Shadow AI: regaining control over employees' AI tools

What is shadow AI, and how far does it reach into your company?

Shadow AI refers to employees' use of AI tools without validation or oversight from the company or its IT department (DSI) [32]. It is the AI counterpart of shadow IT [32]. One source describes it as "unregulated and unapproved" use of AI by employees [3], while another calls generative AI the "first shadow application of the world of work" [5].

What it looks like in practice:

• a personal ChatGPT account opened with a Gmail address [28];
• personal generative AI accounts and unvetted extensions not approved by the IT department [25];
• AI tools used on personal devices for work [7].

The phenomenon is growing because it meets a need: gaining efficiency [33].

What the figures allow us to claim

Our file contains only one academic measurement. According to a study published on Cairn, only 20% of respondents use only the tools provided by their organization [2]. The available excerpt specifies neither the sample nor the date, so retain the order of magnitude, not a generalizable statistic.

Two other figures are circulating, but they come from blogs whose methodology we could not verify:

• 68% of French employees reportedly use AI without informing their manager [22];
• 8 out of 10 companies reportedly lose control of their AI [21].

These two figures do not measure the same thing as the academic study. They cannot be compared with it, and they do not replace an internal assessment: only a mapping of your own usage will show where you stand.

What risks does unauthorized AI pose to the company?

Shadow AI exposes the company to three families of risk: legal (GDPR, AI Act), technical (security breaches, data outside your perimeter) and operational (hallucinations). Most available sources are content from vendors or consulting firms. They identify these risks without quantifying them in a verifiable way.

• GDPR, security breaches, hallucinations. These are the three risks highlighted by content from the company Koul on shadow AI [22].
• AI Act compliance. Content from Adequacy presents shadow AI as a source of security and compliance risks under the AI Act [36]. The available excerpts do not detail the obligations involved or the penalties.
• Data exposure. Shadow AI escapes data security policies [40]. A book on hybrid intelligence notes that, without an appropriate architecture, AI uses can expose the company to risks [4]. The vendor Olfeo speaks of "major risks" for the companies concerned [17].
• A blind spot in governance. A practitioner of sovereign and resilient enterprise AI describes shadow AI as a blind spot he encounters throughout his approach [6].

A 15% overcost to handle with caution

In Europe, the average cost of an AI incident would exceed that of a conventional data breach by 15% (Microsoft, 2024) [40]. This figure is reported by SoSafe, a commercial player in security awareness training. The original study is not in our file and we could not verify it. It should therefore be read as an unconfirmed order of magnitude, not as a reference figure for a budget or an internal business case.

Why is banning AI tools not enough?

Banning AI tools is not enough: a ban shifts usage without eliminating it. According to an article in Solutions Numériques & Cybersécurité, bans have simply moved usage outside the DSI's visible perimeter, making shadow AI even harder to control [35]. An employee blocked on their work computer can switch to a personal device, and the DSI then loses all visibility. An academic source also notes that employees already use AI tools on their personal devices for work [7].

Guide rather than ban. Several sources point in this direction, with similar wording:

• the goal is not to ban AI, but to frame it and make it governable [38];
• it is necessary to regain control without falling into the trap of sterile prohibition [21];
• this must be done without stifling innovation [22][23].

These contents come mainly from consulting firms and vendors, who have an interest in selling framing solutions. Their convergence with the analysis in [35] nonetheless remains clear.

Targeted blocking has its place, provided it is accompanied. A case described by a consulting firm illustrates the balance: public tools are blocked for one position through firewall rules, while the employee is trained on the AI module built into the internal document management software [31]. Blocking only works if it is accompanied by a validated alternative and training. Technical controls (traceability, alerts in case of exfiltration, marking of sensitive documents) complete the arrangement [23].

A change management issue. An analysis published on Cairn recalls that the generalization of generative AI tools implies "a real change" for the company, of which shadow AI, an unregulated and unapproved use, is one symptom [3]. Treating it solely as an infringement amounts to ignoring this change.

