator needs the supervisory judge's permission for a private sale.
A good answer
The right officeholder signs; in US bankruptcy the sale goes through section 363 with an ombudsman where personal data is involved (Spirit Airlines to Google, October 2026); in a Dutch faillissement the curator has rechter-commissaris consent.
A red flag
A founder selling a bankrupt company's archive directly; a wind-down outside court with a privacy policy that forbids transfer.
“If a hearing is required under section 363(b)(1)(B), the court shall order the United States trustee to appoint, not later than 7 days before the commencement of the hearing, 1 disinterested person (other than the United States trustee) to serve as the consumer privacy ombudsman in the case”
“De curator is belast met het beheer en de vereffening van de failliete boedel. ... De goederen worden in het openbaar of met toestemming van de rechter-commissaris ondershands verkocht.”
“Google, at $10m, won the August auction (as summarised by the fetch tool; hearing scheduled 14 October 2026).”
Does the data include controlled technical information, such as semiconductor designs, aerospace or defence engineering, or advanced encryption source code?
Releasing controlled technology or source code to a foreign national, even inside the US, is a deemed export under the EAR, and AI labs employ foreign nationals and use offshore compute.
A good answer
Technical archives screened for export-controlled content (ECCN or USML review) and controlled material excluded.
“Any release in the United States of "technology" or source code to a foreign person is a deemed export to the foreign person's most recent country of citizenship or permanent residency.”
“Customer shall not export, reexport, or transfer the Controlled Subject Matter except as authorized by these laws and regulations.”
Is the buyer, or anyone it may pass the data to, a foreign person, and does the dataset contain US sensitive personal data above the bulk thresholds?
The DOJ Data Security Program bans data brokerage of bulk US sensitive personal data to countries of concern and requires a contractual onward-transfer ban and violation reporting for any other foreign buyer.
A good answer
Buyer screened for ownership by or location in China, Russia, Iran, North Korea, Cuba or Venezuela; data checked against thresholds (for example biometric identifiers over 1,000 US persons, health or financial data over 10,000, covered identifiers over 100,000); onward-transfer clause in the licence for foreign buyers.
A red flag
A buyer with a Hong Kong or mainland China parent; voice or face data on thousands of US persons going to any foreign lab with no clause.
“(b) Biometric identifiers collected about or maintained on more than 1,000 U.S. persons; ... (d) Personal health data collected about or maintained on more than 10,000 U.S. persons; (e) Personal financial data collected about or maintained on more than 10,000 U.S. persons; (f) Covered personal identifiers collected about or maintained on more than 100,000 U.S. persons”
“no U.S. person ... may knowingly engage in any transaction that involves any access by a foreign person to government-related data or bulk U.S. sensitive personal data and that involves data brokerage with any foreign person that is not a covered person unless the U.S. person: (1) Contractually requires that the foreign person refrain from engaging in a subsequent covered data transaction involving data brokerage of the same data with a country of concern or covered person; and (2) Reports any known or suspected violations”
What to fix before a sale (10)
Each is real work, but routine: a consent, a licence list, an assignment, a scan.
Is any of it data or content you licensed from someone else, such as purchased contact lists, research reports, stock media, fonts or design kits?
Licences for bought data and media usually forbid resale or sublicensing, so it usually cannot be passed on to a lab without the licensor's consent.
A good answer
Bought and licensed material is identified and removed before packaging, and the provenance file shows what remains is the company's own.
A red flag
CRM records enriched from a list vendor, stock images inside design files, paywalled reports in a shared drive.
“Review customer, partner, vendor, and data license agreements, identify data held as a service provider or processor, and review platform terms for the systems the data will be extracted from.”
“The licensor agrees to deliver, at the time of license, a structured record of the data's origin”
Do the records include staff discussing pay, working conditions or union matters, or is any workforce covered by a collective bargaining agreement?
US employees have a protected right to act together on working conditions, union-related communications are a category counsel advise excluding, and a bargaining agreement may itself restrict a sale.
A good answer
Union and HR channels and any discussion of pay or conditions are excluded; if a bargaining agreement exists, it has been checked for data and monitoring clauses.
A red flag
Employee organising chats, #random channels full of workplace complaints, or a unionised workforce with no review.
Fixable before a sale. Applies to: company records, text and archives, audio and speech, video and footage. Reasoned, not stated by a source: No source found applying Section 7 directly to the sale of employee communications; the link is an inference from the protected-activity right and counsel's exclusion advice.
“Where employees are unionized, the collective bargaining agreement may restrict the sale, and the company may have bargaining obligations.”
If you have a works council in the Netherlands or Germany, has it been consulted about processing staff data for this sale, and has it agreed where required?
Dutch law requires works council consent for any arrangement on processing employees' personal data, and German works councils co-determine technical systems that can monitor staff behaviour or performance.
A good answer
The works council approved the processing arrangement (the Netherlands) or a works agreement covers the export and anonymisation tooling (Germany), recorded in minutes.
A red flag
A Dutch or German entity with a works council and no consultation; decisions taken without consent can be void in the Netherlands if the council invokes it.
Fixable before a sale. Applies to: company records, text and archives, audio and speech, video and footage, source code. Reasoned, not stated by a source: Whether a one-off export and sale is an 'arrangement' (regeling) under WOR 27(1)(k) or a monitoring system under BetrVG 87(1)6 is not settled in the sources read; the 'void' consequence refers to WOR article 27(5), not re-read. A question for a lawyer.
