Shadow AI: How to Discover and Govern Unsanctioned AI Use

Shadow AI: How to Discover and Govern Unsanctioned AI Use

In many organizations, shadow AI has become the invisible fuel of the daily grind. Employees often paste sensitive data into unauthorized chatbots, install unvetted browser extensions to summarize meetings, or leverage hidden AI features within standard software without a single security policy in place. This uncontrolled sprawl creates significant data exposure risks and fragments your digital environment. I designed this discovery checklist to help you inventory these tools and regain governance without pushing usage underground.

You will learn to track specific usage signals to build your AI inventory and implement the necessary controls.

1. Run an Anonymous Discovery Survey

The biggest obstacle to securing your environment isn't a lack of tools; it’s the fear that if employees reveal their shortcuts, leadership will take them away. Most shadow AI isn't malicious: it’s driven by a desperate need for productivity. If you launch a heavy-handed audit, you only push these tools deeper into the shadows.

To surface the real landscape, make discovery psychologically safe. Run a short, anonymous survey to identify exactly which tools are in use, the specific tasks they solve, and the sensitivity of the data being shared.

Keep It Lightweight

Participation plummets if the survey exceeds three minutes. Limit your request to these six fields:

  • Tool Name: Name of the AI platform or extension.
  • Task: The specific job or workflow it solves.
  • Data Touched: Examples like PII, internal financials, or code.
  • Account Type: Personal or work-affiliated login.
  • Frequency: Usage volume (daily, weekly, or rare).
  • Team: The department using the tool.

Place an explicit “amnesty window” statement at the top: “Discovery is not punishment. We are gathering this inventory to identify how we can officially support your workflow securely. Nothing reported will result in disciplinary action.”

Categorize Immediately

Once data arrives, catalog each entry by Tool, Owner, Purpose, Data Classification, and Risk Tier. This baseline enables you to transition from reactive blocking to proactive, secure enablement.

2. Reconcile Traffic Logs to Identify Repeatable Signals

Surveying employees provides a pulse, but you need telemetry to see actual behavior. You aren't seeking perfect attribution for every request; focus instead on extracting repeatable signals that map usage to specific teams. Chasing one-off hits to obscure domains creates noise. Prioritize patterns that reveal consistent, high-frequency adoption.

Practical Telemetry for AI Visibility

Aggregate data from your secure web gateway, firewall categories, and DNS query logs. Most modern firewalls categorize traffic into "AI SaaS" or "Chatbot" buckets, which simplifies filtering.

Focus on these primary indicators:

  • Volume Spikes: Identify consistent daily traffic to specific AI endpoints. A single user hitting a site is noise; a department hitting it daily is an entrenched workflow.
  • Large Outbound Payloads: Flag significant data transfers to AI domains. This indicates employees are feeding proprietary sets or large documents into external tools.

Building Your AI Inventory

Log every confirmed signal in a master inventory to move from assumptions to defensible proof. Include these fields for every finding:

  • Endpoint/Domain: The specific service in use.
  • Department: The origin of the traffic.
  • Device Type: Whether the traffic routes through managed or unmanaged assets.
  • Frequency: Determine if this is a daily habit or a periodic experiment.
  • Confidence Level: A rating of how clearly the traffic represents a sustained business process.

The Governance Gotcha: Avoid immediate blocking upon discovery. Aggressive restrictions drive shadow AI off-network and onto personal cellular hotspots, where you lose all visibility. Log first to understand the underlying business intent, then implement a governance plan that offers a secure, sanctioned alternative.

3. Audit OAuth Consents to Find Hidden Integrations

Shadow AI is rarely a traditional app installation; it is usually an OAuth consent grant. These permissions allow third-party tools to hook directly into your mail, drive, or calendar. Because data interaction occurs server-to-server, it bypasses firewalls and network perimeters entirely. This is a critical vulnerability for remote workforces that frequently authorize "one-click" AI assistants to boost productivity.

