The signature starts the conversation
A company requires an NDA, restricts documents, and demands care with client information. Then someone copies a proposal, a spreadsheet, or an internal conversation into an artificial intelligence tool. The task may be legitimate and the time saved substantial. But who assessed this new path for the information?
Signing a confidentiality agreement does not, by itself, answer the operational questions about AI use. It should not be treated as automatic permission to send data to any service, either. My point is simple: the attention given to the contract needs to continue when the work begins.
Classify before sharing
A practical policy starts by distinguishing public content, internal information, and restricted material. A corporate statement already published does not call for the same decision as an ongoing negotiation, clients’ personal data, or access credentials. Trade secrets deserve protection even when they do not identify a person.
For each category, define permitted tools, purpose, and who approves exceptions. Employees need to be able to decide when faced with an actual file. A generic rule such as ‘use common sense’ leaves the most important part without guidance.
The necessary context may be smaller than the file
Before attaching an entire database, ask what information the task actually requires. To improve a proposal template, its structure and a fictional example may be enough. Passwords, tokens, and access keys should remain outside this material. Removing someone’s name does not guarantee anonymity when the context still makes that person recognizable.
The Brazilian data protection authority’s Technology Radar highlights limiting personal data to what is necessary, including in prompts and attachments. It provides a reference for considering minimization when preparing a task. (ANPD — generative artificial intelligence)
Assess the service being purchased
It makes no sense to claim that every AI service trains on everything it receives. Conditions vary with the service, contractual arrangement, and settings. The assessment needs to check retention, possible training use, deletion, people with access, and companies involved in processing, including subprocessors. Record the answers and when they were checked.
Not using content for training and not storing content are different conditions. A promise about one does not automatically explain the other. The UK’s voluntary AI security code emphasizes communicating how data is used, accessed, and stored. (UK government — AI security code)
A connector expands the decision
When a tool connects to email, a CRM, or a document repository, the assessment extends beyond what someone typed. Which folders can it consult? Whose permissions does it use? Can access be revoked? An assistant that needs one document should not receive access to the entire operation for convenience.
CISA and its partners recommend limiting autonomy and avoiding broad or unrestricted agent access to sensitive data and critical systems. This helps turn the intention to protect into a technical restriction that can be verified. (CISA — careful adoption of AI agents)
The policy needs to work in everyday practice
Provide approved tools, examples of what may be submitted, and a channel for questions. Define who reviews providers and permissions when use changes. If an inappropriate submission occurs, the team needs to know how to report it, stop access, and involve those responsible for assessing the case. Silence does not help contain exposure.
A company that expects productivity from AI needs to organize the conditions for its use. Confidentiality becomes more consistent when the contract, the tool, and people’s behavior point in the same direction. The question moves beyond ‘did we sign the NDA?’ to include ‘do we know where the data is going?’.
The video behind this reflection
This article develops the topic presented in the video about data sent to AI and protecting information at work. Watch the video on Instagram.