How to detect and govern shadow AI in your organization

To detect and govern shadow AI, organizations need to discover which AI tools employees are using, understand who is using them and why, then turn that visibility into an enforceable AI app policy. The goal is not only to find unsanctioned AI use, but to control which GenAI and AI-enabled apps are allowed, blocked or monitored across the organization.

Start with network, proxy, secure web gateway and DNS logs, because they can reveal many browser-based AI services. Your firewall, secure web gateway or DNS resolver logs outbound connections, and most browser-based AI tools resolve to identifiable domains. Pull a list of destinations and filter for known AI services such as chatgpt.com, claude.ai, gemini.google.com, perplexity.ai and copilot.microsoft.com. Treat this as a first layer of discovery, not a complete shadow AI inventory.

Network telemetry alone can miss or only partially reveal:

• AI features embedded inside sanctioned SaaS apps.
• AI browser extensions.
• Desktop or native AI clients.
• Mobile and unmanaged devices.
• API-based use.
• Private or self-hosted models.
• Traffic obscured by VPNs, encrypted DNS or app-specific infrastructure.

Separate sanctioned from unsanctioned traffic. A sanctioned tool is one your organization has approved, contracted and configured. Everything else is shadow AI, including free-tier versions of tools you have approved only on a paid enterprise plan.

When an unsanctioned domain appears, record the user or device, the frequency of access and --- where telemetry supports it --- upload volume or data-transfer indicators. DNS logs alone will not give you meaningful upload volume; proxy and secure web gateway tooling generally will. Do not block immediately. A sudden block tells employees nothing and pushes them to a personal phone or home machine. Log the finding, confirm it is genuine AI use, and add it to your inventory before you decide on enforcement.

Network logs show domains but not what happens inside the browser, so review browser activity directly on managed endpoints. Browser history reveals which AI tools an employee opens and how regularly, and extension logs reveal AI tools running as add-ons rather than as visited sites.

Use browser and extension telemetry only on managed endpoints and in line with company policy, employee notice and applicable privacy requirements. Depending on jurisdiction, employment agreements and endpoint management setup, this kind of inspection may require prior notice, legal review or explicit acceptable-use policy coverage.

Flag AI browser extensions specifically. Sidebar assistants, writing helpers and summarizers request broad page-level permissions, and some have been caught harvesting chat content and internal URLs. Inventory every AI-themed extension installed across your fleet, note its permission scope, and treat unknown publishers as suspect until reviewed.

Look for personal account sessions on corporate endpoints. An employee logged into a personal AI account on a work machine is one of the hardest leaks to see, because the traffic looks identical to sanctioned use. Check whether sign-ins use corporate single sign-on or a personal email. A personal login on a corporate endpoint is a clear shadow AI signal worth a direct conversation.

Shadow AI is not only standalone chatbots. AI features now ship inside SaaS tools you have already approved, and they often escape governance because the parent app is sanctioned. Microsoft 365 Copilot, Google Workspace AI features and Slack AI each process company content through a large language model, and each needs its own review.

Find which features are active. Check admin consoles for AI toggles, recently enabled add-ons and per-user licensing for AI capabilities. Many features default to on after a vendor update, so re-check after every major release.

Ask vendors specific questions about data handling: Is prompt content used to train models. Where is data processed and stored. Is content retained, and for how long. Can the feature be disabled per user or per group. Treat any embedded AI feature you cannot answer these questions for as ungoverned until you can.

Logs can miss a major category of risk: employees using personal AI accounts on unmanaged devices. A browser paste into a personal AI account on a personal phone or home machine never touches your network or your endpoints, so the only way to surface it is to ask.

Run a short, blameless survey. Ask which AI tools employees use for work, what tasks they use them for, and whether they use a personal or work account. Frame it as planning for safe access rather than catching wrongdoing, because honest answers depend on employees not fearing punishment. Make clear you intend to provide approved tools, not just take tools away. The responses will reveal tools and use cases no log can show.

Identity logs are an underused detection source. Your single sign-on provider, identity provider and email system record when an employee creates or accesses an account with an AI service. Search authentication logs and OAuth grant logs for AI vendor names.

Pay attention to sign-ups that use a corporate email address against a consumer AI plan. That pattern means company identity is tied to a tool with no enterprise contract, no data protection terms and no admin control. Search inbox rules and mailbox traffic for verification and welcome emails from AI vendors as a second signal.

When you find sign-ups, add each to your inventory with the account owner and the email used, then decide whether to migrate the user to a sanctioned plan or revoke access.

A one-time audit is a snapshot. Detection becomes durable when you know what normal looks like. Establish a baseline of expected AI use: which tools, how many employees, typical request volume and the usual data types involved.

Define an anomaly threshold against that baseline. Examples include a single endpoint contacting many distinct AI domains in a day, a sharp spike in upload volume to an AI service, or a new AI domain appearing for the first time. These thresholds turn raw logs into alerts you can act on.

Set the baseline before you enforce anything. Enforcement without a baseline generates false positives, frustrates employees and erodes trust in the program. Measure first, then tune thresholds, then enforce.

