Shadow AI: Have Businesses Already Lost Control of Their AI Usage?

In most businesses, introducing a new technology follows a relatively predictable path. A need is identified, a solution assessed, the risks analysed, and the service is then deployed in a managed environment.

Generative artificial intelligence has largely bypassed this model.

Today, an employee can create a personal account, install a browser extension and give it access to business pages, documents or applications within minutes. The tool promises to summarise a financial report, rephrase a contract or prepare a presentation. The adoption decision is no longer necessarily taken by the IT department. It is shifting to the user.

This reversal lies at the heart of Shadow AI: the professional use of artificial intelligence tools without sufficient approval, oversight or visibility from the organisation.

The phenomenon is generally presented as a cybersecurity problem. Yet it reveals a deeper transformation: the business no longer always controls the pace, or even the scope, of its technology adoption.

01In Switzerland, Usage Is Growing Faster than Controls

The available figures should be read with caution, as they do not all measure the same reality. Some surveys cover the entire population, others focus on employees, managers or technical environments observed by security vendors.

They nevertheless converge on one point: generative AI is no longer a marginal practice.

According to the Federal Statistical Office, 43% of the Swiss population aged 15 to 88 reported having used a generative AI application in 2025. Among people with above-basic digital skills, this proportion reached 66%, compared with 17% among those with low digital skills. [bfs.admin.ch]

Another survey, the AI Barometer 2025 by ZHAW and VOICETECHHUB, indicates that 43% of the employed people surveyed already use generative AI daily in their work. While 85% believe these technologies increase productivity, only 21% consider them trustworthy and 20% regard them as completely safe. [employes.ch]

This gap is telling. Users appear to be adopting a technology whose limitations in terms of reliability, trust and security they themselves acknowledge.

Businesses therefore face a paradox: the usefulness of AI accelerates its adoption, while the uncertainties surrounding how it works complicate its governance.

02Individual Adoption Versus Incomplete Organisational Readiness

Shadow AI does not stem solely from careless behaviour by employees. It can also be interpreted as the consequence of a gap between operational needs and the institutional response.

A Deloitte Switzerland study published in January 2025 shows that only 32% of the Swiss companies surveyed considered their technical infrastructure well or very well prepared for the adoption of generative AI. Half of the executive board members surveyed believed it would still take between one and three years for this technology to profoundly transform their sector, while 30% expected it to take more than three years. [deloitte.com]

At the same time, everyday usage is already developing on the ground. The contrast is hard to ignore: while management plans a medium-term transformation, employees are experimenting right now with tools available online.

This situation calls for a more nuanced view than the claim that Shadow AI is merely a lack of individual discipline. When an organisation offers no approved solution, no understandable policy and no simple procedure for experimenting, users meet their own needs.

Shadow AI is therefore also a symptom of an organisational inability to keep pace with innovation.

03The Personal Account, the Organisation’s New Blind Spot

The awareness bulletin illustrates a now familiar scenario: an artificial intelligence extension is installed in a browser and used through a personal account. The tool can read the content of a page, access a document or communicate with an external service.

For the user, it is an assistant.

For the IT department, it may be an unknown third-party provider with access to business information, without a contract, without centralised logging and without a controlled revocation procedure.

The international research available gives an indication of the scale of the problem. UpGuard’s 2025 report, based on responses from 500 security leaders and 1,000 employees, indicates that 81% of employees surveyed and 88% of security leaders say they use unapproved AI tools. The same report states that 45% of workers surveyed look for a way around an application being blocked. [upguard.com]

These results cannot be automatically transposed to the Swiss market. They do, however, show the limits of a strategy based solely on technical prohibition.

Blocking can reduce access from a managed environment. It does not necessarily remove the business need that drives that access.

04Data Leakage Is Not a Theoretical Risk, but the Figures Deserve Critical Scrutiny

Cybersecurity vendors regularly publish striking figures on the volume of information sent to AI platforms. These data are useful, but must be put in the context of their methodology and the commercial interests of their authors.

LayerX Security’s 2025 report, based on telemetry from browsers used in large client organisations, indicates that 45% of the users observed actively access AI platforms. The report also states that 40% of files uploaded to generative AI tools contain personally identifiable information or data covered by payment card industry standards. [go.layerxsecurity.com]

The figure is worrying, but it is not an exhaustive measure of all businesses. It describes the environments observed by a vendor specialising in browser security. The selection of client companies, their sectors and their digital maturity may influence the results.

This distinction matters. The debate on Shadow AI sometimes suffers from an accumulation of statistics produced by players who sell precisely the detection and control solutions in question.