What to Review

Audit your M365 or Google Workspace admin consoles for recent OAuth grants. Prioritize apps requesting high-risk scopes:

  • Mail: Permissions to read, write, or send emails on behalf of a user.
  • Drive: Broad read/write access to corporate file repositories.
  • Offline Access: The ability for an app to maintain persistent data access regardless of the user’s session status.

Documenting in Your AI Inventory

Treat every OAuth integration with the same scrutiny as installed software. Log the following for each identified tool:

  • App and Publisher: Verify the legitimacy of the entity.
  • Requested Scopes: Define the specific level of access granted.
  • Approver and User Count: Track who authorized the link and how many users are connected.
  • Data Sources: Identify which specific drives or mailboxes are exposed.

Action Cue: Prioritize Triage

Treat any tool with broad file or email access as a critical security risk. If a third-party AI holds persistent permissions to read your entire corporate inbox or storage, escalate it immediately. Shadow AI risks escalate rapidly when integrations gain deep access to sensitive data, creating pathways for unintended leakage or lateral movement.

4. Uncover Endpoint AI Clients and Browser-Based Scrapers

Shadow AI is rarely confined to sanctioned web services. It frequently resides on local devices as desktop clients or browser extensions. These tools bypass perimeter controls, scraping screen content, recording meeting audio, or exfiltrating local data under the guise of "productivity." Because they operate within the browser or OS, they can harvest sensitive information from your screen before it is ever rendered or transmitted, effectively rendering your sanctioned SaaS policies toothless.

What to Scan For

Prioritize your inventory by targeting tools that run locally and bypass web security:

  • Desktop AI Clients: Background applications that monitor your clipboard or active windows.
  • Meeting Note-Takers: Apps that record or transcribe audio and video, often joining calls automatically without oversight.
  • Screenshot-to-AI Tools: Utilities that capture screen regions, potentially leaking internal credentials or PII.
  • Browser Extensions: Extensions requesting “read and change all data” permissions, which can scrape live SaaS portal content.

How to Do It

Audit your software inventory using RMM or endpoint management tools. If visibility is limited, enforce browser-level allow/deny policies via MDM or Group Policy to restrict extensions to approved entities only.

Inventory Fields

For every client or extension, log these data points:

  • App/Extension Name & Version: Track specific iterations.
  • Install Count: Identify usage trends by department.
  • Permission Set: Define access levels (e.g., screen recording, clipboard, email).
  • Business Owner: Identify the manager overseeing the team using the tool.
  • Data Touched: Note if the app interacts with local documents or cloud-synced files.

Why It Matters

Extension-based shadow AI is the most effective bypass for standard SaaS controls. Because these tools sit on the endpoint, they function outside your perimeter, reading data directly from your monitor. Inventorying these clients is the only way to prevent silent data exfiltration occurring on user devices.

5. Map and Inventory Unmanaged API Usage

Direct API usage represents the most significant form of “AI tool sprawl.” Developers often integrate OpenAI, Anthropic, or similar model APIs directly into internal scripts, CI/CD pipelines, or automation tools. Because these integrations bypass the browser and procurement workflows, they remain invisible, until an unexpected usage bill arrives.

Where to Look

You will not find these integrations in standard software logs. Audit the infrastructure layers where your technical teams operate:

  • Code Repositories: Use secrets scanning tools to detect hardcoded API keys.
  • CI/CD Variables: Inspect configuration variables in GitHub Actions, GitLab, or Jenkins.
  • Internal Wikis/Vaults: Check documentation or password managers for “team-wide” keys.
  • Help Desk Tickets: Review previous requests for “LLM integration” or “model rate limit” assistance.

Capture the Full Picture

Build a technical registry for every key discovered. An incomplete inventory leads to both security gaps and bill shock. Capture the following fields:

  • Provider: (e.g., OpenAI, Anthropic, Mistral).
  • Key Owner: The individual or service account linked to the key.
  • Linked Project: The specific codebase or pipeline utilizing the integration.
  • Environments: Development, staging, or production.
  • Usage Limits: Existing spending caps or rate limit configurations.
  • Billing Owner: The cost center responsible for consumption charges.