“De ondernemer behoeft de instemming van de ondernemingsraad voor elk door hem voorgenomen besluit tot vaststelling, wijziging of intrekking van: ... k. een regeling omtrent het verwerken van alsmede de bescherming van de persoonsgegevens van de in de onderneming werkzame personen”
“Der Betriebsrat hat, soweit eine gesetzliche oder tarifliche Regelung nicht besteht, in folgenden Angelegenheiten mitzubestimmen: ... 6. Einführung und Anwendung von technischen Einrichtungen, die dazu bestimmt sind, das Verhalten oder die Leistung der Arbeitnehmer zu überwachen;”
For US consumer or staff data, does the de-identification meet the CCPA definition, including a public commitment and contract terms binding the buyer?
CCPA 'deidentified' data needs three things from the business holding it: reasonable technical measures, a public no-re-identification commitment, and contracts making recipients comply.
A good answer
Reasonable measures documented, a published commitment not to re-identify, and the licence obliges the buyer (and anyone it passes data to) to do the same.
A red flag
A 'PII removed' claim with no public commitment and no contractual flow-down.
“"Deidentified" means information that cannot reasonably be used to infer information about, or otherwise be linked to, a particular consumer provided that the business that possesses the information: (1) Takes reasonable measures to ensure that the information cannot be associated with a consumer or household. (2) Publicly commits to maintain and use the information in deidentified form and not to attempt to reidentify the information ... (3) Contractually obligates any recipients of the information to comply with all provisions of this subdivision.”
Does the code include third-party open-source code, and do you have a list of each licence?
Open-source licences travel with the code, and the breach-of-licence claims against GitHub Copilot survived and remain pending even after the DMCA claim failed in September 2026.
A good answer
A software bill of materials or licence scan; vendored and copied third-party code is either removed or delivered with its licence and attribution; only the company's own code is warranted.
A red flag
Vendored GPL/AGPL code mixed into the repo, copied snippets without attribution, no licence scan.
Did every employee and contractor who created the code, documents or recordings sign an agreement assigning their work to the company?
A company can only license what it owns, and SaaS terms say the customer owns its content only as between it and the vendor, not against its own contractors.
A good answer
Signed invention and IP assignment agreements for all contributors, including contractors and agencies; gaps listed and excluded.
A red flag
Freelancers or agencies without written assignment, open-source contributors' pull requests, founders who never assigned pre-incorporation work.
“You represent and warrant that you are either the author of all Customer Content to be provided under this Agreement or have obtained and hold all rights necessary to provide such Customer Content”
“That the licensor owns or has the rights to license the data.”
Does the data include legal advice, litigation material, or trade secrets (yours or other companies')?
Disclosing privileged communications to a buyer can waive privilege, and counsel list privileged material first among categories to exclude.
A good answer
Legal channels, outside counsel email and litigation folders are excluded; trade secrets the company wants to keep are excluded or the buyer's use is restricted contractually.
A red flag
#legal channels, email with outside counsel, board minutes, or a competitor's materials received in a deal.
“Exclude high-risk data entirely rather than relying on de-identification. Consider privileged material, HR and personnel files, payroll and tax records, health and benefits information, union-related communications, customer and partner confidential information, licensed data, and data from regulated systems.”
Does any lender hold security over the company's assets or IP, or have a right to approve asset sales or IP dispositions?
Venture and bank loans commonly include a negative pledge on IP and an asset-sale covenant that lenders watch for disposals of material IP.
A good answer
Loan documents checked; the data licence is a permitted disposition or outbound licence, or the lender consented in writing.
A red flag
Outstanding venture debt with a disposition covenant, or a lien on all assets including IP, and no consent.
“agrees not to grant a security interest in its intellectual property to anyone else (lenders watch the disposition or asset sale covenant for borrowers using it to dispose of material intellectual property)”
Has the board approved the licence, and do investor agreements or charters give investors a veto over selling or licensing key assets?
A sale of all or substantially all assets needs shareholder approval in Delaware, and preferred-stock protective provisions often require investor consent for asset sales or exclusive licences.
A good answer
Board resolution plus any required investor consents on file; for a wind-down, the dissolution plan covers the sale.
A red flag
An exclusive licence of the company's main dataset with no board or investor sign-off.
“Do not share sample data before this review, since a sample itself could be a disclosure.”
What to disclose (1)
Not a block, but a buyer paying for exclusivity will ask.
Has a software vendor already used this content to train its own AI (for example Atlassian since August 2026, GitHub Copilot interactions, HubSpot, ServiceNow Now Assist, Figma)?
It does not stop a sale, but a buyer paying for exclusivity will want to know whether a platform vendor already holds a model trained on the same material.
A good answer
You list each vendor's AI-training setting and when it was switched off, and the licence states exclusivity only against third parties you control.
A red flag
Promising 'never seen by any AI' when the workspace sat on a plan where vendor training defaulted on.
“If your Atlassian organization has an active Enterprise plan, you can opt-out of metadata contribution. If your Atlassian organization’s highest active plan is Free, Standard, or Premium, metadata contribution is always on, and you’re not able to opt-out. ... every customer has the option to opt out of contributing in-app data.”
“We may also use Customer Data to train our AI models in compliance with our obligations under the Agreement.”
Research from public sources, read on the day shown. It is not legal advice: a sale needs a lawyer who has read the records and the contracts.
In this topic
Rights and law
Start hereCan this data legally be sold or licensed for AI training?Privacy law, copyright and fair use, software vendors' terms, de-identification, bankruptcy sales and what a licence agreement holds.Reading
AIData.Broker, “Can it be sold? The rights behind AI training data”, reviewed 11 Oct 2026. https://www.aidata.markets/rights
Free to reuse under CC BY 4.0 if you credit AIData.Broker and link to this page. Each row names the publication that stated it, and words quoted from a source keep that source's own terms.