Blog
Using Generative AI Safely Without Banning It: Zscaler AI Guard
18 November 2026

Most employees now use at least one generative AI tool every day, while most organisations do not know what data goes to those tools. A total ban lowers productivity and pushes use into the shadows, and uncontrolled freedom increases the risk of data leakage. Zscaler AI Guard offers a third way between the two: using AI with rules, inline and visibly.
Where is the real risk in AI use?
The risk of generative AI tools usually lies not in the tool itself but in what is given to it. Text copied to summarise a customer list, find a bug in source code or translate a contract goes to a service outside the organisation's control. Where that data is stored, whether it is used for model training and who can access it is often unclear.
The second risk is visibility. Alongside the one or two AI tools an organisation has approved, employees use dozens more through browser extensions, assistants embedded in SaaS apps and personal accounts. This “shadow AI” does not appear in the inventory, so it does not make it into policy either.
What does Zscaler AI Guard do?
AI Guard is the component of Zscaler's AI security family that protects the user and application side. It has two parts.
AI Guard for Users: who, which tool, what risk?
AI Guard for Users classifies more than 3,500 AI tools and decides access per user, per app and per risk. One of four decisions is applied to every interaction:
- Allow: an approved tool is used freely by an approved user group.
- Block: access to high-risk tools or tools against company policy is cut off.
- Isolate: the tool opens in an isolated browser session; copy, paste and download can be disabled.
- Caution: the user is informed about the company's AI policy before continuing.
An important detail is that different policies can be defined for corporate and personal accounts of the same app. For example, an assistant licensed by the company can be allowed while use of the same assistant with a personal account is isolated. Risk classifications are refreshed continuously from live traffic across the Zscaler cloud.
AI Guard for Apps: what happens inside approved tools?
Approving a tool does not mean every prompt sent to it is safe. AI Guard for Apps inspects prompts and responses in real time inside the tools the organisation has approved. Even in an allowed tool, a prompt containing sensitive data is detected and handled according to policy.
Stopping data leakage at the prompt
Zscaler's data loss prevention (DLP) capabilities also apply to generative AI interactions. Prompts containing sensitive data can be blocked before they are sent, and browser isolation keeps copy, paste and download under control. The organisation does not have to rebuild its existing DLP dictionaries and classifications separately for AI traffic.
The strength of this approach comes from inspection happening on the traffic path. AI Guard runs inline on the Zscaler Zero Trust SASE platform; the same policy applies whether the user is in the office, at home or on the road, and no separate AI agent has to be installed on the endpoint.
Visibility: an AI inventory
AI Guard usage data feeds an AI inventory. Which tool is used by whom, how often and with which data types can be tracked on dashboards. That visibility makes it possible to write policy based on real usage rather than guesswork: a heavily used tool can be moved to a corporate licence, and a risky trend can be spotted early.
Agents and in-house AI applications
Zscaler's AI security family is not limited to the user side. AI Access Graph maps which data each approved application can reach; AI Gateway authorises each action taken by AI agents individually and covers agent communications such as MCP and A2A. Red-teaming capabilities that test in-house AI applications against attack scenarios are also offered. Organisations can roll these layers out gradually as their needs grow.
| Need | Zscaler component |
|---|---|
| Which AI tools are in use? | AI Guard for Users + AI inventory |
| Who can access which tool? | AI Guard for Users (allow / block / isolate / caution) |
| Is sensitive data reaching approved tools? | AI Guard for Apps + DLP |
| What data do applications reach? | AI Access Graph |
| How are agent actions authorised? | AI Gateway |
Where to start when rolling it out
- Visibility first: in the first phase, current AI use is monitored and inventoried without blocking.
- Classify the policy: tools are grouped as approved, restricted and prohibited, and the split between corporate and personal accounts is defined.
- Connect data rules: existing DLP rules are applied to AI prompts, with blocking enabled for critical data types.
- Inform users: the caution decision is the most effective way to build awareness before banning.
- Review regularly: new tools appear every week, so the inventory and policy are updated periodically.
This order brings AI use under control without cutting it off overnight or disrupting how employees work. The goal is not to block AI but to keep the productivity gains while protecting the organisation's data.
If you would like to make AI use in your organisation visible and design a safe-use policy together, contact our cybersecurity solutions team.