How to stop software vendors from training AI on your company data

Where do vendors get the right to train on your data?#

Software vendors get the right to train AI on your data from the agreements and settings you accepted, usually through clauses that never mention AI. The common sources are a service improvement clause, a right to create aggregated or de-identified data, a broad license to customer data, and product-specific AI terms that sit outside the main agreement and are incorporated by link.

Separate customer content from usage data when you read. Many terms treat telemetry, metadata and how your team uses the product as the vendor's own information while protecting the content you store, so the training answer can differ for each.

Which tools should you review first?#

The tools to review first are the ones holding your most distinctive operating records: CRM, helpdesk, email and chat, document storage, code hosting, call recording and meeting transcription, and your ERP or field service platform. These hold the history of how your company actually works, which is the kind of material model developers look for.

Build the list from single sign-on applications, accounts payable vendors and expense reports rather than memory. Teams often buy transcription, note-taking and writing tools on a company card, and those tools can end up holding client calls, pricing discussions and internal strategy.

How do you switch off training where a setting exists?#

Switching off training where a setting exists starts in the admin console, not in individual user profiles. Look for workspace-level controls over AI features, model improvement and data sharing, record what each setting says and the date you changed it, and save a screenshot to the contract file.

Settings are a partial fix. A vendor can rename, move or redefine a setting in a later release, and switching one off does not always reach data already used. Treat the setting as the immediate step and the contract as the lasting one.

What should a no-training clause say?#

A no-training clause should state that the vendor may not use your data, or anything derived from it, to train, tune or improve models for anyone other than you without your prior written consent. The table shows the main points and the fallback vendors commonly offer.

Fallbacks are negotiable. A large customer can often get the stricter version; a smaller one may accept the vendor's standard enterprise terms if those already exclude training on content and flow the restriction down to model providers.

Why vendor training can breach your own commitments#

Vendor training can breach your own commitments because much of the data in your tools belongs to, or describes, your customers. If your contracts or DPAs promise customers that their data is used only to serve them, and a vendor you chose trains on that data, the gap sits between you and your customer, not just between you and the vendor.

Check three places: the confidentiality and data-use clauses in your customer contracts, the subprocessor list you publish or share, and your privacy notice. A vendor that uses data for its own model training may not fit the processor or service provider role you described, which can turn a procurement oversight into a notice or contract problem.

The safest order is to fix the vendor terms first, then confirm that what you tell customers matches what your vendors are actually allowed to do.

How do you keep the protection current?#

Protection stays current only when someone owns it, because vendor terms and features change after signature. Fold the checks into processes that already exist, such as procurement, renewals and security reviews, rather than running a one-time audit.

• 1Add an AI and data-use check to procurement and renewal reviews.
• 2Assign a named owner to watch terms-change notices for critical tools.
• 3Re-check settings whenever a vendor launches an AI feature.
• 4Publish an employee policy on which AI tools may receive company and customer data.
• 5File the review record with the contract: terms version, settings, date and reviewer.

Illustrative: a consulting firm reviews its stack#

Illustrative: a fictional management consulting firm finds that several partners use a meeting transcription tool on client calls, bought on personal cards. The tool's terms let the vendor use de-identified content to improve its models, and the firm's client agreements promise confidentiality.

The firm moves the tool to an enterprise plan with training switched off and a no-training clause, asks the vendor to delete transcripts already stored, and adds the tool to an approved list. Its review of the CRM, document storage and proposal archive finds no training rights. The firm's proposals and engagement playbooks stay under its own control, which keeps open the option of licensing prepared versions of that internal knowledge later.

Why are your records worth protecting?#

Your records are worth protecting because the operating history in them, how problems were diagnosed, how jobs were scoped and how exceptions were resolved, is the kind of material model developers seek to license. A vendor whose terms let it train on those records captures that value without a separate agreement, a defined scope or any payment to you.

SourceX helps companies license prepared records on their own terms through the SourceX five-step transaction; the company signs off at each step, and the records are licensed, not sold. The SourceX Enterprise Data Value Framework helps a company judge which record families matter most before deciding what to protect and what it might license.