This is the logic behind Oupi: a single secure framework for all of the team's AI uses, instead of scattered personal subscriptions.

How do you identify and map the AI uses actually in place?

To identify shadow AI, you first need to establish who uses which AI tools, then cross-check this finding against three technical traces. The approach consists of mapping existing uses by identifying which employees use which tools [12], then sharing the results with the teams, since mapping goes hand in hand with awareness [11].

Three checkpoints are recommended for the technical side [26]:

• Access to third-party accounts: to which external services do the company's credentials or accounts give access?
• Data exports to external services: which data leaves the information system, and to what destination?
• Plugins and extensions installed on workstations: which modules are present on browsers and workstations?

These checks target uses that do not appear in the official application catalog. A personal generative AI account opened with a Gmail address is the typical example [28]. Bans without visibility do not help: they have simply moved usage outside the DSI's visible perimeter [35].

If these manual checks are not enough, detection platforms can identify real usage, including usage that escapes traditional channels, and help measure the organization's exposure [20]. On the network side, a cybersecurity vendor promotes an SSE (Secure Service Edge) approach to secure AI uses [17].

Reading caution: most of these recommendations come from commercial players, vendors, service providers or consulting firms [11][12][17][20][26]. None of these sources demonstrates the comparative effectiveness of the methods or tools. They provide a framework for approach to be checked against your own environment, not an independent benchmark.

Which validated AI tool should you offer to replace personal subscriptions?

The most effective answer is an internal offering validated by the DSI, which gives employees a secure alternative to their personal subscriptions. The sources consulted point in this direction: they recommend offering secure internal AI tools validated by the DSI, combined with clear rules on data (use, storage, sharing) [29]. These recommendations come from specialized media and consulting firms, not from a regulatory authority.

The starting point is well identified: shadow AI most often goes through personal accounts, without validation or company framework [15]. A ChatGPT account opened with a Gmail address is the typical example [28].

Criteria for a validated tool

A replacement tool must meet two requirements:

• Configuration through precise rules: some generative AI platforms allow usage to be configured according to defined rules [23];
• Data guarantees: some platforms ensure that data will be neither stored nor used to train models [23].

These points must be verified for each solution considered, both contractually and technically.

Oupi's positioning

Oupi offers a single secure framework for all of the team's AI uses, instead of a scattering of personal subscriptions. For a DSI, the benefit is consolidating usage fragmented across several individual accounts into a single entry point that it can validate and oversee.

This offer does not solve everything: it must be accompanied by the usage rules and technical controls described in the next section. But without a credible alternative, a ban alone has mainly shifted usage outside the DSI's visible perimeter [35].

Which rules, training and technical controls should be put in place?

Four levers secure the approach: rules on data, a framework co-built with the teams, practical training and technical controls. Apart from the government study [1], these recommendations come from non-primary sources (vendors, consulting firms, specialized media). They are recommended practices, not proven standards.

Clear rules on data. The aim is to define what may be used, stored and shared with an AI tool [29]. For the framework to be adopted, the government study suggests organizing workshops that co-define the quality of work with AI and anticipate its transformations, then building a framework of trust based on clarity [1].

Practical training. Awareness goes beyond reminding people of the charter. One vendor recommends training teams to detect risks, adopt good prompt hygiene, verify the results proposed and avoid copying and pasting without discernment [27]. Another adds digital security and GDPR compliance [29].

Technical controls. A specialized media outlet cites several measures: automated governance tools, marking of sensitive documents, traceability and alerts in case of information exfiltration [23]. It also mentions generative AI platforms that can be configured according to precise rules, or that guarantee data is neither stored nor used to train models [23]. On the DSI side, a security vendor recommends an SSE approach to secure AI uses [17]. These controls presuppose knowing who uses what, which is why prior mapping matters [11].

These three levers rely more easily on a single framework than on scattered personal subscriptions: this is the logic behind Oupi, which brings all of the team's AI uses together in one secure framework.