When an engagement is governed by an NDA, AI usage becomes a control problem. Not a moral problem. Your job is to define what information can leave your environment, what cannot, and how you prevent accidental leakage in day-to-day work.

‍

A sensible starting point is “default closed.” If a tool sends prompts or code to a third party, treat it as external disclosure unless you have a contractual enterprise arrangement that explicitly covers data handling. Under NDA constraints, it’s usually safer to assume public AI tools are not approved for client code, credentials, architecture details, incident logs, or anything that would identify the client or their systems.

‍

What can be allowed is still meaningful. AI can be used for generic work that does not contain client-sensitive content: drafting internal checklists, rewriting generic documentation, generating boilerplate that you then adapt in your own environment, or exploring general technical patterns. It can also be allowed for client-specific work if you have an approved tool with strong controls, and the client has explicitly agreed to that workflow in writing.

‍

The biggest risk is not intentional misuse. It’s convenience. Engineers paste a snippet to “save time” and forget it contains a URL, an internal hostname, a customer identifier, or a secret embedded in a config. Policy language is not enough. You need enforcement mechanisms that make safe behavior the default.

‍

Practical enforcement tends to look like this. Maintain an approved-tools list. Document a clear rule for what data is prohibited. Provide a safe alternative for when people hit the boundary. Train the team with examples that reflect real mistakes. Confirm the workflow during onboarding. And periodically audit for tool sprawl, because unapproved tools appear quietly.

‍

If you operate under strict NDAs, the policy should also include a redaction expectation. If you ever need to use AI on sensitive context, you remove identifiers and secrets first, and you accept that redaction reduces accuracy. The policy should state that clearly, so teams don’t pretend redaction is free.

‍

Axveria view: Under NDA constraints, the safest AI stance is not “never.” It is “approved tools, explicit boundaries, and enforceable defaults.”