Is AI already training on my company's data without permission?

Three places your data can be, and what decides each#

Whether AI trains on company data depends on three branches: what is public, what your vendors hold and what stays in your private systems. Each branch has a different answer and a different set of controls, so a single yes or no is misleading.

Work through the branches in order. The first is mostly about accepting what has already happened, the second about reading contracts and settings, and the third about how your own people use AI tools.

Branch one: your public web pages#

Public web pages are the branch most likely to have been used already. Marketing pages, product documentation, help-center articles, blog posts, press releases and job posts are all reachable by crawlers, and AI developers have collected public web text at large scale.

Draw the line carefully. A public help-center article about how to reset a device is public; the support conversations that led someone to write it are not. Most of what makes a company's records distinctive never appears on its website.

Some AI developers publish the names of their crawlers and explain how to block them in robots.txt. A robots.txt file is a request, not a lock: it works only for crawlers that choose to honor it, and it applies going forward. Content already collected stays collected, and removing data from a trained model is difficult.

Website terms of use are the other lever. They can state that automated collection for AI training is not permitted, which records the company's position even though a crawler may never read them. Whether such terms can be enforced against a particular developer is unsettled, so treat them as a statement of intent and have counsel review the wording.

Branch two: data your software vendors hold#

Data held by your software vendors may be used for AI depending on their terms and your settings. Helpdesks, CRMs, call recording tools, field service platforms and productivity suites keep adding AI features, and their contracts differ on whether customer data is used only to serve you or also to improve models shared across customers.

Read the documents rather than the marketing page. The answer usually sits in a combination of the main terms, a data processing agreement, an AI-specific addendum and an admin setting.

• 1Find the clause on using customer data for product improvement or model training.
• 2Check whether the vendor treats de-identified or aggregated data differently from your raw records.
• 3Look for an opt-out or admin setting, apply it if you choose to, and keep a screenshot.
• 4Note whether the vendor can change its terms by notice, and who in your company receives those notices.
• 5Check whether your plan tier changes the commitments, since business and consumer plans often differ.

Branch three: your private systems#

Private systems such as your ERP, file servers, internal wikis and archived helpdesks are not reachable by AI developers unless access is granted or lost. No crawler can read a closed system. Apart from software vendors acting under terms you accepted, which is branch two, the legitimate route for an AI developer to obtain internal records is an agreement your company approves.

Exposure in this branch often comes from inside rather than outside. The common paths are employees pasting customer emails, estimates or code into consumer AI tools; drive or wiki links set to anyone with the link; contractors working on personal devices; and, less often, a security breach. Each has an ordinary control: an acceptable-use policy with approved tools, link-sharing settings, contractor agreements and basic security hygiene.

Old systems deserve a look too. A helpdesk or CRM the company stopped using may still hold years of records on a vendor account nobody monitors, under terms nobody has reread since signing. Either bring those accounts into the vendor review or archive the records and close the account properly.

A quick exposure check#

A quick exposure check needs one owner, a list and a few documents. It turns a vague worry into a short written record that management and the board can rely on.

• 1Open your robots.txt and website terms, and decide whether they reflect your current position on AI crawlers.
• 2List every vendor that holds customer, employee or operational records.
• 3Read each vendor's AI and data use clauses and check the admin settings.
• 4Ask team leads which AI tools staff use, and with what kinds of records.
• 5Search for publicly shared links to company drives and wikis, and close the ones that should not be open.
• 6Write down what you found, what you changed and who owns each follow-up.

Illustrative: a roofing company owner checks all three branches#

Illustrative: the owner of a fictional roofing and restoration company hears that a competitor is generating estimates with AI and wonders whether the company's own work helped train the tool. The office manager runs the exposure check.

The public branch shows that the company's project galleries and blog have been online for years and were likely crawled; the owner accepts that and updates robots.txt going forward. The vendor branch turns up a call recording tool whose terms allow it to use de-identified recordings to improve its AI; an admin opt-out exists and is applied. The private branch reveals that estimators had been pasting customer emails into a free chatbot to draft replies.

The owner adopts an approved AI tool on business terms, adds a one-page policy and confirms that the company's estimates, job photos and warranty history have never left its own systems. Those private records remain something the company could choose to license later, on its own terms.

Where licensing fits, and how SourceX approaches it#

Licensing is the route by which private operational records reach AI developers with the company's consent, and it requires an authorized signature. The company keeps ownership and decides the permitted use, because the records are licensed rather than sold outright.

SourceX runs that route as a managed transaction in five steps, Supply, Rights, Preparation, Approval and Delivery, known as the SourceX five-step transaction. Nothing is shared during the initial assessment, the supplier approves every step, and the SourceX Evidence Packet records the permitted use and release authorization for anything that is licensed. That written trail is the opposite of the untracked exposure described in the three branches above.