Will It Leak My Data?

What Happens to What I Type?

The most common question we get in any room, from any owner, is some version of this: what happens to what my people type into these tools? It is the right question. Your pipeline, your pricing, your customer list, your legal exposure, all of it flows through whatever your team pastes into a chat window.

And your team is already pasting. Employees adopt these tools long before the company writes a policy. The risk is rarely a dramatic breach. It is convenience. Convenience is how data leaves buildings.

Consumer Tools and Business Tools Play by Different Rules

Consumer AI tools and business AI tools operate under different rules, and most people conflate them. The free consumer tier of a tool like ChatGPT may use your conversations to improve future models unless you find the setting and turn it off. Enterprise and API versions of the same technology typically carry contractual no-training commitments. Same brand, opposite data postures.

A second distinction matters just as much: reach. A model that only sees text you paste is one exposure. A tool wired into your email, your files, and your customer database is a different animal. A model's intelligence is a capability question. Its access is a security question, and access is the one that decides how much you can lose.

• Consumer tiers may train on your inputs by default. Check the settings, then change them.
• Enterprise and API terms usually include contractual no-training commitments. Read them before rollout.
• Access matters more than intelligence. A tool connected to email and files can leak more than one you paste into.
• Your team is already using these tools. A written policy beats a quiet ban everyone ignores.

The three tiers

Route data by sensitivity, not by convenience. The question for every tool is which tier it belongs to.

Data Routing

Data routing is the practice of classifying your information by sensitivity and deciding, in advance, which class is allowed to touch which system. We run three tiers.

Commodity work, meaning drafting, research, and code with nothing sensitive in it, goes to the best frontier models available, because capability wins when nothing is at stake. Business-sensitive material, our CRM, financials, and deal flow, lives in local databases and gets queried in place; only derived summaries ever reach an outside API. Crown jewels, our credentials and proprietary methods, never touch any AI service at all.

The design question for your business is the same one: which rooms does each tool get a key to.

The Contractor and the Keycard

We treat our own agent platform as a contractor, not a partner. Contractors get keycards to the rooms they need, never master keys to the building. In practice, that means our credential store is locked by default, even to our own agents.

During one working session, an agent asked to list the directory where credentials live and was refused by policy, automatically, no human in the loop. That refusal is the system working. If your AI setup cannot refuse its own agent, it cannot refuse a compromised one. The boundary lives in configuration, where it holds, rather than in anyone's memory or good intentions.

Write Your Three Classes on One Page

Before you approve any AI tool, write down your three data classes on one page: what is public, what is business-sensitive, what can never leave. Then ask the vendor one question per class. Where does this data go, is it used for training, and can you show me the contractual language?

If the answer to data security is a link to a compliance page, treat that as a no. Most businesses find they can capture the large majority of AI's leverage while their most sensitive material never touches an outside service. Routing is how you keep both capability and control.