Published · Updated · Reviewed by Horacio Poblete, CEO and founder
Picture an ordinary Tuesday. Late in the afternoon, a director receives the draft of a contract with a supplier, needs a summary for the next morning’s meeting and pastes the whole document into the AI app on their phone, logged into their personal account. Maybe nothing bad happens. But the company doesn’t know where that text is, under what terms, how long it’s kept or whether the confidentiality clause with the supplier allowed it. Each decision missing from that scene has an owner: some belong to the executive, others to the company.
Where the risk shows up in practice
- Personal accounts: without an enterprise contract, the individual plan’s terms apply, and the company has no control or visibility over what was sent.
- History and training: in individual accounts, what happens to conversations depends on the plan’s terms and the settings the person chose, which they don’t always check.
- Connected apps and extensions: browser extensions that read the open page and apps linked to email or drives see far more than the user imagines.
- Shared conversation links: whoever receives the link sees the whole conversation, including what was pasted at the start.
- Agents with access to systems: an agent that reads email, documents or the CRM can end up exposing what it read to people who shouldn’t see it.
- Wrong output treated as fact: a made-up number in a board report is also an incident, even if no data leaked.
The risk of an AI system revealing information it shouldn’t has a name in the OWASP Top 10 for LLM Applications 2025: LLM02:2025, sensitive information disclosure. It applies to any application built on a language model, including the internal agents the company creates. The practical rule is to give each agent access only to what it needs and to review those permissions the way you review a person’s access.
Three categories of information
Whether it’s safe to put company data into ChatGPT or another AI depends, above all, on what data it is. Before pasting anything, the question is which of these categories it falls into:
| Category | What information | Examples |
|---|---|---|
| Fine anywhere | Public, or unrelated to the company | Asking for the structure of a presentation, reviewing a text already published, understanding a market concept |
| Only in the approved enterprise tool | Internal, not public | A meeting agenda, the draft of the monthly report, a competitor analysis done by the team |
| Never, in any tool | Sensitive, protected by contract or not yet disclosed | Sensitive personal data, client data protected by contract, results before disclosure, ongoing negotiations and mergers, passwords and credentials |
Exceptions to the last category only with formal approval.
The generative AI use policy is the place to write down this classification with your company’s examples, so no one has to guess.
Habits that protect the executive
Most of the protection lies in small gestures, repeated every week:
- Anonymize: replace the names of clients, people and companies with labels (Client A, Supplier B) before pasting.
- Describe instead of pasting: often an unidentified summary is enough, such as "long-term supply contract with a regional exclusivity clause".
- Check the contract and NDA: before pasting a supplier’s proposal, partner material or a due diligence document, see whether the confidentiality agreement allows it.
- Use the enterprise account on your phone too: it’s in a hurry, out of the office, that the personal account tends to come into play.
- Don’t share the conversation: copy the useful part into the document instead of sending the link.
- Check what comes out: numbers, names, dates, quotes and clauses verified before passing them on.
Back to Tuesday’s director. With these habits, the same task looks like this: they open the enterprise tool, confirm that the supplier’s confidentiality clause doesn’t prevent the use, ask for the summary and, before taking it to the meeting, check every deadline and every amount against the original draft. It takes almost the same time, with a different level of risk.
What the company needs to provide
Individual habits don’t hold up without structure. The company needs to do its part:
- An approved tool that works well: if the enterprise version is worse than the personal one, people go back to the personal one.
- Contract terms checked: the major vendors’ enterprise plans usually state in the contract that data isn’t used to train models, as the pages of OpenAI, Microsoft and Google say; check your tool’s contract, along with retention and storage location.
- Single sign-on and access control: with corporate login (SSO), access for someone who leaves the company ends along with their other accounts, and each area sees only what it needs.
- A written rule: a short policy, with the data classification and real examples.
- An owner: someone is responsible for approved tools, questions and incidents, and has time for it.
Data protection law and personal data
When the information involves people, data protection law comes in; in Brazil, the LGPD. Pasting a client list, a résumé or a customer service history into an AI tool is a way of processing personal data, and the law applies to it as to any other system. Sensitive personal data (racial or ethnic origin, religious belief, political opinion, union membership, health, sex life, genetic or biometric data) calls for extra care and, under this guide’s rule, falls into the "never" category.
Legal basis, purpose and retention are matters for legal and the data protection officer. This guide organizes day-to-day use and doesn’t replace that guidance.
The example comes from the board
A security rule the board doesn’t follow becomes a rule for show. If the CEO pastes the strategic plan into a personal account, no policy will convince the team otherwise. The reverse also holds: when leadership works in the approved tools, talks openly about how it uses AI and asks before pasting a third party’s document, the behavior spreads without the need for an audit.
A company only becomes AI-First when its leadership works AI-First.
That’s why the AI-First Executive Journey starts with the board’s routine: security in using AI is, above all, a habit of the people who decide. To see how that habit fits into an executive’s week, the guide on how an executive uses AI day to day shows the routine of emails, reports, meetings and decisions.