Quick Control Recommendation

Do not wait for a full audit to secure these assets. Immediately rotate keys that access sensitive datasets. Move all active keys into a centralized secrets vault and apply strict tags to separate experimental workloads from production pipelines. This ensures immediate visibility over your security posture and ongoing AI spend.

6. Inventory Embedded AI Features in Enterprise SaaS

Shadow AI is rarely a standalone app. Increasingly, it exists as silent, opt-in features hidden inside the enterprise software you already pay for. CRM platforms, HRIS suites, email clients, and document editors aggressively roll out embedded AI that enables by default or via a single click. Because these features integrate into your trusted stack, they often bypass procurement vetting and security reviews entirely.

Spotting Silent Adoptions

The "we didn't know we had AI" scenario is a major organizational blind spot. To uncover these features, audit where you manage vendor relationships:

  • Admin Consoles: Audit the "AI" or "Labs" sections in your SaaS admin dashboards. Check feature toggles for generative summaries, writing assistants, or predictive analytics updated automatically by vendors.
  • Vendor Release Notes: Review update logs for key platforms like M365, Google Workspace, or Salesforce. New AI-powered tools often appear in these updates, waiting for users to toggle them on.
  • License Add-ons: Scan billing statements for new SKU tiers or AI "credits" providing access to premium model features that were previously unavailable.

AI Inventory Requirements

Treat embedded features with the same scrutiny as standalone software. Your inventory must document:

  • Parent System: The core platform (e.g., Salesforce, Slack).
  • AI Feature Name: The specific tool (e.g., Einstein Copilot).
  • Data Sources: Which datasets the AI accesses (e.g., private docs, customer records).
  • Training Settings: Whether the vendor uses your input to train their global models.
  • Enabled Groups: Which teams hold access.

Why This Matters

Mapping embedded features helps you separate "shadow" AI (unauthorized, high-risk tools) from "unsanctioned-but-already-licensed" software. By inventorying these features, you can formally sanction them, apply proper data-use policies, and manage sprawl without disrupting essential workflows.

7. Build a Standardized AI Inventory Framework

You cannot govern what you cannot see, yet most organizations fail by overcomplicating the intake process. Do not start with expensive GRC software. Start with a simple spreadsheet that acts as your living, single source of truth. Sophistication follows once you establish the habit of tracking.

The Minimum Viable Inventory

Keep your columns focused on what is usable for decision-making. If a field does not help you decide whether to block, tolerate, or sanction a tool, remove it. Your spreadsheet must capture:

  • Identity: Tool name, department, and business owner.
  • Usage: Primary purpose or specific use case.
  • Data Profile: Data types touched (public, internal, confidential, regulated) and account type (personal vs. corporate).
  • Vendor Intelligence: Training/retention stance and integration points (e.g., OAuth, API connectors).
  • Control Layer: Risk tier (1 to 4), required security controls, and decision status (sanction, tolerate, or prohibit).

The 4-Tier Risk Triage

Assign every tool a risk tier to clarify your governance workload:

  • Tier 1 (Low Risk): Public-facing data, marketing copy, or generic brainstorming.
  • Tier 2 (Internal): Non-sensitive internal meetings, general project coordination, or low-stakes drafting.
  • Tier 3 (Sensitive): Proprietary IP, internal financials, or sensitive personal data.
  • Tier 4 (Regulated): Tools touching data under contractual or legal restrictions like PHI or SOC 2 scopes.

This inventory acts as your single source of truth for audits and governance. By centralizing this data, you stop playing "whack-a-mole" with shadow AI and begin managing it as a transparent, categorized element of your operational stack.

8. Consolidate AI Sprawl to Reduce Attack Surface and Cost

Shadow AI is a cumulative operational tax. Individual subscriptions for AI assistants, writing aids, and meeting transcribers quickly morph into a fragmented web of services. This sprawl bloats budgets and exponentially increases your attack surface by creating orphaned accounts and unmonitored renewals.

