Shadow AI: how to detect usage and build the right framework

Collaborateur utilisant un outil d’IA personnel au travail

Shadow AI: how to detect usage and build the right framework

An employee drafts a document with their personal account. A team tests an assistant without approval. An AI feature appears in software that is already deployed. These uses often meet real needs, but they fall outside the intended framework. That is what shadow AI is all about.

What is shadow AI?

Shadow AI refers to the professional use of artificial intelligence tools or features that have not been approved under the organization’s rules: a free service, a personal subscription, a browser extension, a feature turned on without review.

The provider’s name isn’t enough to qualify the situation. An enterprise plan may be authorized while a personal account on the same service is not. A tool may be accepted for public information and excluded for internal documents. Everything depends on the account, the context and the data.

A broader phenomenon than you might think

According to Productiv (2026, US companies), a company uses an average of 14 AI tools, while its IT department knows about only 4 or 5. And employees know it: in a survey published by PagerDuty in June 2026, 66% of respondents say they have already used AI tools at work that they believed were unauthorized.

The study covers 1,250 office workers at companies with at least $500 million in revenue in Australia, Japan, the United Kingdom and the United States. The trend is clear: the need often comes before the decision to provide the tools.

Situations to qualify differently

An unknown tool for a legitimate task

A team prepares content from public information. You need to examine the need, the service’s terms and whether an approved solution exists. An unknown tool calls for an assessment, not a sanction.

A personal account used for work

The service is known, but the account is outside the company’s administration and contractual terms. The questions then concern access, data retention and what happens when the employee leaves.

Confidential information at stake

A contract, code or personal data require particular attention. Observing a connection to a service doesn’t tell you what was sent: you need to distinguish the presence of a tool, a risk signal and a confirmed exposure.

Providers located in other jurisdictions

Where is the data entrusted to an AI tool processed? The provider’s headquarters isn’t enough to answer: you need to examine the service’s terms and its configuration. Mapping where the providers you use are located becomes a question of data residency and control.

Cross-check sources, without overestimating any of them

SourceWhat it revealsIts limit
Invoices, corporate cards and expense reportsPurchases and subscriptionsDon’t include free usage
Vendor consolesManaged accounts and their activityCoverage varies, and no conversational usage
Network and proxy logsServices reached from the monitored networkShow neither the account nor the content; going further often requires decrypting traffic
Workstation measurementActual usage by tool and by plan, in the office and when working remotely without VPN, through passive observation and without intercepting encrypted trafficLimited to enrolled computers and supported tools
DLP tool recordsSensitive data exposed in promptsOnly if a DLP tool is already in place
List of approved toolsThe rules for qualifying observationsMust specify the accepted accounts and conditions

Cross-checking makes gaps readable. Also document the areas that aren’t observed: a pilot doesn’t represent the whole organization. A risk score helps prioritize checks, provided its criteria are explicit. It doesn’t, however, turn an observation into an incident.

Five steps to bring shadow AI under control

01. Build a contextualized inventory

Bring together observed tools, managed accounts and known purchases, specifying the period and coverage of each source.

02. Understand the needs

Talk with teams about their tasks and the limits of the tools available. A poorly suited alternative will leave the need unmet.

03. Define rules people can use

Approved tools, account types, permitted data, uses subject to approval, and a simple process to assess new needs.

04. Offer alternatives and inform

Everyone needs to know where to find an approved tool and whom to ask. Any measurement setup must also be disclosed to employees: in the United States, several states require employers to give notice before monitoring electronic activity, and in Europe employee representatives must be consulted. In all cases, it must remain justified and proportionate.

05. Review at the right pace

Tools, plans and needs change fast. Set a review frequency based on risks and observed changes.

Make the inventory a starting point

An unapproved tool may lead to an assessment, a change of account, an official alternative or a security investigation. The goal isn’t to ban everything, but to make usage understandable and rules applicable, so that business units, IT, Procurement and Security finally talk about the same thing. To take stock of your situation: five questions to ask about AI.

How Saaswedo Can Help

With sobrAIvo, Saaswedo measures actual AI usage on employee computers with its AI Trace software, through passive observation, in the office and when working remotely, without intercepting traffic and without any conversation content leaving the computer. Cross-checked in the portal with your network logs, HR file and DLP tool records, it reveals shadow AI, personal accounts and where your providers are located, with a risk score for each tool. Your data remains hosted on AWS, in the United States for US customers and in France for French customers, encrypted and isolated for each customer.

1600 900 SAASWEDO | Telecom Expense Management & Managed Mobility Services