This does not mean the risk is exaggerated or non-existent. It means that management must examine, at the same time:

• the origin of the data;
• the population studied;
• the definition of Shadow AI used;
• the collection method;
• the economic interests of the source.

Responsible governance should rest neither on denial nor on alarmism.

05“Employees Are Left to Fend for Themselves”

For Beni Eugster, Cloud Operation and Security Officer at Swisscom, the phenomenon emerges when businesses do not provide the tools employees expect. Employees then use the solutions of their choice with private accounts.

The expert believes the potential consequences concern data security, compliance and the company’s reputation. He notes, however, that the scale of the problem varies depending on the sector, the size of the organisation, its IT structure and the criticality of its data. [swisscom.ch]

This nuance is essential.

An extension used to rephrase a public text does not present the same level of risk as an assistant connected to a mailbox, a document repository or human resources files. Likewise, a tool backed by contractual guarantees and administrative controls cannot automatically be equated with a service used anonymously or with a private account.

Not all AI tools are equal. Nor are all uses.

The real challenge, therefore, is not to place each tool in a binary category, authorised or dangerous. It is to assess the combination of the service’s capabilities, the permissions granted, the data processed and the contractual framework.

06Employees in French-Speaking Switzerland Already Use AI, Sometimes Without Saying So

A Qualinsight survey reported by RTS in June 2026 indicates that 72% of working people in French-speaking Switzerland use artificial intelligence at work. According to the report, some of this usage remains hidden for fear of professional judgement or of being replaced. [rts.ch]

The testimony of Julia Vaez, functional analyst at the Fondation Asile des Aveugles, illustrates another ambiguity. She explains that checking work produced by AI can take almost as long as carrying out the task in the first place. She also questions the risk of gradually losing the ability to perform an activity that is systematically delegated. [rts.ch]

Shadow AI is therefore not only a confidentiality issue.

It also raises questions of quality, skills retention and the transformation of work:

• is the time saving real once verification is taken into account?
• who remains responsible for an error generated by the system?
• which skills risk being eroded?
• how can useful automation be distinguished from uncontrolled delegation?
• does covert use prevent the business from measuring the promised gains?

These questions are rarely covered by a traditional cybersecurity policy.

07The Illusion of Automatically Achieved Productivity

The commercial discourse around AI often rests on a simple equation: automation reduces the time needed to complete a task.

In practice, productivity depends on many factors: the quality of the output, verification time, the user’s level of expertise, the sensitivity of the content and the ability to integrate the response into an existing process.

A summary produced in ten seconds may require twenty minutes of checking. A convincing analysis may contain an error that is hard to detect. A fluent translation may change the meaning of a contractual clause.

The true return on AI should therefore not be measured solely in generation time. It should include the cost of validation, correction, governance and, where applicable, incidents.

This is where Shadow AI creates an additional difficulty: since the usage is not declared, the organisation can measure neither the real benefits nor the hidden costs.

08Unapproved AI Can Also Reveal Needs IT Has Not Identified

A purely defensive reading of Shadow AI overlooks part of the phenomenon.

The qualitative study conducted by Inria and datacraft across 14 organisations describes informal uses as practices that allow employees to gain efficiency, creativity or autonomy. The authors consider that these experiments can become a source of collective intelligence, provided they are brought out of the shadows in a negotiated way. [datacraft.paris]

The study also highlights that only 20% of the AI projects examined in the work cited reach industrialisation. It attributes this result in particular to unsuitable infrastructure, poor data quality, insufficiently defined objectives and a mismatch with the work actually carried out. [datacraft.paris]

The contrast is instructive.

Official projects can fail because they are designed too far from operational needs. Informal uses can succeed locally precisely because they start from a concrete problem.

The IT department must therefore strike a balance. Neutralising every experiment would risk suppressing a valuable signal about business needs. Tolerating all uses would expose the organisation to an uncontrolled proliferation of tools, accounts and data flows.

Shadow AI is both a security risk and an informal mechanism for innovation.

09The Swiss Framework Already Exists, Even Without a General AI Act

Switzerland does not yet have a general law devoted exclusively to artificial intelligence. This does not mean that processing carried out by these systems escapes the law in force.

The Federal Data Protection and Information Commissioner (FDPIC) points out that the Federal Act on Data Protection applies directly as soon as an AI system processes personal data. Manufacturers, providers and users must in particular ensure transparency regarding the purpose, operation and data sources used in the processing. [edoeb.admin.ch]

The FDPIC also states that users of language models must be able to know whether the information they enter is reused to improve the programs or for other purposes. Where there is a high risk, a data protection impact assessment may be required. [edoeb.admin.ch]

In February 2025, the Federal Council adopted a regulatory approach intended to strengthen Switzerland as an innovation hub, protect fundamental rights and increase trust in AI. Incorporating the Council of Europe Convention into Swiss law is to be accompanied by adjustments that are as sector-specific as possible. [bakom.admin.ch], [bj.admin.ch]

For businesses, the current absence of general AI legislation is therefore not an invitation to wait. Obligations relating to data protection, confidentiality, security and contractual commitments continue to apply.