The output of the previous steps is one living document: an AI tool inventory. For each tool, record the tool name, the number of users, the data types submitted, a business justification, and a governance status --- approved, monitored, restricted or blocked.

Use the inventory as a decision tool. Tools with clear business value and acceptable data handling move toward approved status with a contract and controls. Tools that carry real value but also real risk can sit in a monitored or restricted state while usage is watched. Tools with no justification or unacceptable risk move toward blocked status and retirement, with an approved alternative offered in their place.

The inventory is the foundation of an enforceable AI usage policy. A policy written without it is guesswork; a policy built on it reflects how employees actually work and can be enforced, because you know what you are governing and which controls apply to each tool. Visibility produces the list; the governance status and the controls behind it are what actually reduce risk.

The seven-step audit works, but it requires continuous monitoring that a small IT team cannot sustain by hand. New tools launch weekly, employees switch accounts, and a snapshot goes stale within days.

Acronis GenAI Protection helps MSPs move from one-time shadow AI audits to continuous visibility and governance across client environments. It discovers and monitors browser-based GenAI tools by user, tool and frequency, giving MSPs a clearer view of adoption patterns, risky usage and potential policy gaps, and it applies anomaly detection against normal usage patterns with extractable reports for review and policy decisions.

With governance controls, MSPs can help clients define which GenAI and AI apps are approved, monitored or restricted, keeping AI usage aligned with internal policy. Where policy enforcement is enabled, MSPs can control access to unwanted or unsanctioned AI tools, including through URL-based controls where applicable, keeping AI usage aligned with client policy.

This sits alongside Acronis GenAI Protection's broader capabilities for sensitive data protection and harmful prompt prevention, helping MSPs support safer AI adoption without adding another standalone point solution. Prompts can be inspected for sensitive content such as PII and PHI before submission to public or unsanctioned AI tools, and malicious prompt techniques such as prompt injection can be detected and blocked.

Acronis GenAI Protection runs on Windows 10, Windows 11 and macOS. It is delivered and managed through the Acronis platform, so discovery, monitoring, governance and reporting sit alongside existing endpoint protection rather than in a separate console.

• An audit of network logs, browsers, SaaS features, employee surveys and identity logs is the starting point for finding shadow AI, but no single source gives a complete inventory.
• Shadow AI detection should lead to governance: an approved, monitored, restricted or blocked status for each AI tool.
• A living AI app inventory is the foundation for an enforceable AI usage policy.
• Visibility without governance only produces a list; visibility plus enforcement reduces unmanaged AI risk.
• MSPs can help clients move from shadow AI to secure AI by monitoring usage, controlling AI app access and reporting on policy alignment.
• Shadow AI is part of a broader GenAI risk model that also includes sensitive data leakage and harmful prompts.
• Organizations with high levels of shadow AI faced an average of $670,000 in additional breach costs compared with those that had low or no shadow AI, according to the IBM Cost of a Data Breach Report 2025.
• AI use is now near-universal: 88% of survey respondents said their organizations regularly use AI in at least one business function, according to McKinsey's The State of AI in 2025.
• Establishing a baseline of normal AI use before enforcement reduces false positives and keeps employee trust intact.

What is the fastest way to start detecting shadow AI?

Pull your network, proxy or DNS logs and filter outbound traffic for known AI domains such as chatgpt.com and claude.ai. This surfaces many browser-based AI tools within an afternoon, using data you already collect, and gives you a ranked list of tools and users to investigate before you invest in any dedicated tooling. Treat the result as a starting point rather than a full inventory, since it will not capture AI features inside sanctioned SaaS apps, unmanaged devices or API-based use.

Can I detect shadow AI without a dedicated tool?

Yes, to a point. Network logs, browser history, SaaS admin consoles, identity logs and employee surveys all expose AI use with no new software. The limit is sustainability and enforcement: manual audits go stale fast, miss pastes to personal accounts on unmanaged machines, and give you no mechanism to control which AI apps are actually allowed. Continuous monitoring with policy enforcement is what turns a periodic audit into governance.

What should I do when I find an unsanctioned AI tool?

Do not block on sight. Record the tool, the users, the data involved and the business reason, then add it to your inventory and assign a governance status. Decide whether to approve it with a contract and controls, monitor or restrict it while you assess, or block it and offer an approved alternative. Blocking without an alternative pushes employees to personal accounts you cannot see.

How often should I audit for shadow AI?

Treat discovery as continuous, not annual. AI tools launch and change weekly, so a quarterly manual review is a practical minimum, with the approved tool list checked more often. Continuous automated monitoring is the stronger model because it flags new tools and usage spikes as they happen rather than months later.

What is the difference between shadow AI detection and shadow AI governance?

Detection is finding which AI tools employees use, who uses them, and where data may be flowing. Governance is deciding which tools are approved, monitored, restricted or blocked --- then enforcing that decision through policy and controls. Detection comes first because you cannot govern what you cannot see, but detection without governance only produces a list, not risk reduction.