What data has already left an acquired company through AI tools?

Where does data leave a company through AI tools?#

Data leaves a company through AI tools by four main channels: AI features inside the business software it already pays for, vendor terms that let those vendors train on customer records, browser extensions and personal AI accounts used by staff, and bulk exports or integrations that move records into outside AI services.

None of these requires bad intent. A support lead turns on reply suggestions in the help desk; a sales rep installs a notetaker; an analyst uploads a customer export to summarize churn reasons. For an acquirer, the questions are what happened before closing and what is still happening now.

The audit checklist#

The audit checklist covers each channel with where to look and who owns the check. Run it soon after closing, alongside the system register, so both draw on the same list of systems.

SaaS AI features and vendor training settings#

SaaS AI features are often the largest channel because they process records at volume and run quietly once enabled. Many help desk, CRM and call platforms now offer AI summaries, reply drafting or call analysis, and whether the vendor may use that data to improve its models varies by vendor, contract and plan.

Check both the feature switch and the training setting, and note the date of each change if the console shows a history. Where it does not, record the current state and the date you checked. Counsel should read the vendor's data processing terms, since marketing pages are not a reliable guide to what the contract permits.

Browser tools, personal accounts and notetakers#

Browser tools and personal accounts are the hardest channel to see because they sit outside the company's systems. Managed browser reports show extensions on company devices, expense claims show AI subscriptions bought on company cards, and short interviews with department heads fill in the rest.

Ask concrete questions rather than general ones. Which tools did you use to draft customer emails? Did anyone paste ticket threads, contracts or code into a chat tool? Did a notetaker join customer calls, and where did transcripts go? People answer specific questions honestly when the tone is fact-finding rather than disciplinary.

Code needs a separate check. Developers may have used AI coding assistants on personal accounts, sending repository content to a vendor under consumer terms the company never reviewed. GitHub's Copilot documentation, for example, says that from April 24, 2026, interactions on individual Copilot plans (Free, Pro, Pro+ and Max) may be used to train and improve AI models unless the user turns off the 'Allow GitHub to use my data for AI model training' setting. Record which plan each developer used and how that setting was configured.

Why earlier disclosure matters for licensing#

Earlier disclosure matters because a later license is easier to value when the company can say exactly who already holds the records. If an AI vendor received support tickets under terms that allowed training, a future licensee of those tickets may ask whether it is paying for something another developer has already used.

Exclusivity is the sharpest issue. A company cannot credibly offer exclusive rights to records it has already disclosed under training-permissive terms, and an acquirer or licensee may ask for representations on exactly that point. Disclosure does not always erase value, since prepared, linked and documented records differ from raw uploads, but it has to be known and disclosed.

What to do with what you find#

Act on findings in a fixed order, so the audit stops exposure without stalling the business.

• 1Stop new exposure: switch off unreviewed AI features and training settings, and pause notetakers and integrations pending review.
• 2Document the past: record each finding with the system, dates, record types and the terms in force at the time.
• 3Assess with counsel: decide whether personal data, customer confidentiality or contract obligations are affected, and whether any notice is required.
• 4Request deletion where the vendor's terms provide for it, and keep the written confirmation.
• 5Update the AI use policy and approved tool list, and give staff a sanctioned alternative.
• 6Carry the findings into the integration file and any later licensing or exit diligence.

Illustrative: a portfolio CTO audits a transportation software company#

Illustrative: a fictional sponsor acquires a transportation management software company, and the portfolio CTO runs the checklist soon after closing. The company uses Zendesk for support, HubSpot for sales, a call recording platform and GitHub for code.

The audit finds AI reply suggestions enabled in the help desk with the vendor's model improvement setting left at its default. Two sales reps used a personal notetaker on customer calls. An engineer connected a personal AI coding assistant to a private repository. A customer success manager uploaded a ticket export to a consumer chat tool to draft a quarterly review.

The CTO switches off the training setting, ends personal notetaker use, moves coding assistants to a company plan with reviewed terms, and asks counsel to assess the ticket upload. When the sponsor later considers licensing the ticket history, the file lets the rights review state which records were exposed, and the affected accounts are left out of scope.

How SourceX treats prior AI tool exposure#

SourceX asks about prior exposure in the Rights step of the SourceX five-step transaction, because buyers ask about it too. A documented audit lets the company state what was disclosed, to whom and under what terms, and scope any package around it. The audit findings can be discussed as metadata, since no records move during the initial assessment.

For each package that proceeds, the SourceX Evidence Packet records provenance and licensing rights, including any known prior disclosure that bears on exclusivity.