Why Sprawl is a CFO-Level Concern

Unmanaged subscriptions are latent security incidents. Accounts linked to offboarded employees often remain active, providing entry points for credential harvesting. Furthermore, "paid personal" reimbursement patterns mask rogue expenses, preventing the firm from leveraging enterprise pricing or enforcing security standards. Discovery isn’t just security theater; it is the path to consolidation and predictable OPEX.

Metrics That Matter

To translate discovery into savings, track these indicators:

  • Functional Duplication: Audit tools used for writing, meetings, research, and coding. Multiple transcribers across departments signal wasted spend on redundant capabilities.
  • Orphaned Access: Flag accounts that remain active after staff departures, specifically those using personal email or OAuth connectors.
  • Expense Anomalies: Scan for recurring, small-dollar charges to AI domains in expense reports. These represent self-funded tools that bypassed security vetting.

Inventory Best Practices

Update your AI inventory to drive cost-control. Each entry should include:

  • Cost Owner: Department head responsible for the budget.
  • Seat vs. Usage: The delta between licenses purchased and active users.
  • Renewal Date: Essential to prevent auto-renewals.
  • Overlap Note: Identify tools duplicating M365 or Google Workspace capabilities.

By centralizing this data, you move from chaotic shadow spending toward a predictable IT model. For a deeper dive into mitigating these financial risks, see our CFO playbook on IT cost-control.

9. Define Your AI Detection and Monitoring Posture

Treating all AI tools as equal threats leads to wasted resources. While traditional network logs and firewalls effectively track public chat usage, they remain blind to the data flowing into embedded model connectors, RAG pipelines, and agentic automation scripts.

Use this decision rule to right-size your detection approach:

  • For general usage visibility: If your goal is tracking "who is using public chat tools," standard firewall categories, proxy logs, and identity reviews suffice. Focus on reinforcing authenticated access and baseline security policies.
  • For data-in-use visibility: If you need to monitor "what data is flowing into prompts and connectors," traditional perimeters fail. Deploy AI Security Posture Management (AI-SPM) or an LLM-specific proxy. These tools integrate with DLP to inspect prompt content and API responses in real-time.

Inventory your current detection coverage by risk tier. Document exactly what you can see versus where you remain blind, such as payload contents within encrypted API calls. This transparency prevents blind spending and helps you prioritize investments in deep inspection tools.

Integrate this detection strategy with robust endpoint security. Device-level telemetry, managed through our endpoint security guidance, serves as your ultimate safety net, capturing local AI scraping and browser-based data leakage that network-layer monitoring inevitably misses. By aligning detection depth with actual risk, you gain control without over-engineering your security stack.

The 4-Week AI Governance Execution Plan

Shadow AI is a governance problem that requires a repeatable discovery process. A blunt ban only drives activity into private, unmonitored channels. This four-week plan prioritizes discovery and rapid consolidation to ensure your IT environment remains secure, visible, and optimized.

Prerequisites: Your Governance Toolkit

Before you begin, prepare these two assets:

  • The AI Inventory Spreadsheet: Create a centralized document tracking Tool Name, Owner, Risk Tier, and Data Classification.
  • The Amnesty Comms Template: Prepare a clear, reassuring email announcing that the organization is mapping AI usage to improve security rather than punish productivity.

Week 1: Survey and Signal Gathering

Build your baseline inventory.

  1. Launch the Amnesty Survey: Distribute the anonymous survey to all staff. Frame this as an enablement project.
  2. Extract Quick Logs: Perform a rapid pull of firewall and DNS logs. Identify traffic spikes related to AI SaaS to pinpoint which departments are the most active.
  3. Initial Entry: Create the first 20 to 50 rows in your inventory spreadsheet.
  • Outcome: You will have a draft inventory list that correlates observed traffic with survey responses.

Week 2: OAuth and Endpoint Deep Dive

Identify the "hidden" AI that bypasses the network perimeter.

