Back to the blog
AI & Smart Assistants

How to Let Employees Use AI Without Leaking Data to ChatGPT (A Guide for CISOs)

Employees are already pasting documents, code and customer data into ChatGPT. Blocking it does not stop that; it just moves it to their personal phones. This guide explains how to protect company data when people use AI: a usage policy, technical controls, and a secure replacement that runs on the organization's own server.

21/11/202610 min readRelated to WISHGOODS
לקריאת המאמר בעברית
How to Let Employees Use AI Without Leaking Data to ChatGPT (A Guide for CISOs)

In almost every organization today there are employees pasting a contract into ChatGPT to summarize it, a spreadsheet to analyze it, or a piece of code to find a bug. They are not being careless. They do it because it saves them an hour, and nobody has given them a better way.

To a CISO or an IT manager this looks like a data leak happening every day, and the instinctive response is to block it. This guide explains why blocking alone does not work, and how to reach a state where employees get powerful AI and the data stays inside the organization.

Why blocking ChatGPT at work does not solve the problem

Blocking ChatGPT and similar AI tools on the corporate network only stops their use on work computers. The employee who needs to summarize a document simply photographs it with a personal phone, or forwards it to a personal email account and carries on from there. That is shadow AI: AI use the organization cannot see, log or control.

In that situation the risk has not gone down, it has only gone out of sight. So the right goal is not "zero AI" but an approved, convenient path for using AI, one that is also the easiest option for the employee.

What data employees leak to ChatGPT without noticing

The common leaks are not state secrets but everyday work: customer documents and contracts, financial data and reports, information about employees and salaries, source code with access keys left inside it, and internal correspondence with identifying details.

In public tools, and in personal accounts above all, the data is processed on an external vendor's infrastructure and retained according to the vendor's policy. Even when the vendor commits not to train models on it, the organization no longer controls where the data is or who can access it.

Step 1: an AI usage policy and data classification

Start with a short document every employee can read in five minutes: which AI tools are approved, what data may go into each tool, and what is forbidden in any case. A policy nobody can remember will not be followed.

A good baseline is three levels: public data, allowed in any tool; internal data, allowed only in tools the organization has approved; and sensitive data (customer, medical, financial, legal, defense) allowed only in a tool that runs inside the organization.

Step 2: technical controls to protect company data in AI use

On the technical side there are several layers: DLP systems that detect sensitive data being copied to external sites, restricting AI tools to company accounts instead of personal ones, and monitoring usage to understand which needs employees are trying to meet.

Enterprise editions of cloud services, such as ChatGPT Enterprise and Microsoft 365 Copilot, give better control than a personal account. But the data is still processed outside the organization, so for the most sensitive level of data they are not always enough.

Step 3: a secure ChatGPT replacement that runs on your own server

The complete answer is to give employees an AI assistant on par with ChatGPT that is installed on the organization's own server or private cloud. When the model runs locally, questions, documents and answers never leave the organization, and the employee has no reason to look for another tool.

This is exactly what Zoe for Business by WISHGOODS was built for: an enterprise AI assistant that ships as a Docker bundle on the organization's server, with the option of a local model that sends nothing out, AES-256-GCM encryption for connection tokens, a security log with an integrity check on every entry, and human approval before every outgoing action.

A checklist for CISOs before choosing an AI tool

Where is the data processed and stored, and who can access it? Can the model run inside the organization? Are there role-based permissions, so each employee sees only what they are allowed to? Is every question and action recorded in an audit log that can be exported to a SIEM?

And further: can the tool take actions (send an email, change a record) without approval? Is there SSO and MFA? What happens to the data when you stop using the tool? A tool without clear answers to these questions is a risk, however convenient it is.

Regulation: privacy law and regulated sectors

Privacy regulation keeps raising organizations' responsibility for the personal data they hold, including data passed to third parties. In Israel, Amendment 13 to the Privacy Protection Law tightens that responsibility; elsewhere, laws such as the GDPR play a similar role. A public AI tool into which employees paste customer data is exactly such a case.

In hospitals and health funds, in financial institutions and insurers, and in the defense industry the requirements are stricter still, and keeping data inside the organization is often not a preference but an obligation. In those cases an on-premise installation is the simple way to meet the requirements and still give employees AI.

How much does it cost to set up private AI in an organization

Enterprise cloud services are usually priced as a monthly subscription per user, so the cost grows with every employee who joins. Direct access to models through an API is priced by usage (tokens), and the cost is hard to predict.

With private AI on the organization's own server, the main cost is one-time: hardware or a private cloud, setup and adaptation, and then maintenance. There is no charge per question or per user, so the more employees use it, the lower the cost per employee. The break-even point depends on the number of users and the kind of tasks, and that is what a pilot measures.

Key takeaways

  • Blocking ChatGPT alone does not prevent data leaks; it moves usage to personal phones (shadow AI).
  • A short policy with three data levels, backed by technical controls such as DLP and company-only accounts.
  • For the most sensitive data, the secure replacement is an AI assistant running on the organization's own server, with an audit log and human approval.

Next step

Zoe for Business is an enterprise AI assistant for on-premise installation: the data stays on your server, every action is logged, and nothing is sent without approval. We would be glad to show you how it looks in your organization, in a two-week pilot.

Meet Zoe for Business