10Prohibit, Observe or Support?

Three approaches can be seen in businesses.

The first is to prohibit unapproved tools. This response may be necessary to protect sensitive environments, but it risks shifting usage to personal accounts or devices.

The second is to tolerate experimentation without framing it. It facilitates innovation, but leaves significant blind spots.

The third seeks to organise controlled experimentation: a catalogue of approved solutions, a test environment, data classification, a rapid assessment procedure and reporting without automatic sanctions for errors made in good faith.

This last approach appears more demanding. It requires IT, security, legal and business teams to work together. It does, however, have the advantage of addressing the cause of Shadow AI rather than merely its symptoms.

It is not enough to tell employees what they must not do. They must also be given the means to work differently.

11From Prior Control to Continuous Governance

The artificial intelligence ecosystem is evolving too quickly to be governed solely by a static list of authorised tools.

An extension can change owner. A provider can change its retention terms. A feature initially limited to text can gain access to files, email or connected applications.

Governance must therefore cover several dimensions:

• Identity: managed business account or personal account.
• Data: public, internal, confidential or personal content.
• Permissions: read, modify, transfer or connect to applications.
• Provider: sub-processors, location, guarantees and contractual terms.
• Use case: writing assistance, analysis, decision-making or automation.
• Accountability: human validation, logging and the possibility of redress.

A solution authorised today should not enjoy unlimited trust. It must remain subject to reassessment proportionate to changes in its features and in the data it accesses.

12Shadow AI Is Also a Crisis of Internal Trust

The most delicate question may not be a technological one.

If employees hide their use of AI, the business can neither protect its data, nor measure the gains, nor identify training needs. Secrecy turns a potentially useful experiment into an undocumented risk.

Conversely, excessive monitoring can deepen mistrust and encourage workarounds.

The answer therefore does not lie solely in detection tools. It requires a clear organisational contract:

• the business explains the risks in concrete terms;
• employees declare their usage;
• business teams take part in assessing solutions;
• IT offers credible alternatives;
• errors made in good faith are reported quickly;
• restriction decisions are justified by risk.

This human dimension distinguishes AI governance from a simple web-filtering policy.

13Bringing AI out of the Shadows Without Stifling Innovation

Shadow AI is neither a phenomenon to be trivialised nor a threat that would justify an indiscriminate ban.

The figures show a rapid spread of AI across the population and the Swiss workplace. They also show a trust deficit and organisational readiness that remains uneven. [employes.ch], [bfs.admin.ch], [deloitte.com]

International studies point to significant volumes of unapproved usage and transfers of potentially sensitive data. But these results, often produced by security vendors, must be examined in the light of their methodology and should not be applied mechanically to all Swiss businesses. [upguard.com], [go.layerxsecurity.com]

The conclusion is less spectacular, but probably more useful: businesses cannot govern AI solely through fear of data leakage.

They need to understand why employees adopt these tools, what problems they are trying to solve and what conditions would make it possible to turn scattered initiatives into controlled collective capabilities.

The real challenge of Shadow AI is not to push every experiment back into the shadows.

It is to create enough trust, visibility and accountability for innovation to finally operate in the open.

14Key Figures

• 43% of the Swiss population aged 15 to 88 reported having used generative AI in 2025. [bfs.admin.ch]
• 43% of employed people surveyed in the AI Barometer 2025 used generative AI daily in their work. [employes.ch]
• 85% of respondents believed AI increased productivity, but only 21% considered it trustworthy and 20% completely safe. [employes.ch]
• 32% of Swiss companies surveyed by Deloitte Switzerland considered their infrastructure well or very well prepared for the adoption of generative AI. [deloitte.com]
• 72% of working people in French-speaking Switzerland reportedly use AI at work, according to the Qualinsight study reported by RTS in June 2026. [rts.ch]
• 81% of employees surveyed in UpGuard’s international study said they used unapproved AI tools. This international figure does not specifically measure the situation in Switzerland. [upguard.com]

15Sources and Further Reading

• Federal Statistical Office: Artificial intelligence
• FDPIC (PFPDT): AI and data protection
• OFCOM: overview and Swiss regulatory approach
• Deloitte Switzerland: generative AI adoption in Swiss companies
• Swisscom: interview on the risks of Shadow AI
• RTS: the use of AI at work in French-speaking Switzerland
• Inria and datacraft: study on Shadow AI and informal uses