  1. OAuth Audit: Scan admin consoles for third-party AI integrations requesting high-risk scopes like Mail or Drive access.
  2. Endpoint Scan: Utilize your RMM or MDM tools to identify browser extensions and desktop AI clients.
  3. Tagging: Categorize these items as Tier 3 or 4 based on their level of system access.
  • Outcome: You will have a list of high-risk integrations ready for immediate triage or removal.

Week 3: API Hunt and Reconciliation

Address the technical debt created by your data-heavy teams.

  1. Secrets Scan: Use scanning tools to locate hardcoded API keys in your repositories or CI/CD pipelines.
  2. Embedded AI Review: Audit platforms like Salesforce or M365 to identify default-enabled AI features.
  3. Owner Reconciliation: Contact the identified owners to document the business purpose for each tool.
  • Outcome: You will have a complete registry of technical AI integrations and embedded tools with assigned business owners.

Week 4: Sprawl Assessment and Strategy

Finalize your governance structure.

  1. Detection Gap Decision: Compare your inventory against your monitoring capabilities. Decide which tools require specialized security tools and which can be managed via standard proxy logs.
  2. Draft Sanctioned List: Create a list of approved, enterprise-grade AI tools.
  3. Intake Form: Build a low-friction request form for new tools.
  • Outcome: You will have a finalized AI policy that governs how to request, review, and approve future tools. Pair it with employee cybersecurity training to ensure staff knows the requirements for safe AI use.

Convert, Don't Ban: The Governance Outcome

Apply one of these three outcomes to every tool in your inventory:

  1. Approve: If the tool meets security standards, provide it under an enterprise license.
  2. Tolerate: For low-risk tools that lack an enterprise alternative, apply restrictions, such as disabling history, and mandate security training.
  3. Prohibit: If the tool presents an unacceptable risk, block it, but only after providing a viable, sanctioned alternative that achieves the same business goal.

The logo of Cortavo.

Need help running this discovery, auditing your stack, or building your AI guardrails? Contact our team today to learn how Cortavo manages your IT security.

Frequently Asked Questions

What is shadow AI, and how does it differ from shadow IT?

Shadow AI refers to the use of artificial intelligence tools (such as chatbots, browser extensions, or embedded features) without formal IT approval or security review. While shadow IT covers all unauthorized hardware and software, shadow AI is specifically dangerous because it often involves employees feeding proprietary company data, code, or sensitive customer information into external models that may use that data for future training.

What are the biggest shadow AI risks for SMBs and mid-market companies?

The primary risks are data leakage and contractual liability. When employees use unvetted AI, they may inadvertently expose intellectual property or PII to third-party providers. Furthermore, companies often face compliance and audit failures because they cannot prove they have control over where their data flows, potentially violating privacy regulations or confidentiality agreements with clients. See Section 7 for the risk triage framework to assess these threats.

Won’t an AI inventory become outdated immediately?

It will, which is why your inventory should be a living document rather than a static report. Start with a minimal spreadsheet to capture your baseline, then update it quarterly. Use automated signals from your OAuth reviews, endpoint inventory, and secure web gateway (SWG) logs to refresh entries. By tying inventory updates to existing lifecycle tasks, like software audits or offboarding, you keep visibility high without creating constant manual overhead.

Should we block ChatGPT, Claude, and Gemini on the network?

Blocking these tools usually does more harm than good. A hard network block simply drives employees to use personal cellular hotspots, which removes all visibility and control. Instead, provide a sanctioned, enterprise-grade AI alternative, establish clear internal policies for what data can be used, and implement monitoring to ensure compliance. Strategy should focus on enablement, not prohibition.

We found 12 AI tools doing the same job. What now?

This is a prime opportunity for consolidation. Use your inventory to identify redundant subscriptions, which are often a sign of wasted spend and fragmented security. By standardizing on one or two enterprise-approved tools, you can cut AI tool sprawl, negotiate better licensing terms, and tighten your security posture. For a deeper breakdown on managing this financial waste, see our IT cost-control resource for CFO stakeholders.

Need help building visibility into your AI usage or standardizing your technology stack? Contact Cortavo to learn how our all-inclusive model keeps your business secure, efficient, and growth-ready.